├── figures
├── acq.png
├── logo.png
├── rea.png
├── tasks.png
└── results.png
├── context_data
├── PaulGrahamEssays
│ ├── rss.txt
│ ├── pow.txt
│ ├── todo.txt
│ ├── nft.txt
│ ├── weird.txt
│ ├── rootsoflisp.txt
│ ├── foundervisa.txt
│ ├── iflisp.txt
│ ├── sun.txt
│ ├── want.txt
│ ├── unions.txt
│ ├── bias.txt
│ ├── know.txt
│ ├── mod.txt
│ ├── island.txt
│ ├── diff.txt
│ ├── founders.txt
│ ├── vw.txt
│ ├── copy.txt
│ ├── goodtaste.txt
│ ├── ecw.txt
│ ├── corpdev.txt
│ ├── addiction.txt
│ ├── newideas.txt
│ ├── aord.txt
│ ├── vcsqueeze.txt
│ ├── vb.txt
│ ├── hubs.txt
│ ├── gba.txt
│ ├── apple.txt
│ └── submarine.txt
├── a_stars.txt
└── r_stars.txt
├── LICENSE
├── README.md
├── viz.ipynb
└── gen_test_data.ipynb
/figures/acq.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nick7nlp/Counting-Stars/HEAD/figures/acq.png
--------------------------------------------------------------------------------
/figures/logo.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nick7nlp/Counting-Stars/HEAD/figures/logo.png
--------------------------------------------------------------------------------
/figures/rea.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nick7nlp/Counting-Stars/HEAD/figures/rea.png
--------------------------------------------------------------------------------
/figures/tasks.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nick7nlp/Counting-Stars/HEAD/figures/tasks.png
--------------------------------------------------------------------------------
/figures/results.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nick7nlp/Counting-Stars/HEAD/figures/results.png
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/rss.txt:
--------------------------------------------------------------------------------
1 | Aaron Swartz created a scraped
2 | feed
3 | of the essays page.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/pow.txt:
--------------------------------------------------------------------------------
1 | January 2017People who are powerful but uncharismatic will tend to be disliked.
2 | Their power makes them a target for criticism that they don't have
3 | the charisma to disarm. That was Hillary Clinton's problem. It also
4 | tends to be a problem for any CEO who is more of a builder than a
5 | schmoozer. And yet the builder-type CEO is (like Hillary) probably
6 | the best person for the job.I don't think there is any solution to this problem. It's human
7 | nature. The best we can do is to recognize that it's happening, and
8 | to understand that being a magnet for criticism is sometimes a sign
9 | not that someone is the wrong person for a job, but that they're
10 | the right one.
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | MIT License
2 |
3 | Copyright (c) 2024 Mingyang
4 |
5 | Permission is hereby granted, free of charge, to any person obtaining a copy
6 | of this software and associated documentation files (the "Software"), to deal
7 | in the Software without restriction, including without limitation the rights
8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9 | copies of the Software, and to permit persons to whom the Software is
10 | furnished to do so, subject to the following conditions:
11 |
12 | The above copyright notice and this permission notice shall be included in all
13 | copies or substantial portions of the Software.
14 |
15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 | SOFTWARE.
22 |
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/todo.txt:
--------------------------------------------------------------------------------
1 | April 2012A palliative care nurse called Bronnie Ware made a list of the
2 | biggest regrets
3 | of the dying. Her list seems plausible. I could see
4 | myself — can see myself — making at least 4 of these
5 | 5 mistakes.If you had to compress them into a single piece of advice, it might
6 | be: don't be a cog. The 5 regrets paint a portrait of post-industrial
7 | man, who shrinks himself into a shape that fits his circumstances,
8 | then turns dutifully till he stops.The alarming thing is, the mistakes that produce these regrets are
9 | all errors of omission. You forget your dreams, ignore your family,
10 | suppress your feelings, neglect your friends, and forget to be
11 | happy. Errors of omission are a particularly dangerous type of
12 | mistake, because you make them by default.I would like to avoid making these mistakes. But how do you avoid
13 | mistakes you make by default? Ideally you transform your life so
14 | it has other defaults. But it may not be possible to do that
15 | completely. As long as these mistakes happen by default, you probably
16 | have to be reminded not to make them. So I inverted the 5 regrets,
17 | yielding a list of 5 commands
18 |
19 | Don't ignore your dreams; don't work too much; say what you
20 | think; cultivate friendships; be happy.
21 |
22 | which I then put at the top of the file I use as a todo list.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/nft.txt:
--------------------------------------------------------------------------------
1 | May 2021Noora Health, a nonprofit I've
2 | supported for years, just launched
3 | a new NFT. It has a dramatic name, Save Thousands of Lives,
4 | because that's what the proceeds will do.Noora has been saving lives for 7 years. They run programs in
5 | hospitals in South Asia to teach new mothers how to take care of
6 | their babies once they get home. They're in 165 hospitals now. And
7 | because they know the numbers before and after they start at a new
8 | hospital, they can measure the impact they have. It is massive.
9 | For every 1000 live births, they save 9 babies.This number comes from a study
10 | of 133,733 families at 28 different
11 | hospitals that Noora conducted in collaboration with the Better
12 | Birth team at Ariadne Labs, a joint center for health systems
13 | innovation at Brigham and Womens Hospital and Harvard T.H. Chan
14 | School of Public Health.Noora is so effective that even if you measure their costs in the
15 | most conservative way, by dividing their entire budget by the number
16 | of lives saved, the cost of saving a life is the lowest I've seen.
17 | $1,235.For this NFT, they're going to issue a public report tracking how
18 | this specific tranche of money is spent, and estimating the number
19 | of lives saved as a result.NFTs are a new territory, and this way of using them is especially
20 | new, but I'm excited about its potential. And I'm excited to see
21 | what happens with this particular auction, because unlike an NFT
22 | representing something that has already happened,
23 | this NFT gets better as the price gets higher.The reserve price was about $2.5 million, because that's what it
24 | takes for the name to be accurate: that's what it costs to save
25 | 2000 lives. But the higher the price of this NFT goes, the more
26 | lives will be saved. What a sentence to be able to write.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/weird.txt:
--------------------------------------------------------------------------------
1 | August 2021When people say that in their experience all programming languages
2 | are basically equivalent, they're making a statement not about
3 | languages but about the kind of programming they've done.99.5% of programming consists of gluing together calls to library
4 | functions. All popular languages are equally good at this. So one
5 | can easily spend one's whole career operating in the intersection
6 | of popular programming languages.But the other .5% of programming is disproportionately interesting.
7 | If you want to learn what it consists of, the weirdness of weird
8 | languages is a good clue to follow.Weird languages aren't weird by accident. Not the good ones, at
9 | least. The weirdness of the good ones usually implies the existence
10 | of some form of programming that's not just the usual gluing together
11 | of library calls.A concrete example: Lisp macros. Lisp macros seem weird even to
12 | many Lisp programmers. They're not only not in the intersection of
13 | popular languages, but by their nature would be hard to implement
14 | properly in a language without turning it into a dialect of
15 | Lisp. And macros are definitely evidence of techniques that go
16 | beyond glue programming. For example, solving problems by first
17 | writing a language for problems of that type, and then writing
18 | your specific application in it. Nor is this all you can do with
19 | macros; it's just one region in a space of program-manipulating
20 | techniques that even now is far from fully explored.So if you want to expand your concept of what programming can be,
21 | one way to do it is by learning weird languages. Pick a language
22 | that most programmers consider weird but whose median user is smart,
23 | and then focus on the differences between this language and the
24 | intersection of popular languages. What can you say in this language
25 | that would be impossibly inconvenient to say in others? In the
26 | process of learning how to say things you couldn't previously say,
27 | you'll probably be learning how to think things you couldn't
28 | previously think.
29 | Thanks to Trevor Blackwell, Patrick Collison, Daniel Gackle, Amjad
30 | Masad, and Robert Morris for reading drafts of this.
31 |
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/rootsoflisp.txt:
--------------------------------------------------------------------------------
1 | May 2001
2 |
3 | (I wrote this article to help myself understand exactly
4 | what McCarthy discovered. You don't need to know this stuff
5 | to program in Lisp, but it should be helpful to
6 | anyone who wants to
7 | understand the essence of Lisp both in the sense of its
8 | origins and its semantic core. The fact that it has such a core
9 | is one of Lisp's distinguishing features, and the reason why,
10 | unlike other languages, Lisp has dialects.)In 1960, John
11 | McCarthy published a remarkable paper in
12 | which he did for programming something like what Euclid did for
13 | geometry. He showed how, given a handful of simple
14 | operators and a notation for functions, you can
15 | build a whole programming language.
16 | He called this language Lisp, for "List Processing,"
17 | because one of his key ideas was to use a simple
18 | data structure called a list for both
19 | code and data.It's worth understanding what McCarthy discovered, not
20 | just as a landmark in the history of computers, but as
21 | a model for what programming is tending to become in
22 | our own time. It seems to me that there have been
23 | two really clean, consistent models of programming so
24 | far: the C model and the Lisp model.
25 | These two seem points of high ground, with swampy lowlands
26 | between them. As computers have grown more powerful,
27 | the new languages being developed have been moving
28 | steadily toward the Lisp model. A popular recipe
29 | for new programming languages in the past 20 years
30 | has been to take the C model of computing and add to
31 | it, piecemeal, parts taken from the Lisp model,
32 | like runtime typing and garbage collection.In this article I'm going to try to explain in the
33 | simplest possible terms what McCarthy discovered.
34 | The point is not just to learn about an interesting
35 | theoretical result someone figured out forty years ago,
36 | but to show where languages are heading.
37 | The unusual thing about Lisp in fact, the defining
38 | quality of Lisp is that it can be written in
39 | itself. To understand what McCarthy meant by this,
40 | we're going to retrace his steps, with his mathematical
41 | notation translated into running Common Lisp code.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/foundervisa.txt:
--------------------------------------------------------------------------------
1 |
2 |
3 | April 2009I usually avoid politics, but since we now seem to have an administration that's open to suggestions, I'm going to risk making one. The single biggest thing the government could do to increase the number of startups in this country is a policy that would cost nothing: establish a new class of visa for startup founders.The biggest constraint on the number of new startups that get created in the US is not tax policy or employment law or even Sarbanes-Oxley. It's that we won't let the people who want to start them into the country.Letting just 10,000 startup founders into the country each year could have a visible effect on the economy. If we assume 4 people per startup, which is probably an overestimate, that's 2500 new companies. Each year. They wouldn't all grow as big as Google, but out of 2500 some would come close.By definition these 10,000 founders wouldn't be taking jobs from Americans: it could be part of the terms of the visa that they couldn't work for existing companies, only new ones they'd founded. In fact they'd cause there to be
4 | more jobs for Americans, because the companies they started would hire more employees as they grew.The tricky part might seem to be how one defined a startup. But that could be solved quite easily: let the market decide. Startup investors work hard to find the best startups. The government could not do better than to piggyback on their expertise, and use investment by recognized startup investors as the test of whether a company was a real startup.How would the government decide who's a startup investor? The same way they decide what counts as a university for student visas. We'll establish our own accreditation procedure. We know who one another are.10,000 people is a drop in the bucket by immigration standards, but would represent a huge increase in the pool of startup founders. I think this would have such a visible effect on the economy that it would make the legislator who introduced the bill famous. The only way to know for sure would be to try it, and that would cost practically nothing.
5 | Thanks to Trevor Blackwell, Paul Buchheit, Jeff Clavier, David Hornik, Jessica Livingston, Greg Mcadoo, Aydin Senkut, and Fred Wilson for reading drafts of this.Related:
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/iflisp.txt:
--------------------------------------------------------------------------------
1 | May 2003If Lisp is so great, why don't more people use it? I was
2 | asked this question by a student in the audience at a
3 | talk I gave recently. Not for the first time, either.In languages, as in so many things, there's not much
4 | correlation between popularity and quality. Why does
5 | John Grisham (King of Torts sales rank, 44) outsell
6 | Jane Austen (Pride and Prejudice sales rank, 6191)?
7 | Would even Grisham claim that it's because he's a better
8 | writer?Here's the first sentence of Pride and Prejudice:
9 |
10 | It is a truth universally acknowledged, that a single man
11 | in possession of a good fortune must be in want of a
12 | wife.
13 |
14 | "It is a truth universally acknowledged?" Long words for
15 | the first sentence of a love story.Like Jane Austen, Lisp looks hard. Its syntax, or lack
16 | of syntax, makes it look completely unlike
17 | the languages
18 | most people are used to. Before I learned Lisp, I was afraid
19 | of it too. I recently came across a notebook from 1983
20 | in which I'd written:
21 |
22 | I suppose I should learn Lisp, but it seems so foreign.
23 |
24 | Fortunately, I was 19 at the time and not too resistant to learning
25 | new things. I was so ignorant that learning
26 | almost anything meant learning new things.People frightened by Lisp make up other reasons for not
27 | using it. The standard
28 | excuse, back when C was the default language, was that Lisp
29 | was too slow. Now that Lisp dialects are among
30 | the faster
31 | languages available, that excuse has gone away.
32 | Now the standard excuse is openly circular: that other languages
33 | are more popular.(Beware of such reasoning. It gets you Windows.)Popularity is always self-perpetuating, but it's especially
34 | so in programming languages. More libraries
35 | get written for popular languages, which makes them still
36 | more popular. Programs often have to work with existing programs,
37 | and this is easier if they're written in the same language,
38 | so languages spread from program to program like a virus.
39 | And managers prefer popular languages, because they give them
40 | more leverage over developers, who can more easily be replaced.Indeed, if programming languages were all more or less equivalent,
41 | there would be little justification for using any but the most
42 | popular. But they aren't all equivalent, not by a long
43 | shot. And that's why less popular languages, like Jane Austen's
44 | novels, continue to survive at all. When everyone else is reading
45 | the latest John Grisham novel, there will always be a few people
46 | reading Jane Austen instead.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/sun.txt:
--------------------------------------------------------------------------------
1 | September 2017The most valuable insights are both general and surprising.
2 | F = ma for example. But general and surprising is a hard
3 | combination to achieve. That territory tends to be picked
4 | clean, precisely because those insights are so valuable.Ordinarily, the best that people can do is one without the
5 | other: either surprising without being general (e.g.
6 | gossip), or general without being surprising (e.g.
7 | platitudes).Where things get interesting is the moderately valuable
8 | insights. You get those from small additions of whichever
9 | quality was missing. The more common case is a small
10 | addition of generality: a piece of gossip that's more than
11 | just gossip, because it teaches something interesting about
12 | the world. But another less common approach is to focus on
13 | the most general ideas and see if you can find something new
14 | to say about them. Because these start out so general, you
15 | only need a small delta of novelty to produce a useful
16 | insight.A small delta of novelty is all you'll be able to get most
17 | of the time. Which means if you take this route, your ideas
18 | will seem a lot like ones that already exist. Sometimes
19 | you'll find you've merely rediscovered an idea that did
20 | already exist. But don't be discouraged. Remember the huge
21 | multiplier that kicks in when you do manage to think of
22 | something even a little new.Corollary: the more general the ideas you're talking about,
23 | the less you should worry about repeating yourself. If you
24 | write enough, it's inevitable you will. Your brain is much
25 | the same from year to year and so are the stimuli that hit
26 | it. I feel slightly bad when I find I've said something
27 | close to what I've said before, as if I were plagiarizing
28 | myself. But rationally one shouldn't. You won't say
29 | something exactly the same way the second time, and that
30 | variation increases the chance you'll get that tiny but
31 | critical delta of novelty.And of course, ideas beget ideas. (That sounds
32 | familiar.)
33 | An idea with a small amount of novelty could lead to one
34 | with more. But only if you keep going. So it's doubly
35 | important not to let yourself be discouraged by people who
36 | say there's not much new about something you've discovered.
37 | "Not much new" is a real achievement when you're talking
38 | about the most general ideas. It's not true that there's nothing new under the sun. There
39 | are some domains where there's almost nothing new. But
40 | there's a big difference between nothing and almost nothing,
41 | when it's multiplied by the area under the sun.
42 | Thanks to Sam Altman, Patrick Collison, and Jessica
43 | Livingston for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/want.txt:
--------------------------------------------------------------------------------
1 | November 2022Since I was about 9 I've been puzzled by the apparent contradiction
2 | between being made of matter that behaves in a predictable way, and
3 | the feeling that I could choose to do whatever I wanted. At the
4 | time I had a self-interested motive for exploring the question. At
5 | that age (like most succeeding ages) I was always in trouble with
6 | the authorities, and it seemed to me that there might possibly be
7 | some way to get out of trouble by arguing that I wasn't responsible
8 | for my actions. I gradually lost hope of that, but the puzzle
9 | remained: How do you reconcile being a machine made of matter with
10 | the feeling that you're free to choose what you do?
11 | [1]The best way to explain the answer may be to start with a slightly
12 | wrong version, and then fix it. The wrong version is: You can do
13 | what you want, but you can't want what you want. Yes, you can control
14 | what you do, but you'll do what you want, and you can't control
15 | that.The reason this is mistaken is that people do sometimes change what
16 | they want. People who don't want to want something — drug addicts,
17 | for example — can sometimes make themselves stop wanting it. And
18 | people who want to want something — who want to like classical
19 | music, or broccoli — sometimes succeed.So we modify our initial statement: You can do what you want, but
20 | you can't want to want what you want.That's still not quite true. It's possible to change what you want
21 | to want. I can imagine someone saying "I decided to stop wanting
22 | to like classical music." But we're getting closer to the truth.
23 | It's rare for people to change what they want to want, and the more
24 | "want to"s we add, the rarer it gets.We can get arbitrarily close to a true statement by adding more "want
25 | to"s in much the same way we can get arbitrarily close to 1 by adding
26 | more 9s to a string of 9s following a decimal point. In practice
27 | three or four "want to"s must surely be enough. It's hard even to
28 | envision what it would mean to change what you want to want to want
29 | to want, let alone actually do it.So one way to express the correct answer is to use a regular
30 | expression. You can do what you want, but there's some statement
31 | of the form "you can't (want to)* want what you want" that's true.
32 | Ultimately you get back to a want that you don't control.
33 | [2]
34 | Notes[1]
35 | I didn't know when I was 9 that matter might behave randomly,
36 | but I don't think it affects the problem much. Randomness destroys
37 | the ghost in the machine as effectively as determinism.[2]
38 | If you don't like using an expression, you can make the same
39 | point using higher-order desires: There is some n such that you
40 | don't control your nth-order desires.
41 | Thanks to Trevor Blackwell,
42 | Jessica Livingston, Robert Morris, and
43 | Michael Nielsen for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/unions.txt:
--------------------------------------------------------------------------------
1 | May 2007People who worry about the increasing gap between rich and poor
2 | generally look back on the mid twentieth century as a golden age.
3 | In those days we had a large number of high-paying union manufacturing
4 | jobs that boosted the median income. I wouldn't quite call the
5 | high-paying union job a myth, but I think people who dwell on it
6 | are reading too much into it.Oddly enough, it was working with startups that made me realize
7 | where the high-paying union job came from. In a rapidly growing
8 | market, you don't worry too much about efficiency. It's more
9 | important to grow fast. If there's some mundane problem getting
10 | in your way, and there's a simple solution that's somewhat expensive,
11 | just take it and get on with more important things. EBay didn't
12 | win by paying less for servers than their competitors.Difficult though it may be to imagine now, manufacturing was a
13 | growth industry in the mid twentieth century. This was an era when
14 | small firms making everything from cars to candy were getting
15 | consolidated into a new kind of corporation with national reach and
16 | huge economies of scale. You had to grow fast or die. Workers
17 | were for these companies what servers are for an Internet startup.
18 | A reliable supply was more important than low cost.If you looked in the head of a 1950s auto executive, the attitude
19 | must have been: sure, give 'em whatever they ask for, so long as
20 | the new model isn't delayed.In other words, those workers were not paid what their work was
21 | worth. Circumstances being what they were, companies would have
22 | been stupid to insist on paying them so little.If you want a less controversial example of this phenomenon, ask
23 | anyone who worked as a consultant building web sites during the
24 | Internet Bubble. In the late nineties you could get paid huge sums
25 | of money for building the most trivial things. And yet does anyone
26 | who was there have any expectation those days will ever return? I
27 | doubt it. Surely everyone realizes that was just a temporary
28 | aberration.The era of labor unions seems to have been the same kind of aberration,
29 | just spread
30 | over a longer period, and mixed together with a lot of ideology
31 | that prevents people from viewing it with as cold an eye as they
32 | would something like consulting during the Bubble.Basically, unions were just Razorfish.People who think the labor movement was the creation of heroic union
33 | organizers have a problem to explain: why are unions shrinking now?
34 | The best they can do is fall back on the default explanation of
35 | people living in fallen civilizations. Our ancestors were giants.
36 | The workers of the early twentieth century must have had a moral
37 | courage that's lacking today.In fact there's a simpler explanation. The early twentieth century
38 | was just a fast-growing startup overpaying for infrastructure. And
39 | we in the present are not a fallen people, who have abandoned
40 | whatever mysterious high-minded principles produced the high-paying
41 | union job. We simply live in a time when the fast-growing companies
42 | overspend on different things.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/bias.txt:
--------------------------------------------------------------------------------
1 | October 2015This will come as a surprise to a lot of people, but in some cases
2 | it's possible to detect bias in a selection process without knowing
3 | anything about the applicant pool. Which is exciting because among
4 | other things it means third parties can use this technique to detect
5 | bias whether those doing the selecting want them to or not.You can use this technique whenever (a) you have at least
6 | a random sample of the applicants that were selected, (b) their
7 | subsequent performance is measured, and (c) the groups of
8 | applicants you're comparing have roughly equal distribution of ability.How does it work? Think about what it means to be biased. What
9 | it means for a selection process to be biased against applicants
10 | of type x is that it's harder for them to make it through. Which
11 | means applicants of type x have to be better to get selected than
12 | applicants not of type x.
13 | [1]
14 | Which means applicants of type x
15 | who do make it through the selection process will outperform other
16 | successful applicants. And if the performance of all the successful
17 | applicants is measured, you'll know if they do.Of course, the test you use to measure performance must be a valid
18 | one. And in particular it must not be invalidated by the bias you're
19 | trying to measure.
20 | But there are some domains where performance can be measured, and
21 | in those detecting bias is straightforward. Want to know if the
22 | selection process was biased against some type of applicant? Check
23 | whether they outperform the others. This is not just a heuristic
24 | for detecting bias. It's what bias means.For example, many suspect that venture capital firms are biased
25 | against female founders. This would be easy to detect: among their
26 | portfolio companies, do startups with female founders outperform
27 | those without? A couple months ago, one VC firm (almost certainly
28 | unintentionally) published a study showing bias of this type. First
29 | Round Capital found that among its portfolio companies, startups
30 | with female founders outperformed
31 | those without by 63%.
32 | [2]The reason I began by saying that this technique would come as a
33 | surprise to many people is that we so rarely see analyses of this
34 | type. I'm sure it will come as a surprise to First Round that they
35 | performed one. I doubt anyone there realized that by limiting their
36 | sample to their own portfolio, they were producing a study not of
37 | startup trends but of their own biases when selecting companies.I predict we'll see this technique used more in the future. The
38 | information needed to conduct such studies is increasingly available.
39 | Data about who applies for things is usually closely guarded by the
40 | organizations selecting them, but nowadays data about who gets
41 | selected is often publicly available to anyone who takes the trouble
42 | to aggregate it.
43 | Notes[1]
44 | This technique wouldn't work if the selection process looked
45 | for different things from different types of applicants—for
46 | example, if an employer hired men based on their ability but women
47 | based on their appearance.[2]
48 | As Paul Buchheit points out, First Round excluded their most
49 | successful investment, Uber, from the study. And while it
50 | makes sense to exclude outliers from some types of studies,
51 | studies of returns from startup investing, which is all about
52 | hitting outliers, are not one of them.
53 | Thanks to Sam Altman, Jessica Livingston, and Geoff Ralston for reading
54 | drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/know.txt:
--------------------------------------------------------------------------------
1 | December 2014I've read Villehardouin's chronicle of the Fourth Crusade at least
2 | two times, maybe three. And yet if I had to write down everything
3 | I remember from it, I doubt it would amount to much more than a
4 | page. Multiply this times several hundred, and I get an uneasy
5 | feeling when I look at my bookshelves. What use is it to read all
6 | these books if I remember so little from them?A few months ago, as I was reading Constance Reid's excellent
7 | biography of Hilbert, I figured out if not the answer to this
8 | question, at least something that made me feel better about it.
9 | She writes:
10 |
11 | Hilbert had no patience with mathematical lectures which filled
12 | the students with facts but did not teach them how to frame a
13 | problem and solve it. He often used to tell them that "a perfect
14 | formulation of a problem is already half its solution."
15 |
16 | That has always seemed to me an important point, and I was even
17 | more convinced of it after hearing it confirmed by Hilbert.But how had I come to believe in this idea in the first place? A
18 | combination of my own experience and other things I'd read. None
19 | of which I could at that moment remember! And eventually I'd forget
20 | that Hilbert had confirmed it too. But my increased belief in the
21 | importance of this idea would remain something I'd learned from
22 | this book, even after I'd forgotten I'd learned it.Reading and experience train your model of the world. And even if
23 | you forget the experience or what you read, its effect on your model
24 | of the world persists. Your mind is like a compiled program you've
25 | lost the source of. It works, but you don't know why.The place to look for what I learned from Villehardouin's chronicle
26 | is not what I remember from it, but my mental models of the crusades,
27 | Venice, medieval culture, siege warfare, and so on. Which doesn't
28 | mean I couldn't have read more attentively, but at least the harvest
29 | of reading is not so miserably small as it might seem.This is one of those things that seem obvious in retrospect. But
30 | it was a surprise to me and presumably would be to anyone else who
31 | felt uneasy about (apparently) forgetting so much they'd read.Realizing it does more than make you feel a little better about
32 | forgetting, though. There are specific implications.For example, reading and experience are usually "compiled" at the
33 | time they happen, using the state of your brain at that time. The
34 | same book would get compiled differently at different points in
35 | your life. Which means it is very much worth reading important
36 | books multiple times. I always used to feel some misgivings about
37 | rereading books. I unconsciously lumped reading together with work
38 | like carpentry, where having to do something again is a sign you
39 | did it wrong the first time. Whereas now the phrase "already read"
40 | seems almost ill-formed.Intriguingly, this implication isn't limited to books. Technology
41 | will increasingly make it possible to relive our experiences. When
42 | people do that today it's usually to enjoy them again (e.g. when
43 | looking at pictures of a trip) or to find the origin of some bug in
44 | their compiled code (e.g. when Stephen Fry succeeded in remembering
45 | the childhood trauma that prevented him from singing). But as
46 | technologies for recording and playing back your life improve, it
47 | may become common for people to relive experiences without any goal
48 | in mind, simply to learn from them again as one might when rereading
49 | a book.Eventually we may be able not just to play back experiences but
50 | also to index and even edit them. So although not knowing how you
51 | know things may seem part of being human, it may not be.
52 | Thanks to Sam Altman, Jessica Livingston, and Robert Morris for reading
53 | drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/mod.txt:
--------------------------------------------------------------------------------
1 | December 2019There are two distinct ways to be politically moderate: on purpose
2 | and by accident. Intentional moderates are trimmers, deliberately
3 | choosing a position mid-way between the extremes of right and left.
4 | Accidental moderates end up in the middle, on average, because they
5 | make up their own minds about each question, and the far right and
6 | far left are roughly equally wrong.You can distinguish intentional from accidental moderates by the
7 | distribution of their opinions. If the far left opinion on some
8 | matter is 0 and the far right opinion 100, an intentional moderate's
9 | opinion on every question will be near 50. Whereas an accidental
10 | moderate's opinions will be scattered over a broad range, but will,
11 | like those of the intentional moderate, average to about 50.Intentional moderates are similar to those on the far left and the
12 | far right in that their opinions are, in a sense, not their own.
13 | The defining quality of an ideologue, whether on the left or the
14 | right, is to acquire one's opinions in bulk. You don't get to pick
15 | and choose. Your opinions about taxation can be predicted from your
16 | opinions about sex. And although intentional moderates
17 | might seem to be the opposite of ideologues, their beliefs (though
18 | in their case the word "positions" might be more accurate) are also
19 | acquired in bulk. If the median opinion shifts to the right or left,
20 | the intentional moderate must shift with it. Otherwise they stop
21 | being moderate.Accidental moderates, on the other hand, not only choose their own
22 | answers, but choose their own questions. They may not care at all
23 | about questions that the left and right both think are terribly
24 | important. So you can only even measure the politics of an accidental
25 | moderate from the intersection of the questions they care about and
26 | those the left and right care about, and this can
27 | sometimes be vanishingly small.It is not merely a manipulative rhetorical trick to say "if you're
28 | not with us, you're against us," but often simply false.Moderates are sometimes derided as cowards, particularly by
29 | the extreme left. But while it may be accurate to call intentional
30 | moderates cowards, openly being an accidental moderate requires the
31 | most courage of all, because you get attacked from both right and
32 | left, and you don't have the comfort of being an orthodox member
33 | of a large group to sustain you.Nearly all the most impressive people I know are accidental moderates.
34 | If I knew a lot of professional athletes, or people in the entertainment
35 | business, that might be different. Being on the far left or far
36 | right doesn't affect how fast you run or how well you sing. But
37 | someone who works with ideas has to be independent-minded to do it
38 | well.Or more precisely, you have to be independent-minded about the ideas
39 | you work with. You could be mindlessly doctrinaire in your politics
40 | and still be a good mathematician. In the 20th century, a lot of
41 | very smart people were Marxists just no one who was smart about
42 | the subjects Marxism involves. But if the ideas you use in your
43 | work intersect with the politics of your time, you have two choices:
44 | be an accidental moderate, or be mediocre.Notes[1] It's possible in theory for one side to be entirely right and
45 | the other to be entirely wrong. Indeed, ideologues must always
46 | believe this is the case. But historically it rarely has been.[2] For some reason the far right tend to ignore moderates rather
47 | than despise them as backsliders. I'm not sure why. Perhaps it
48 | means that the far right is less ideological than the far left. Or
49 | perhaps that they are more confident, or more resigned, or simply
50 | more disorganized. I just don't know.[3] Having heretical opinions doesn't mean you have to express
51 | them openly. It may be
52 | easier to have them if you don't.
53 | Thanks to Austen Allred, Trevor Blackwell, Patrick Collison, Jessica Livingston,
54 | Amjad Masad, Ryan Petersen, and Harj Taggar for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/island.txt:
--------------------------------------------------------------------------------
1 | July 2006I've discovered a handy test for figuring out what you're addicted
2 | to. Imagine you were going to spend the weekend at a friend's house
3 | on a little island off the coast of Maine. There are no shops on
4 | the island and you won't be able to leave while you're there. Also,
5 | you've never been to this house before, so you can't assume it will
6 | have more than any house might.What, besides clothes and toiletries, do you make a point of packing?
7 | That's what you're addicted to. For example, if you find yourself
8 | packing a bottle of vodka (just in case), you may want to stop and
9 | think about that.For me the list is four things: books, earplugs, a notebook, and a
10 | pen.There are other things I might bring if I thought of it, like music,
11 | or tea, but I can live without them. I'm not so addicted to caffeine
12 | that I wouldn't risk the house not having any tea, just for a
13 | weekend.Quiet is another matter. I realize it seems a bit eccentric to
14 | take earplugs on a trip to an island off the coast of Maine. If
15 | anywhere should be quiet, that should. But what if the person in
16 | the next room snored? What if there was a kid playing basketball?
17 | (Thump, thump, thump... thump.) Why risk it? Earplugs are small.Sometimes I can think with noise. If I already have momentum on
18 | some project, I can work in noisy places. I can edit an essay or
19 | debug code in an airport. But airports are not so bad: most of the
20 | noise is whitish. I couldn't work with the sound of a sitcom coming
21 | through the wall, or a car in the street playing thump-thump music.And of course there's another kind of thinking, when you're starting
22 | something new, that requires complete quiet. You never
23 | know when this will strike. It's just as well to carry plugs.The notebook and pen are professional equipment, as it were. Though
24 | actually there is something druglike about them, in the sense that
25 | their main purpose is to make me feel better. I hardly ever go
26 | back and read stuff I write down in notebooks. It's just that if
27 | I can't write things down, worrying about remembering one idea gets
28 | in the way of having the next. Pen and paper wick ideas.The best notebooks I've found are made by a company called Miquelrius.
29 | I use their smallest size, which is about 2.5 x 4 in.
30 | The secret to writing on such
31 | narrow pages is to break words only when you run out of space, like
32 | a Latin inscription. I use the cheapest plastic Bic ballpoints,
33 | partly because their gluey ink doesn't seep through pages, and
34 | partly so I don't worry about losing them.I only started carrying a notebook about three years ago. Before
35 | that I used whatever scraps of paper I could find. But the problem
36 | with scraps of paper is that they're not ordered. In a notebook
37 | you can guess what a scribble means by looking at the pages
38 | around it. In the scrap era I was constantly finding notes I'd
39 | written years before that might say something I needed to remember,
40 | if I could only figure out what.As for books, I know the house would probably have something to
41 | read. On the average trip I bring four books and only read one of
42 | them, because I find new books to read en route. Really bringing
43 | books is insurance.I realize this dependence on books is not entirely good—that what
44 | I need them for is distraction. The books I bring on trips are
45 | often quite virtuous, the sort of stuff that might be assigned
46 | reading in a college class. But I know my motives aren't virtuous.
47 | I bring books because if the world gets boring I need to be able
48 | to slip into another distilled by some writer. It's like eating
49 | jam when you know you should be eating fruit.There is a point where I'll do without books. I was walking in
50 | some steep mountains once, and decided I'd rather just think, if I
51 | was bored, rather than carry a single unnecessary ounce. It wasn't
52 | so bad. I found I could entertain myself by having ideas instead
53 | of reading other people's. If you stop eating jam, fruit starts
54 | to taste better.So maybe I'll try not bringing books on some future trip. They're
55 | going to have to pry the plugs out of my cold, dead ears, however.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/diff.txt:
--------------------------------------------------------------------------------
1 | December 2001 (rev. May 2002)
2 |
3 | (This article came about in response to some questions on
4 | the LL1 mailing list. It is now
5 | incorporated in Revenge of the Nerds.)When McCarthy designed Lisp in the late 1950s, it was
6 | a radical departure from existing languages,
7 | the most important of which was Fortran.Lisp embodied nine new ideas:
8 | 1. Conditionals. A conditional is an if-then-else
9 | construct. We take these for granted now. They were
10 | invented
11 | by McCarthy in the course of developing Lisp.
12 | (Fortran at that time only had a conditional
13 | goto, closely based on the branch instruction in the
14 | underlying hardware.) McCarthy, who was on the Algol committee, got
15 | conditionals into Algol, whence they spread to most other
16 | languages.2. A function type. In Lisp, functions are first class
17 | objects-- they're a data type just like integers, strings,
18 | etc, and have a literal representation, can be stored in variables,
19 | can be passed as arguments, and so on.3. Recursion. Recursion existed as a mathematical concept
20 | before Lisp of course, but Lisp was the first programming language to support
21 | it. (It's arguably implicit in making functions first class
22 | objects.)4. A new concept of variables. In Lisp, all variables
23 | are effectively pointers. Values are what
24 | have types, not variables, and assigning or binding
25 | variables means copying pointers, not what they point to.5. Garbage-collection.6. Programs composed of expressions. Lisp programs are
26 | trees of expressions, each of which returns a value.
27 | (In some Lisps expressions
28 | can return multiple values.) This is in contrast to Fortran
29 | and most succeeding languages, which distinguish between
30 | expressions and statements.It was natural to have this
31 | distinction in Fortran because (not surprisingly in a language
32 | where the input format was punched cards) the language was
33 | line-oriented. You could not nest statements. And
34 | so while you needed expressions for math to work, there was
35 | no point in making anything else return a value, because
36 | there could not be anything waiting for it.This limitation
37 | went away with the arrival of block-structured languages,
38 | but by then it was too late. The distinction between
39 | expressions and statements was entrenched. It spread from
40 | Fortran into Algol and thence to both their descendants.When a language is made entirely of expressions, you can
41 | compose expressions however you want. You can say either
42 | (using Arc syntax)(if foo (= x 1) (= x 2))or(= x (if foo 1 2))7. A symbol type. Symbols differ from strings in that
43 | you can test equality by comparing a pointer.8. A notation for code using trees of symbols.9. The whole language always available.
44 | There is
45 | no real distinction between read-time, compile-time, and runtime.
46 | You can compile or run code while reading, read or run code
47 | while compiling, and read or compile code at runtime.Running code at read-time lets users reprogram Lisp's syntax;
48 | running code at compile-time is the basis of macros; compiling
49 | at runtime is the basis of Lisp's use as an extension
50 | language in programs like Emacs; and reading at runtime
51 | enables programs to communicate using s-expressions, an
52 | idea recently reinvented as XML.
53 | When Lisp was first invented, all these ideas were far
54 | removed from ordinary programming practice, which was
55 | dictated largely by the hardware available in the late 1950s.Over time, the default language, embodied
56 | in a succession of popular languages, has
57 | gradually evolved toward Lisp. 1-5 are now widespread.
58 | 6 is starting to appear in the mainstream.
59 | Python has a form of 7, though there doesn't seem to be
60 | any syntax for it.
61 | 8, which (with 9) is what makes Lisp macros
62 | possible, is so far still unique to Lisp,
63 | perhaps because (a) it requires those parens, or something
64 | just as bad, and (b) if you add that final increment of power,
65 | you can no
66 | longer claim to have invented a new language, but only
67 | to have designed a new dialect of Lisp ; -)Though useful to present-day programmers, it's
68 | strange to describe Lisp in terms of its
69 | variation from the random expedients other languages
70 | adopted. That was not, probably, how McCarthy
71 | thought of it. Lisp wasn't designed to fix the mistakes
72 | in Fortran; it came about more as the byproduct of an
73 | attempt to axiomatize computation.
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 |
2 |
Counting-Stars (★): A Multi-evidence, Position-aware, and Scalable Benchmark for Evaluating Long-Context Large Language Models
3 |
4 |
5 |
6 |
7 |
48 | 
49 |
50 |
51 | ## Evaluation Results
52 |
53 |
54 | 
55 |
56 |
57 |
58 | 
59 |
60 |
61 | > Visualization of the results on the Chinese version of the Counting-Stars-32-(Multi-evidence Acquisition).
62 |
63 |
64 | 
65 |
66 |
67 | > Visualization of the results on the Chinese version of the Counting-Stars-32-(Multi-evidence Reasoning).
68 |
69 | ## Cite
70 | If you find our work helpful, feel free to give us a cite.
71 |
72 | ```
73 | @inproceedings{song-etal-2025-counting,
74 | title = "Counting-Stars: A Multi-evidence, Position-aware, and Scalable Benchmark for Evaluating Long-Context Large Language Models",
75 | author = "Song, Mingyang and Zheng, Mao and Luo, Xuan",
76 | booktitle = "Proceedings of the 31st International Conference on Computational Linguistics",
77 | year = "2025",
78 | address = "Abu Dhabi, UAE",
79 | publisher = "Association for Computational Linguistics",
80 | url = "https://aclanthology.org/2025.coling-main.253",
81 | pages = "3753--3763"
82 | }
83 | ```
84 |
85 | ## CONTACT
86 | For any questions, feel free to create an issue, and we will try our best to solve it. \
87 | **If the problem is more urgent**, you can email me simultaneously (I check email almost daily).
88 | ```
89 | NAME: Mingyang Song
90 | EMAIL: nickmysong@tencent.com
91 | ```
92 | Our visualization code is built on the source code from [NeedleInAHaystack](https://github.com/gkamradt/LLMTest_NeedleInAHaystack). Thanks for their work.
93 |
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/founders.txt:
--------------------------------------------------------------------------------
1 |
2 |
3 | Want to start a startup? Get funded by
4 | Y Combinator.
5 |
6 |
7 |
8 |
9 | October 2010
10 |
11 | (I wrote this for Forbes, who asked me to write something
12 | about the qualities we look for in founders. In print they had to cut
13 | the last item because they didn't have room.)1. DeterminationThis has turned out to be the most important quality in startup
14 | founders. We thought when we started Y Combinator that the most
15 | important quality would be intelligence. That's the myth in the
16 | Valley. And certainly you don't want founders to be stupid. But
17 | as long as you're over a certain threshold of intelligence, what
18 | matters most is determination. You're going to hit a lot of
19 | obstacles. You can't be the sort of person who gets demoralized
20 | easily.Bill Clerico and Rich Aberman of WePay
21 | are a good example. They're
22 | doing a finance startup, which means endless negotiations with big,
23 | bureaucratic companies. When you're starting a startup that depends
24 | on deals with big companies to exist, it often feels like they're
25 | trying to ignore you out of existence. But when Bill Clerico starts
26 | calling you, you may as well do what he asks, because he is not
27 | going away.
28 | 2. FlexibilityYou do not however want the sort of determination implied by phrases
29 | like "don't give up on your dreams." The world of startups is so
30 | unpredictable that you need to be able to modify your dreams on the
31 | fly. The best metaphor I've found for the combination of determination
32 | and flexibility you need is a running back.
33 | He's determined to get
34 | downfield, but at any given moment he may need to go sideways or
35 | even backwards to get there.The current record holder for flexibility may be Daniel Gross of
36 | Greplin. He applied to YC with
37 | some bad ecommerce idea. We told
38 | him we'd fund him if he did something else. He thought for a second,
39 | and said ok. He then went through two more ideas before settling
40 | on Greplin. He'd only been working on it for a couple days when
41 | he presented to investors at Demo Day, but he got a lot of interest.
42 | He always seems to land on his feet.
43 | 3. ImaginationIntelligence does matter a lot of course. It seems like the type
44 | that matters most is imagination. It's not so important to be able
45 | to solve predefined problems quickly as to be able to come up with
46 | surprising new ideas. In the startup world, most good ideas
47 | seem
48 | bad initially. If they were obviously good, someone would already
49 | be doing them. So you need the kind of intelligence that produces
50 | ideas with just the right level of craziness.Airbnb is that kind of idea.
51 | In fact, when we funded Airbnb, we
52 | thought it was too crazy. We couldn't believe large numbers of
53 | people would want to stay in other people's places. We funded them
54 | because we liked the founders so much. As soon as we heard they'd
55 | been supporting themselves by selling Obama and McCain branded
56 | breakfast cereal, they were in. And it turned out the idea was on
57 | the right side of crazy after all.
58 | 4. NaughtinessThough the most successful founders are usually good people, they
59 | tend to have a piratical gleam in their eye. They're not Goody
60 | Two-Shoes type good. Morally, they care about getting the big
61 | questions right, but not about observing proprieties. That's why
62 | I'd use the word naughty rather than evil. They delight in
63 | breaking
64 | rules, but not rules that matter. This quality may be redundant
65 | though; it may be implied by imagination.Sam Altman of Loopt
66 | is one of the most successful alumni, so we
67 | asked him what question we could put on the Y Combinator application
68 | that would help us discover more people like him. He said to ask
69 | about a time when they'd hacked something to their advantage—hacked in the sense of beating the system, not breaking into
70 | computers. It has become one of the questions we pay most attention
71 | to when judging applications.
72 | 5. FriendshipEmpirically it seems to be hard to start a startup with just
73 | one
74 | founder. Most of the big successes have two or three. And the
75 | relationship between the founders has to be strong. They must
76 | genuinely like one another, and work well together. Startups do
77 | to the relationship between the founders what a dog does to a sock:
78 | if it can be pulled apart, it will be.Emmett Shear and Justin Kan of Justin.tv
79 | are a good example of close
80 | friends who work well together. They've known each other since
81 | second grade. They can practically read one another's minds. I'm
82 | sure they argue, like all founders, but I have never once sensed
83 | any unresolved tension between them.Thanks to Jessica Livingston and Chris Steiner for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/vw.txt:
--------------------------------------------------------------------------------
1 | January 2012A few hours before the Yahoo acquisition was announced in June 1998
2 | I took a snapshot of Viaweb's
3 | site. I thought it might be interesting to look at one day.The first thing one notices is is how tiny the pages are. Screens
4 | were a lot smaller in 1998. If I remember correctly, our frontpage
5 | used to just fit in the size window people typically used then.Browsers then (IE 6 was still 3 years in the future) had few fonts
6 | and they weren't antialiased. If you wanted to make pages that
7 | looked good, you had to render display text as images.You may notice a certain similarity between the Viaweb and Y Combinator logos. We did that
8 | as an inside joke when we started YC. Considering how basic a red
9 | circle is, it seemed surprising to me when we started Viaweb how
10 | few other companies used one as their logo. A bit later I realized
11 | why.On the Company
12 | page you'll notice a mysterious individual called John McArtyem.
13 | Robert Morris (aka Rtm) was so publicity averse after the
14 | Worm that he
15 | didn't want his name on the site. I managed to get him to agree
16 | to a compromise: we could use his bio but not his name. He has
17 | since relaxed a bit
18 | on that point.Trevor graduated at about the same time the acquisition closed, so in the
19 | course of 4 days he went from impecunious grad student to millionaire
20 | PhD. The culmination of my career as a writer of press releases
21 | was one celebrating
22 | his graduation, illustrated with a drawing I did of him during
23 | a meeting.(Trevor also appears as Trevino
24 | Bagwell in our directory of web designers merchants could hire
25 | to build stores for them. We inserted him as a ringer in case some
26 | competitor tried to spam our web designers. We assumed his logo
27 | would deter any actual customers, but it did not.)Back in the 90s, to get users you had to get mentioned in magazines
28 | and newspapers. There were not the same ways to get found online
29 | that there are today. So we used to pay a PR
30 | firm $16,000 a month to get us mentioned in the press. Fortunately
31 | reporters liked
32 | us.In our advice about
33 | getting traffic from search engines (I don't think the term SEO
34 | had been coined yet), we say there are only 7 that matter: Yahoo,
35 | AltaVista, Excite, WebCrawler, InfoSeek, Lycos, and HotBot. Notice
36 | anything missing? Google was incorporated that September.We supported online transactions via a company called
37 | Cybercash,
38 | since if we lacked that feature we'd have gotten beaten up in product
39 | comparisons. But Cybercash was so bad and most stores' order volumes
40 | were so low that it was better if merchants processed orders like phone orders. We had a page in our site trying to talk merchants
41 | out of doing real time authorizations.The whole site was organized like a funnel, directing people to the
42 | test drive.
43 | It was a novel thing to be able to try out software online. We put
44 | cgi-bin in our dynamic urls to fool competitors about how our
45 | software worked.We had some well
46 | known users. Needless to say, Frederick's of Hollywood got the
47 | most traffic. We charged a flat fee of $300/month for big stores,
48 | so it was a little alarming to have users who got lots of traffic.
49 | I once calculated how much Frederick's was costing us in bandwidth,
50 | and it was about $300/month.Since we hosted all the stores, which together were getting just
51 | over 10 million page views per month in June 1998, we consumed what
52 | at the time seemed a lot of bandwidth. We had 2 T1s (3 Mb/sec)
53 | coming into our offices. In those days there was no AWS. Even
54 | colocating servers seemed too risky, considering how often things
55 | went wrong with them. So we had our servers in our offices. Or
56 | more precisely, in Trevor's office. In return for the unique
57 | privilege of sharing his office with no other humans, he had to
58 | share it with 6 shrieking tower servers. His office was nicknamed
59 | the Hot Tub on account of the heat they generated. Most days his
60 | stack of window air conditioners could keep up.For describing pages, we had a template language called RTML, which
61 | supposedly stood for something, but which in fact I named after
62 | Rtm. RTML was Common Lisp augmented by some macros and libraries,
63 | and concealed under a structure editor that made it look like it
64 | had syntax.Since we did continuous releases, our software didn't actually have
65 | versions. But in those days the trade press expected versions, so
66 | we made them up. If we wanted to get lots of attention, we made
67 | the version number an
68 | integer. That "version 4.0" icon was generated by our own
69 | button generator, incidentally. The whole Viaweb site was made
70 | with our software, even though it wasn't an online store, because
71 | we wanted to experience what our users did.At the end of 1997, we released a general purpose shopping search
72 | engine called Shopfind. It
73 | was pretty advanced for the time. It had a programmable crawler
74 | that could crawl most of the different stores online and pick out
75 | the products.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/copy.txt:
--------------------------------------------------------------------------------
1 | July 2006
2 | When I was in high school I spent a lot of time imitating bad
3 | writers. What we studied in English classes was mostly fiction,
4 | so I assumed that was the highest form of writing. Mistake number
5 | one. The stories that seemed to be most admired were ones in which
6 | people suffered in complicated ways. Anything funny or
7 | gripping was ipso facto suspect, unless it was old enough to be hard to
8 | understand, like Shakespeare or Chaucer. Mistake number two. The
9 | ideal medium seemed the short story, which I've since learned had
10 | quite a brief life, roughly coincident with the peak of magazine
11 | publishing. But since their size made them perfect for use in
12 | high school classes, we read a lot of them, which gave us the
13 | impression the short story was flourishing. Mistake number three.
14 | And because they were so short, nothing really had to happen; you
15 | could just show a randomly truncated slice of life, and that was
16 | considered advanced. Mistake number four. The result was that I
17 | wrote a lot of stories in which nothing happened except that someone
18 | was unhappy in a way that seemed deep.For most of college I was a philosophy major. I was very impressed
19 | by the papers published in philosophy journals. They were so
20 | beautifully typeset, and their tone was just captivating—alternately
21 | casual and buffer-overflowingly technical. A fellow would be walking
22 | along a street and suddenly modality qua modality would spring upon
23 | him. I didn't ever quite understand these papers, but I figured
24 | I'd get around to that later, when I had time to reread them more
25 | closely. In the meantime I tried my best to imitate them. This
26 | was, I can now see, a doomed undertaking, because they weren't
27 | really saying anything. No philosopher ever refuted another, for
28 | example, because no one said anything definite enough to refute.
29 | Needless to say, my imitations didn't say anything either.In grad school I was still wasting time imitating the wrong things.
30 | There was then a fashionable type of program called an expert system,
31 | at the core of which was something called an inference engine. I
32 | looked at what these things did and thought "I could write that in
33 | a thousand lines of code." And yet eminent professors were writing
34 | books about them, and startups were selling them for a year's salary
35 | a copy. What an opportunity, I thought; these impressive things
36 | seem easy to me; I must be pretty sharp. Wrong. It was simply a
37 | fad. The books the professors wrote about expert systems are now
38 | ignored. They were not even on a path to anything interesting.
39 | And the customers paying so much for them were largely the same
40 | government agencies that paid thousands for screwdrivers and toilet
41 | seats.How do you avoid copying the wrong things? Copy only what you
42 | genuinely like. That would have saved me in all three cases. I
43 | didn't enjoy the short stories we had to read in English classes;
44 | I didn't learn anything from philosophy papers; I didn't use expert
45 | systems myself. I believed these things were good because they
46 | were admired.It can be hard to separate the things you like from the things
47 | you're impressed with. One trick is to ignore presentation. Whenever
48 | I see a painting impressively hung in a museum, I ask myself: how
49 | much would I pay for this if I found it at a garage sale, dirty and
50 | frameless, and with no idea who painted it? If you walk around a
51 | museum trying this experiment, you'll find you get some truly
52 | startling results. Don't ignore this data point just because it's
53 | an outlier.Another way to figure out what you like is to look at what you enjoy
54 | as guilty pleasures. Many things people like, especially if they're
55 | young and ambitious, they like largely for the feeling of virtue
56 | in liking them. 99% of people reading Ulysses are thinking
57 | "I'm reading Ulysses" as they do it. A guilty pleasure is
58 | at least a pure one. What do you read when you don't feel up to being
59 | virtuous? What kind of book do you read and feel sad that there's
60 | only half of it left, instead of being impressed that you're half
61 | way through? That's what you really like.Even when you find genuinely good things to copy, there's another
62 | pitfall to be avoided. Be careful to copy what makes them good,
63 | rather than their flaws. It's easy to be drawn into imitating
64 | flaws, because they're easier to see, and of course easier to copy
65 | too. For example, most painters in the eighteenth and nineteenth
66 | centuries used brownish colors. They were imitating the great
67 | painters of the Renaissance, whose paintings by that time were brown
68 | with dirt. Those paintings have since been cleaned, revealing
69 | brilliant colors; their imitators are of course still brown.It was painting, incidentally, that cured me of copying the wrong
70 | things. Halfway through grad school I decided I wanted to try being
71 | a painter, and the art world was so manifestly corrupt that it
72 | snapped the leash of credulity. These people made philosophy
73 | professors seem as scrupulous as mathematicians. It was so clearly
74 | a choice of doing good work xor being an insider that I was forced
75 | to see the distinction. It's there to some degree in almost every
76 | field, but I had till then managed to avoid facing it.That was one of the most valuable things I learned from painting:
77 | you have to figure out for yourself what's
78 | good. You can't trust
79 | authorities. They'll lie to you on this one.
80 |
81 | Comment on this essay.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/goodtaste.txt:
--------------------------------------------------------------------------------
1 | November 2021(This essay is derived from a talk at the Cambridge Union.)When I was a kid, I'd have said there wasn't. My father told me so.
2 | Some people like some things, and other people like other things,
3 | and who's to say who's right?It seemed so obvious that there was no such thing as good taste
4 | that it was only through indirect evidence that I realized my father
5 | was wrong. And that's what I'm going to give you here: a proof by
6 | reductio ad absurdum. If we start from the premise that there's no
7 | such thing as good taste, we end up with conclusions that are
8 | obviously false, and therefore the premise must be wrong.We'd better start by saying what good taste is. There's a narrow
9 | sense in which it refers to aesthetic judgements and a broader one
10 | in which it refers to preferences of any kind. The strongest proof
11 | would be to show that taste exists in the narrowest sense, so I'm
12 | going to talk about taste in art. You have better taste than me if
13 | the art you like is better than the art I like.If there's no such thing as good taste, then there's no such thing
14 | as good art. Because if there is such a
15 | thing as good art, it's
16 | easy to tell which of two people has better taste. Show them a lot
17 | of works by artists they've never seen before and ask them to
18 | choose the best, and whoever chooses the better art has better
19 | taste.So if you want to discard the concept of good taste, you also have
20 | to discard the concept of good art. And that means you have to
21 | discard the possibility of people being good at making it. Which
22 | means there's no way for artists to be good at their jobs. And not
23 | just visual artists, but anyone who is in any sense an artist. You
24 | can't have good actors, or novelists, or composers, or dancers
25 | either. You can have popular novelists, but not good ones.We don't realize how far we'd have to go if we discarded the concept
26 | of good taste, because we don't even debate the most obvious cases.
27 | But it doesn't just mean we can't say which of two famous painters
28 | is better. It means we can't say that any painter is better than a
29 | randomly chosen eight year old.That was how I realized my father was wrong. I started studying
30 | painting. And it was just like other kinds of work I'd done: you
31 | could do it well, or badly, and if you tried hard, you could get
32 | better at it. And it was obvious that Leonardo and Bellini were
33 | much better at it than me. That gap between us was not imaginary.
34 | They were so good. And if they could be good, then art could be
35 | good, and there was such a thing as good taste after all.Now that I've explained how to show there is such a thing as good
36 | taste, I should also explain why people think there isn't. There
37 | are two reasons. One is that there's always so much disagreement
38 | about taste. Most people's response to art is a tangle of unexamined
39 | impulses. Is the artist famous? Is the subject attractive? Is this
40 | the sort of art they're supposed to like? Is it hanging in a famous
41 | museum, or reproduced in a big, expensive book? In practice most
42 | people's response to art is dominated by such extraneous factors.And the people who do claim to have good taste are so often mistaken.
43 | The paintings admired by the so-called experts in one generation
44 | are often so different from those admired a few generations later.
45 | It's easy to conclude there's nothing real there at all. It's only
46 | when you isolate this force, for example by trying to paint and
47 | comparing your work to Bellini's, that you can see that it does in
48 | fact exist.The other reason people doubt that art can be good is that there
49 | doesn't seem to be any room in the art for this goodness. The
50 | argument goes like this. Imagine several people looking at a work
51 | of art and judging how good it is. If being good art really is a
52 | property of objects, it should be in the object somehow. But it
53 | doesn't seem to be; it seems to be something happening in the heads
54 | of each of the observers. And if they disagree, how do you choose
55 | between them?The solution to this puzzle is to realize that the purpose of art
56 | is to work on its human audience, and humans have a lot in common.
57 | And to the extent the things an object acts upon respond in the
58 | same way, that's arguably what it means for the object to have the
59 | corresponding property. If everything a particle interacts with
60 | behaves as if the particle had a mass of m, then it has a mass of
61 | m. So the distinction between "objective" and "subjective" is not
62 | binary, but a matter of degree, depending on how much the subjects
63 | have in common. Particles interacting with one another are at one
64 | pole, but people interacting with art are not all the way at the
65 | other; their reactions aren't random.Because people's responses to art aren't random, art can be designed
66 | to operate on people, and be good or bad depending on how effectively
67 | it does so. Much as a vaccine can be. If someone were talking about
68 | the ability of a vaccine to confer immunity, it would seem very
69 | frivolous to object that conferring immunity wasn't really a property
70 | of vaccines, because acquiring immunity is something that happens
71 | in the immune system of each individual person. Sure, people's
72 | immune systems vary, and a vaccine that worked on one might not
73 | work on another, but that doesn't make it meaningless to talk about
74 | the effectiveness of a vaccine.The situation with art is messier, of course. You can't measure
75 | effectiveness by simply taking a vote, as you do with vaccines.
76 | You have to imagine the responses of subjects with a deep knowledge
77 | of art, and enough clarity of mind to be able to ignore extraneous
78 | influences like the fame of the artist. And even then you'd still
79 | see some disagreement. People do vary, and judging art is hard,
80 | especially recent art. There is definitely not a total order either
81 | of works or of people's ability to judge them. But there is equally
82 | definitely a partial order of both. So while it's not possible to
83 | have perfect taste, it is possible to have good taste.
84 | Thanks to the Cambridge Union for inviting me, and to Trevor
85 | Blackwell, Jessica Livingston, and Robert Morris for reading drafts
86 | of this.
87 |
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/ecw.txt:
--------------------------------------------------------------------------------
1 | December 2014If the world were static, we could have monotonically increasing
2 | confidence in our beliefs. The more (and more varied) experience
3 | a belief survived, the less likely it would be false. Most people
4 | implicitly believe something like this about their opinions. And
5 | they're justified in doing so with opinions about things that don't
6 | change much, like human nature. But you can't trust your opinions
7 | in the same way about things that change, which could include
8 | practically everything else.When experts are wrong, it's often because they're experts on an
9 | earlier version of the world.Is it possible to avoid that? Can you protect yourself against
10 | obsolete beliefs? To some extent, yes. I spent almost a decade
11 | investing in early stage startups, and curiously enough protecting
12 | yourself against obsolete beliefs is exactly what you have to do
13 | to succeed as a startup investor. Most really good startup ideas
14 | look like bad ideas at first, and many of those look bad specifically
15 | because some change in the world just switched them from bad to
16 | good. I spent a lot of time learning to recognize such ideas, and
17 | the techniques I used may be applicable to ideas in general.The first step is to have an explicit belief in change. People who
18 | fall victim to a monotonically increasing confidence in their
19 | opinions are implicitly concluding the world is static. If you
20 | consciously remind yourself it isn't, you start to look for change.Where should one look for it? Beyond the moderately useful
21 | generalization that human nature doesn't change much, the unfortunate
22 | fact is that change is hard to predict. This is largely a tautology
23 | but worth remembering all the same: change that matters usually
24 | comes from an unforeseen quarter.So I don't even try to predict it. When I get asked in interviews
25 | to predict the future, I always have to struggle to come up with
26 | something plausible-sounding on the fly, like a student who hasn't
27 | prepared for an exam.
28 | [1]
29 | But it's not out of laziness that I haven't
30 | prepared. It seems to me that beliefs about the future are so
31 | rarely correct that they usually aren't worth the extra rigidity
32 | they impose, and that the best strategy is simply to be aggressively
33 | open-minded. Instead of trying to point yourself in the right
34 | direction, admit you have no idea what the right direction is, and
35 | try instead to be super sensitive to the winds of change.It's ok to have working hypotheses, even though they may constrain
36 | you a bit, because they also motivate you. It's exciting to chase
37 | things and exciting to try to guess answers. But you have to be
38 | disciplined about not letting your hypotheses harden into anything
39 | more.
40 | [2]I believe this passive m.o. works not just for evaluating new ideas
41 | but also for having them. The way to come up with new ideas is not
42 | to try explicitly to, but to try to solve problems and simply not
43 | discount weird hunches you have in the process.The winds of change originate in the unconscious minds of domain
44 | experts. If you're sufficiently expert in a field, any weird idea
45 | or apparently irrelevant question that occurs to you is ipso facto
46 | worth exploring.
47 | [3]
48 | Within Y Combinator, when an idea is described
49 | as crazy, it's a compliment—in fact, on average probably a
50 | higher compliment than when an idea is described as good.Startup investors have extraordinary incentives for correcting
51 | obsolete beliefs. If they can realize before other investors that
52 | some apparently unpromising startup isn't, they can make a huge
53 | amount of money. But the incentives are more than just financial.
54 | Investors' opinions are explicitly tested: startups come to them
55 | and they have to say yes or no, and then, fairly quickly, they learn
56 | whether they guessed right. The investors who say no to a Google
57 | (and there were several) will remember it for the rest of their
58 | lives.Anyone who must in some sense bet on ideas rather than merely
59 | commenting on them has similar incentives. Which means anyone who
60 | wants such incentives can have them, by turning their comments into
61 | bets: if you write about a topic in some fairly durable and public
62 | form, you'll find you worry much more about getting things right
63 | than most people would in a casual conversation.
64 | [4]Another trick I've found to protect myself against obsolete beliefs
65 | is to focus initially on people rather than ideas. Though the nature
66 | of future discoveries is hard to predict, I've found I can predict
67 | quite well what sort of people will make them. Good new ideas come
68 | from earnest, energetic, independent-minded people.Betting on people over ideas saved me countless times as an investor.
69 | We thought Airbnb was a bad idea, for example. But we could tell
70 | the founders were earnest, energetic, and independent-minded.
71 | (Indeed, almost pathologically so.) So we suspended disbelief and
72 | funded them.This too seems a technique that should be generally applicable.
73 | Surround yourself with the sort of people new ideas come from. If
74 | you want to notice quickly when your beliefs become obsolete, you
75 | can't do better than to be friends with the people whose discoveries
76 | will make them so.It's hard enough already not to become the prisoner of your own
77 | expertise, but it will only get harder, because change is accelerating.
78 | That's not a recent trend; change has been accelerating since the
79 | paleolithic era. Ideas beget ideas. I don't expect that to change.
80 | But I could be wrong.
81 | Notes[1]
82 | My usual trick is to talk about aspects of the present that
83 | most people haven't noticed yet.[2]
84 | Especially if they become well enough known that people start
85 | to identify them with you. You have to be extra skeptical about
86 | things you want to believe, and once a hypothesis starts to be
87 | identified with you, it will almost certainly start to be in that
88 | category.[3]
89 | In practice "sufficiently expert" doesn't require one to be
90 | recognized as an expert—which is a trailing indicator in any
91 | case. In many fields a year of focused work plus caring a lot would
92 | be enough.[4]
93 | Though they are public and persist indefinitely, comments on
94 | e.g. forums and places like Twitter seem empirically to work like
95 | casual conversation. The threshold may be whether what you write
96 | has a title.
97 | Thanks to Sam Altman, Patrick Collison, and Robert Morris
98 | for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/corpdev.txt:
--------------------------------------------------------------------------------
1 | January 2015Corporate Development, aka corp dev, is the group within companies
2 | that buys other companies. If you're talking to someone from corp
3 | dev, that's why, whether you realize it yet or not.It's usually a mistake to talk to corp dev unless (a) you want to
4 | sell your company right now and (b) you're sufficiently likely to
5 | get an offer at an acceptable price. In practice that means startups
6 | should only talk to corp dev when they're either doing really well
7 | or really badly. If you're doing really badly, meaning the company
8 | is about to die, you may as well talk to them, because you have
9 | nothing to lose. And if you're doing really well, you can safely
10 | talk to them, because you both know the price will have to be high,
11 | and if they show the slightest sign of wasting your time, you'll
12 | be confident enough to tell them to get lost.The danger is to companies in the middle. Particularly to young
13 | companies that are growing fast, but haven't been doing it for long
14 | enough to have grown big yet. It's usually a mistake for a promising
15 | company less than a year old even to talk to corp dev.But it's a mistake founders constantly make. When someone from
16 | corp dev wants to meet, the founders tell themselves they should
17 | at least find out what they want. Besides, they don't want to
18 | offend Big Company by refusing to meet.Well, I'll tell you what they want. They want to talk about buying
19 | you. That's what the title "corp dev" means. So before agreeing
20 | to meet with someone from corp dev, ask yourselves, "Do we want to
21 | sell the company right now?" And if the answer is no, tell them
22 | "Sorry, but we're focusing on growing the company." They won't be
23 | offended. And certainly the founders of Big Company won't be
24 | offended. If anything they'll think more highly of you. You'll
25 | remind them of themselves. They didn't sell either; that's why
26 | they're in a position now to buy other companies.
27 | [1]Most founders who get contacted by corp dev already know what it
28 | means. And yet even when they know what corp dev does and know
29 | they don't want to sell, they take the meeting. Why do they do it?
30 | The same mix of denial and wishful thinking that underlies most
31 | mistakes founders make. It's flattering to talk to someone who wants
32 | to buy you. And who knows, maybe their offer will be surprisingly
33 | high. You should at least see what it is, right?No. If they were going to send you an offer immediately by email,
34 | sure, you might as well open it. But that is not how conversations
35 | with corp dev work. If you get an offer at all, it will be at the
36 | end of a long and unbelievably distracting process. And if the
37 | offer is surprising, it will be surprisingly low.Distractions are the thing you can least afford in a startup. And
38 | conversations with corp dev are the worst sort of distraction,
39 | because as well as consuming your attention they undermine your
40 | morale. One of the tricks to surviving a grueling process is not
41 | to stop and think how tired you are. Instead you get into a sort
42 | of flow.
43 | [2]
44 | Imagine what it would do to you if at mile 20 of a
45 | marathon, someone ran up beside you and said "You must feel really
46 | tired. Would you like to stop and take a rest?" Conversations
47 | with corp dev are like that but worse, because the suggestion of
48 | stopping gets combined in your mind with the imaginary high price
49 | you think they'll offer.And then you're really in trouble. If they can, corp dev people
50 | like to turn the tables on you. They like to get you to the point
51 | where you're trying to convince them to buy instead of them trying
52 | to convince you to sell. And surprisingly often they succeed.This is a very slippery slope, greased with some of the most powerful
53 | forces that can work on founders' minds, and attended by an experienced
54 | professional whose full time job is to push you down it.Their tactics in pushing you down that slope are usually fairly
55 | brutal. Corp dev people's whole job is to buy companies, and they
56 | don't even get to choose which. The only way their performance is
57 | measured is by how cheaply they can buy you, and the more ambitious
58 | ones will stop at nothing to achieve that. For example, they'll
59 | almost always start with a lowball offer, just to see if you'll
60 | take it. Even if you don't, a low initial offer will demoralize you
61 | and make you easier to manipulate.And that is the most innocent of their tactics. Just wait till
62 | you've agreed on a price and think you have a done deal, and then
63 | they come back and say their boss has vetoed the deal and won't do
64 | it for more than half the agreed upon price. Happens all the time.
65 | If you think investors can behave badly, it's nothing compared to
66 | what corp dev people can do. Even corp dev people at companies
67 | that are otherwise benevolent.I remember once complaining to a
68 | friend at Google about some nasty trick their corp dev people had
69 | pulled on a YC startup."What happened to Don't be Evil?" I asked."I don't think corp dev got the memo," he replied.The tactics you encounter in M&A conversations can be like nothing
70 | you've experienced in the otherwise comparatively
71 | upstanding world
72 | of Silicon Valley. It's as if a chunk of genetic material from the
73 | old-fashioned robber baron business world got incorporated into the
74 | startup world.
75 | [3]The simplest way to protect yourself is to use the trick that John
76 | D. Rockefeller, whose grandfather was an alcoholic, used to protect
77 | himself from becoming one. He once told a Sunday school class
78 |
79 | Boys, do you know why I never became a drunkard? Because I never
80 | took the first drink.
81 |
82 | Do you want to sell your company right now? Not eventually, right
83 | now. If not, just don't take the first meeting. They won't be
84 | offended. And you in turn will be guaranteed to be spared one of
85 | the worst experiences that can happen to a startup.If you do want to sell, there's another set of
86 | techniques
87 | for doing
88 | that. But the biggest mistake founders make in dealing with corp
89 | dev is not doing a bad job of talking to them when they're ready
90 | to, but talking to them before they are. So if you remember only
91 | the title of this essay, you already know most of what you need to
92 | know about M&A in the first year.Notes[1]
93 | I'm not saying you should never sell. I'm saying you should
94 | be clear in your own mind about whether you want to sell or not,
95 | and not be led by manipulation or wishful thinking into trying to
96 | sell earlier than you otherwise would have.[2]
97 | In a startup, as in most competitive sports, the task at hand
98 | almost does this for you; you're too busy to feel tired. But when
99 | you lose that protection, e.g. at the final whistle, the fatigue
100 | hits you like a wave. To talk to corp dev is to let yourself feel
101 | it mid-game.[3]
102 | To be fair, the apparent misdeeds of corp dev people are magnified
103 | by the fact that they function as the face of a large organization
104 | that often doesn't know its own mind. Acquirers can be surprisingly
105 | indecisive about acquisitions, and their flakiness is indistinguishable
106 | from dishonesty by the time it filters down to you.Thanks to Marc Andreessen, Jessica Livingston, Geoff
107 | Ralston, and Qasar Younis for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/addiction.txt:
--------------------------------------------------------------------------------
1 | July 2010What hard liquor, cigarettes, heroin, and crack have in common is
2 | that they're all more concentrated forms of less addictive predecessors.
3 | Most if not all the things we describe as addictive are. And the
4 | scary thing is, the process that created them is accelerating.We wouldn't want to stop it. It's the same process that cures
5 | diseases: technological progress. Technological progress means
6 | making things do more of what we want. When the thing we want is
7 | something we want to want, we consider technological progress good.
8 | If some new technique makes solar cells x% more efficient, that
9 | seems strictly better. When progress concentrates something we
10 | don't want to want—when it transforms opium into heroin—it seems
11 | bad. But it's the same process at work.
12 | [1]No one doubts this process is accelerating, which means increasing
13 | numbers of things we like will be transformed into things we like
14 | too much.
15 | [2]As far as I know there's no word for something we like too much.
16 | The closest is the colloquial sense of "addictive." That usage has
17 | become increasingly common during my lifetime. And it's clear why:
18 | there are an increasing number of things we need it for. At the
19 | extreme end of the spectrum are crack and meth. Food has been
20 | transformed by a combination of factory farming and innovations in
21 | food processing into something with way more immediate bang for the
22 | buck, and you can see the results in any town in America. Checkers
23 | and solitaire have been replaced by World of Warcraft and FarmVille.
24 | TV has become much more engaging, and even so it can't compete with Facebook.The world is more addictive than it was 40 years ago. And unless
25 | the forms of technological progress that produced these things are
26 | subject to different laws than technological progress in general,
27 | the world will get more addictive in the next 40 years than it did
28 | in the last 40.The next 40 years will bring us some wonderful things. I don't
29 | mean to imply they're all to be avoided. Alcohol is a dangerous
30 | drug, but I'd rather live in a world with wine than one without.
31 | Most people can coexist with alcohol; but you have to be careful.
32 | More things we like will mean more things we have to be careful
33 | about.Most people won't, unfortunately. Which means that as the world
34 | becomes more addictive, the two senses in which one can live a
35 | normal life will be driven ever further apart. One sense of "normal"
36 | is statistically normal: what everyone else does. The other is the
37 | sense we mean when we talk about the normal operating range of a
38 | piece of machinery: what works best.These two senses are already quite far apart. Already someone
39 | trying to live well would seem eccentrically abstemious in most of
40 | the US. That phenomenon is only going to become more pronounced.
41 | You can probably take it as a rule of thumb from now on that if
42 | people don't think you're weird, you're living badly.Societies eventually develop antibodies to addictive new things.
43 | I've seen that happen with cigarettes. When cigarettes first
44 | appeared, they spread the way an infectious disease spreads through
45 | a previously isolated population. Smoking rapidly became a
46 | (statistically) normal thing. There were ashtrays everywhere. We
47 | had ashtrays in our house when I was a kid, even though neither of
48 | my parents smoked. You had to for guests.As knowledge spread about the dangers of smoking, customs changed.
49 | In the last 20 years, smoking has been transformed from something
50 | that seemed totally normal into a rather seedy habit: from something
51 | movie stars did in publicity shots to something small huddles of
52 | addicts do outside the doors of office buildings. A lot of the
53 | change was due to legislation, of course, but the legislation
54 | couldn't have happened if customs hadn't already changed.It took a while though—on the order of 100 years. And unless the
55 | rate at which social antibodies evolve can increase to match the
56 | accelerating rate at which technological progress throws off new
57 | addictions, we'll be increasingly unable to rely on customs to
58 | protect us.
59 | [3]
60 | Unless we want to be canaries in the coal mine
61 | of each new addiction—the people whose sad example becomes a
62 | lesson to future generations—we'll have to figure out for ourselves
63 | what to avoid and how. It will actually become a reasonable strategy
64 | (or a more reasonable strategy) to suspect
65 | everything new.In fact, even that won't be enough. We'll have to worry not just
66 | about new things, but also about existing things becoming more
67 | addictive. That's what bit me. I've avoided most addictions, but
68 | the Internet got me because it became addictive while I was using
69 | it.
70 | [4]Most people I know have problems with Internet addiction. We're
71 | all trying to figure out our own customs for getting free of it.
72 | That's why I don't have an iPhone, for example; the last thing I
73 | want is for the Internet to follow me out into the world.
74 | [5]
75 | My latest trick is taking long hikes. I used to think running was a
76 | better form of exercise than hiking because it took less time. Now
77 | the slowness of hiking seems an advantage, because the longer I
78 | spend on the trail, the longer I have to think without interruption.Sounds pretty eccentric, doesn't it? It always will when you're
79 | trying to solve problems where there are no customs yet to guide
80 | you. Maybe I can't plead Occam's razor; maybe I'm simply eccentric.
81 | But if I'm right about the acceleration of addictiveness, then this
82 | kind of lonely squirming to avoid it will increasingly be the fate
83 | of anyone who wants to get things done. We'll increasingly be
84 | defined by what we say no to.
85 | Notes[1]
86 | Could you restrict technological progress to areas where you
87 | wanted it? Only in a limited way, without becoming a police state.
88 | And even then your restrictions would have undesirable side effects.
89 | "Good" and "bad" technological progress aren't sharply differentiated,
90 | so you'd find you couldn't slow the latter without also slowing the
91 | former. And in any case, as Prohibition and the "war on drugs"
92 | show, bans often do more harm than good.[2]
93 | Technology has always been accelerating. By Paleolithic
94 | standards, technology evolved at a blistering pace in the Neolithic
95 | period.[3]
96 | Unless we mass produce social customs. I suspect the recent
97 | resurgence of evangelical Christianity in the US is partly a reaction
98 | to drugs. In desperation people reach for the sledgehammer; if
99 | their kids won't listen to them, maybe they'll listen to God. But
100 | that solution has broader consequences than just getting kids to
101 | say no to drugs. You end up saying no to
102 | science as well.
103 | I worry we may be heading for a future in which only a few people
104 | plot their own itinerary through no-land, while everyone else books
105 | a package tour. Or worse still, has one booked for them by the
106 | government.[4]
107 | People commonly use the word "procrastination" to describe
108 | what they do on the Internet. It seems to me too mild to describe
109 | what's happening as merely not-doing-work. We don't call it
110 | procrastination when someone gets drunk instead of working.[5]
111 | Several people have told me they like the iPad because it
112 | lets them bring the Internet into situations where a laptop would
113 | be too conspicuous. In other words, it's a hip flask. (This is
114 | true of the iPhone too, of course, but this advantage isn't as
115 | obvious because it reads as a phone, and everyone's used to those.)Thanks to Sam Altman, Patrick Collison, Jessica Livingston, and
116 | Robert Morris for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/newideas.txt:
--------------------------------------------------------------------------------
1 | May 2021There's one kind of opinion I'd be very afraid to express publicly.
2 | If someone I knew to be both a domain expert and a reasonable person
3 | proposed an idea that sounded preposterous, I'd be very reluctant
4 | to say "That will never work."Anyone who has studied the history of ideas, and especially the
5 | history of science, knows that's how big things start. Someone
6 | proposes an idea that sounds crazy, most people dismiss it, then
7 | it gradually takes over the world.Most implausible-sounding ideas are in fact bad and could be safely
8 | dismissed. But not when they're proposed by reasonable domain
9 | experts. If the person proposing the idea is reasonable, then they
10 | know how implausible it sounds. And yet they're proposing it anyway.
11 | That suggests they know something you don't. And if they have deep
12 | domain expertise, that's probably the source of it.
13 | [1]Such ideas are not merely unsafe to dismiss, but disproportionately
14 | likely to be interesting. When the average person proposes an
15 | implausible-sounding idea, its implausibility is evidence of their
16 | incompetence. But when a reasonable domain expert does it, the
17 | situation is reversed. There's something like an efficient market
18 | here: on average the ideas that seem craziest will, if correct,
19 | have the biggest effect. So if you can eliminate the theory that
20 | the person proposing an implausible-sounding idea is incompetent,
21 | its implausibility switches from evidence that it's boring to
22 | evidence that it's exciting.
23 | [2]Such ideas are not guaranteed to work. But they don't have to be.
24 | They just have to be sufficiently good bets — to have sufficiently
25 | high expected value. And I think on average they do. I think if you
26 | bet on the entire set of implausible-sounding ideas proposed by
27 | reasonable domain experts, you'd end up net ahead.The reason is that everyone is too conservative. The word "paradigm"
28 | is overused, but this is a case where it's warranted. Everyone is
29 | too much in the grip of the current paradigm. Even the people who
30 | have the new ideas undervalue them initially. Which means that
31 | before they reach the stage of proposing them publicly, they've
32 | already subjected them to an excessively strict filter.
33 | [3]The wise response to such an idea is not to make statements, but
34 | to ask questions, because there's a real mystery here. Why has this
35 | smart and reasonable person proposed an idea that seems so wrong?
36 | Are they mistaken, or are you? One of you has to be. If you're the
37 | one who's mistaken, that would be good to know, because it means
38 | there's a hole in your model of the world. But even if they're
39 | mistaken, it should be interesting to learn why. A trap that an
40 | expert falls into is one you have to worry about too.This all seems pretty obvious. And yet there are clearly a lot of
41 | people who don't share my fear of dismissing new ideas. Why do they
42 | do it? Why risk looking like a jerk now and a fool later, instead
43 | of just reserving judgement?One reason they do it is envy. If you propose a radical new idea
44 | and it succeeds, your reputation (and perhaps also your wealth)
45 | will increase proportionally. Some people would be envious if that
46 | happened, and this potential envy propagates back into a conviction
47 | that you must be wrong.Another reason people dismiss new ideas is that it's an easy way
48 | to seem sophisticated. When a new idea first emerges, it usually
49 | seems pretty feeble. It's a mere hatchling. Received wisdom is a
50 | full-grown eagle by comparison. So it's easy to launch a devastating
51 | attack on a new idea, and anyone who does will seem clever to those
52 | who don't understand this asymmetry.This phenomenon is exacerbated by the difference between how those
53 | working on new ideas and those attacking them are rewarded. The
54 | rewards for working on new ideas are weighted by the value of the
55 | outcome. So it's worth working on something that only has a 10%
56 | chance of succeeding if it would make things more than 10x better.
57 | Whereas the rewards for attacking new ideas are roughly constant;
58 | such attacks seem roughly equally clever regardless of the target.People will also attack new ideas when they have a vested interest
59 | in the old ones. It's not surprising, for example, that some of
60 | Darwin's harshest critics were churchmen. People build whole careers
61 | on some ideas. When someone claims they're false or obsolete, they
62 | feel threatened.The lowest form of dismissal is mere factionalism: to automatically
63 | dismiss any idea associated with the opposing faction. The lowest
64 | form of all is to dismiss an idea because of who proposed it.But the main thing that leads reasonable people to dismiss new ideas
65 | is the same thing that holds people back from proposing them: the
66 | sheer pervasiveness of the current paradigm. It doesn't just affect
67 | the way we think; it is the Lego blocks we build thoughts out of.
68 | Popping out of the current paradigm is something only a few people
69 | can do. And even they usually have to suppress their intuitions at
70 | first, like a pilot flying through cloud who has to trust his
71 | instruments over his sense of balance.
72 | [4]Paradigms don't just define our present thinking. They also vacuum
73 | up the trail of crumbs that led to them, making our standards for
74 | new ideas impossibly high. The current paradigm seems so perfect
75 | to us, its offspring, that we imagine it must have been accepted
76 | completely as soon as it was discovered — that whatever the church thought
77 | of the heliocentric model, astronomers must have been convinced as
78 | soon as Copernicus proposed it. Far, in fact, from it. Copernicus
79 | published the heliocentric model in 1532, but it wasn't till the
80 | mid seventeenth century that the balance of scientific opinion
81 | shifted in its favor.
82 | [5]Few understand how feeble new ideas look when they first appear.
83 | So if you want to have new ideas yourself, one of the most valuable
84 | things you can do is to learn what they look like when they're born.
85 | Read about how new ideas happened, and try to get yourself into the
86 | heads of people at the time. How did things look to them, when the
87 | new idea was only half-finished, and even the person who had it was
88 | only half-convinced it was right?But you don't have to stop at history. You can observe big new ideas
89 | being born all around you right now. Just look for a reasonable
90 | domain expert proposing something that sounds wrong.If you're nice, as well as wise, you won't merely resist attacking
91 | such people, but encourage them. Having new ideas is a lonely
92 | business. Only those who've tried it know how lonely. These people
93 | need your help. And if you help them, you'll probably learn something
94 | in the process.Notes[1]
95 | This domain expertise could be in another field. Indeed,
96 | such crossovers tend to be particularly promising.[2]
97 | I'm not claiming this principle extends much beyond math,
98 | engineering, and the hard sciences. In politics, for example,
99 | crazy-sounding ideas generally are as bad as they sound. Though
100 | arguably this is not an exception, because the people who propose
101 | them are not in fact domain experts; politicians are domain experts
102 | in political tactics, like how to get elected and how to get
103 | legislation passed, but not in the world that policy acts upon.
104 | Perhaps no one could be.[3]
105 | This sense of "paradigm" was defined by Thomas Kuhn in his
106 | Structure of Scientific Revolutions, but I also recommend his
107 | Copernican Revolution, where you can see him at work developing the
108 | idea.[4]
109 | This is one reason people with a touch of Asperger's may have
110 | an advantage in discovering new ideas. They're always flying on
111 | instruments.[5]
112 | Hall, Rupert. From Galileo to Newton. Collins, 1963. This
113 | book is particularly good at getting into contemporaries' heads.Thanks to Trevor Blackwell, Patrick Collison, Suhail Doshi, Daniel
114 | Gackle, Jessica Livingston, and Robert Morris for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/aord.txt:
--------------------------------------------------------------------------------
1 | October 2015When I talk to a startup that's been operating for more than 8 or
2 | 9 months, the first thing I want to know is almost always the same.
3 | Assuming their expenses remain constant and their revenue growth
4 | is what it has been over the last several months, do they make it to
5 | profitability on the money they have left? Or to put it more
6 | dramatically, by default do they live or die?The startling thing is how often the founders themselves don't know.
7 | Half the founders I talk to don't know whether they're default alive
8 | or default dead.If you're among that number, Trevor Blackwell has made a handy
9 | calculator you can use to find out.The reason I want to know first whether a startup is default alive
10 | or default dead is that the rest of the conversation depends on the
11 | answer. If the company is default alive, we can talk about ambitious
12 | new things they could do. If it's default dead, we probably need
13 | to talk about how to save it. We know the current trajectory ends
14 | badly. How can they get off that trajectory?Why do so few founders know whether they're default alive or default
15 | dead? Mainly, I think, because they're not used to asking that.
16 | It's not a question that makes sense to ask early on, any more than
17 | it makes sense to ask a 3 year old how he plans to support
18 | himself. But as the company grows older, the question switches from
19 | meaningless to critical. That kind of switch often takes people
20 | by surprise.I propose the following solution: instead of starting to ask too
21 | late whether you're default alive or default dead, start asking too
22 | early. It's hard to say precisely when the question switches
23 | polarity. But it's probably not that dangerous to start worrying
24 | too early that you're default dead, whereas it's very dangerous to
25 | start worrying too late.The reason is a phenomenon I wrote about earlier: the
26 | fatal pinch.
27 | The fatal pinch is default dead + slow growth + not enough
28 | time to fix it. And the way founders end up in it is by not realizing
29 | that's where they're headed.There is another reason founders don't ask themselves whether they're
30 | default alive or default dead: they assume it will be easy to raise
31 | more money. But that assumption is often false, and worse still, the
32 | more you depend on it, the falser it becomes.Maybe it will help to separate facts from hopes. Instead of thinking
33 | of the future with vague optimism, explicitly separate the components.
34 | Say "We're default dead, but we're counting on investors to save
35 | us." Maybe as you say that, it will set off the same alarms in your
36 | head that it does in mine. And if you set off the alarms sufficiently
37 | early, you may be able to avoid the fatal pinch.It would be safe to be default dead if you could count on investors
38 | saving you. As a rule their interest is a function of
39 | growth. If you have steep revenue growth, say over 5x a year, you
40 | can start to count on investors being interested even if you're not
41 | profitable.
42 | [1]
43 | But investors are so fickle that you can never
44 | do more than start to count on them. Sometimes something about your
45 | business will spook investors even if your growth is great. So no
46 | matter how good your growth is, you can never safely treat fundraising
47 | as more than a plan A. You should always have a plan B as well: you
48 | should know (as in write down) precisely what you'll need to do to
49 | survive if you can't raise more money, and precisely when you'll
50 | have to switch to plan B if plan A isn't working.In any case, growing fast versus operating cheaply is far from the
51 | sharp dichotomy many founders assume it to be. In practice there
52 | is surprisingly little connection between how much a startup spends
53 | and how fast it grows. When a startup grows fast, it's usually
54 | because the product hits a nerve, in the sense of hitting some big
55 | need straight on. When a startup spends a lot, it's usually because
56 | the product is expensive to develop or sell, or simply because
57 | they're wasteful.If you're paying attention, you'll be asking at this point not just
58 | how to avoid the fatal pinch, but how to avoid being default dead.
59 | That one is easy: don't hire too fast. Hiring too fast is by far
60 | the biggest killer of startups that raise money.
61 | [2]Founders tell themselves they need to hire in order to grow. But
62 | most err on the side of overestimating this need rather than
63 | underestimating it. Why? Partly because there's so much work to
64 | do. Naive founders think that if they can just hire enough
65 | people, it will all get done. Partly because successful startups have
66 | lots of employees, so it seems like that's what one does in order
67 | to be successful. In fact the large staffs of successful startups
68 | are probably more the effect of growth than the cause. And
69 | partly because when founders have slow growth they don't want to
70 | face what is usually the real reason: the product is not appealing
71 | enough.Plus founders who've just raised money are often encouraged to
72 | overhire by the VCs who funded them. Kill-or-cure strategies are
73 | optimal for VCs because they're protected by the portfolio effect.
74 | VCs want to blow you up, in one sense of the phrase or the other.
75 | But as a founder your incentives are different. You want above all
76 | to survive.
77 | [3]Here's a common way startups die. They make something moderately
78 | appealing and have decent initial growth. They raise their first
79 | round fairly easily, because the founders seem smart and the idea
80 | sounds plausible. But because the product is only moderately
81 | appealing, growth is ok but not great. The founders convince
82 | themselves that hiring a bunch of people is the way to boost growth.
83 | Their investors agree. But (because the product is only moderately
84 | appealing) the growth never comes. Now they're rapidly running out
85 | of runway. They hope further investment will save them. But because
86 | they have high expenses and slow growth, they're now unappealing
87 | to investors. They're unable to raise more, and the company dies.What the company should have done is address the fundamental problem:
88 | that the product is only moderately appealing. Hiring people is
89 | rarely the way to fix that. More often than not it makes it harder.
90 | At this early stage, the product needs to evolve more than to be
91 | "built out," and that's usually easier with fewer people.
92 | [4]Asking whether you're default alive or default dead may save you
93 | from this. Maybe the alarm bells it sets off will counteract the
94 | forces that push you to overhire. Instead you'll be compelled to
95 | seek growth in other ways. For example, by doing
96 | things that don't scale, or by redesigning the product in the
97 | way only founders can.
98 | And for many if not most startups, these paths to growth will be
99 | the ones that actually work.Airbnb waited 4 months after raising money at the end of Y Combinator
100 | before they hired their first employee. In the meantime the founders
101 | were terribly overworked. But they were overworked evolving Airbnb
102 | into the astonishingly successful organism it is now.Notes[1]
103 | Steep usage growth will also interest investors. Revenue
104 | will ultimately be a constant multiple of usage, so x% usage growth
105 | predicts x% revenue growth. But in practice investors discount
106 | merely predicted revenue, so if you're measuring usage you need a
107 | higher growth rate to impress investors.[2]
108 | Startups that don't raise money are saved from hiring too
109 | fast because they can't afford to. But that doesn't mean you should
110 | avoid raising money in order to avoid this problem, any more than
111 | that total abstinence is the only way to avoid becoming an alcoholic.[3]
112 | I would not be surprised if VCs' tendency to push founders
113 | to overhire is not even in their own interest. They don't know how
114 | many of the companies that get killed by overspending might have
115 | done well if they'd survived. My guess is a significant number.[4]
116 | After reading a draft, Sam Altman wrote:"I think you should make the hiring point more strongly. I think
117 | it's roughly correct to say that YC's most successful companies
118 | have never been the fastest to hire, and one of the marks of a great
119 | founder is being able to resist this urge."Paul Buchheit adds:"A related problem that I see a lot is premature scaling—founders
120 | take a small business that isn't really working (bad unit economics,
121 | typically) and then scale it up because they want impressive growth
122 | numbers. This is similar to over-hiring in that it makes the business
123 | much harder to fix once it's big, plus they are bleeding cash really
124 | fast."
125 | Thanks to Sam Altman, Paul Buchheit, Joe Gebbia, Jessica Livingston,
126 | and Geoff Ralston for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/vcsqueeze.txt:
--------------------------------------------------------------------------------
1 | November 2005In the next few years, venture capital funds will find themselves
2 | squeezed from four directions. They're already stuck with a seller's
3 | market, because of the huge amounts they raised at the end of the
4 | Bubble and still haven't invested. This by itself is not the end
5 | of the world. In fact, it's just a more extreme version of the
6 | norm
7 | in the VC business: too much money chasing too few deals.Unfortunately, those few deals now want less and less money, because
8 | it's getting so cheap to start a startup. The four causes: open
9 | source, which makes software free; Moore's law, which makes hardware
10 | geometrically closer to free; the Web, which makes promotion free
11 | if you're good; and better languages, which make development a lot
12 | cheaper.When we started our startup in 1995, the first three were our biggest
13 | expenses. We had to pay $5000 for the Netscape Commerce Server,
14 | the only software that then supported secure http connections. We
15 | paid $3000 for a server with a 90 MHz processor and 32 meg of
16 | memory. And we paid a PR firm about $30,000 to promote our launch.Now you could get all three for nothing. You can get the software
17 | for free; people throw away computers more powerful than our first
18 | server; and if you make something good you can generate ten times
19 | as much traffic by word of mouth online than our first PR firm got
20 | through the print media.And of course another big change for the average startup is that
21 | programming languages have improved-- or rather, the median language has. At most startups ten years
22 | ago, software development meant ten programmers writing code in
23 | C++. Now the same work might be done by one or two using Python
24 | or Ruby.During the Bubble, a lot of people predicted that startups would
25 | outsource their development to India. I think a better model for
26 | the future is David Heinemeier Hansson, who outsourced his development
27 | to a more powerful language instead. A lot of well-known applications
28 | are now, like BaseCamp, written by just one programmer. And one
29 | guy is more than 10x cheaper than ten, because (a) he won't waste
30 | any time in meetings, and (b) since he's probably a founder, he can
31 | pay himself nothing.Because starting a startup is so cheap, venture capitalists now
32 | often want to give startups more money than the startups want to
33 | take. VCs like to invest several million at a time. But as one
34 | VC told me after a startup he funded would only take about half a
35 | million, "I don't know what we're going to do. Maybe we'll just
36 | have to give some of it back." Meaning give some of the fund back
37 | to the institutional investors who supplied it, because it wasn't
38 | going to be possible to invest it all.Into this already bad situation comes the third problem: Sarbanes-Oxley.
39 | Sarbanes-Oxley is a law, passed after the Bubble, that drastically
40 | increases the regulatory burden on public companies. And in addition
41 | to the cost of compliance, which is at least two million dollars a
42 | year, the law introduces frightening legal exposure for corporate
43 | officers. An experienced CFO I know said flatly: "I would not
44 | want to be CFO of a public company now."You might think that responsible corporate governance is an area
45 | where you can't go too far. But you can go too far in any law, and
46 | this remark convinced me that Sarbanes-Oxley must have. This CFO
47 | is both the smartest and the most upstanding money guy I know. If
48 | Sarbanes-Oxley deters people like him from being CFOs of public
49 | companies, that's proof enough that it's broken.Largely because of Sarbanes-Oxley, few startups go public now. For
50 | all practical purposes, succeeding now equals getting bought. Which
51 | means VCs are now in the business of finding promising little 2-3
52 | man startups and pumping them up into companies that cost $100
53 | million to acquire. They didn't mean to be in this business; it's
54 | just what their business has evolved into.Hence the fourth problem: the acquirers have begun to realize they
55 | can buy wholesale. Why should they wait for VCs to make the startups
56 | they want more expensive? Most of what the VCs add, acquirers don't
57 | want anyway. The acquirers already have brand recognition and HR
58 | departments. What they really want is the software and the developers,
59 | and that's what the startup is in the early phase: concentrated
60 | software and developers.Google, typically, seems to have been the first to figure this out.
61 | "Bring us your startups early," said Google's speaker at the Startup School. They're quite
62 | explicit about it: they like to acquire startups at just the point
63 | where they would do a Series A round. (The Series A round is the
64 | first round of real VC funding; it usually happens in the first
65 | year.) It is a brilliant strategy, and one that other big technology
66 | companies will no doubt try to duplicate. Unless they want to have
67 | still more of their lunch eaten by Google.Of course, Google has an advantage in buying startups: a lot of the
68 | people there are rich, or expect to be when their options vest.
69 | Ordinary employees find it very hard to recommend an acquisition;
70 | it's just too annoying to see a bunch of twenty year olds get rich
71 | when you're still working for salary. Even if it's the right thing
72 | for your company to do.The Solution(s)Bad as things look now, there is a way for VCs to save themselves.
73 | They need to do two things, one of which won't surprise them, and
74 | another that will seem an anathema.Let's start with the obvious one: lobby to get Sarbanes-Oxley
75 | loosened. This law was created to prevent future Enrons, not to
76 | destroy the IPO market. Since the IPO market was practically dead
77 | when it passed, few saw what bad effects it would have. But now
78 | that technology has recovered from the last bust, we can see clearly
79 | what a bottleneck Sarbanes-Oxley has become.Startups are fragile plants—seedlings, in fact. These seedlings
80 | are worth protecting, because they grow into the trees of the
81 | economy. Much of the economy's growth is their growth. I think
82 | most politicians realize that. But they don't realize just how
83 | fragile startups are, and how easily they can become collateral
84 | damage of laws meant to fix some other problem.Still more dangerously, when you destroy startups, they make very
85 | little noise. If you step on the toes of the coal industry, you'll
86 | hear about it. But if you inadvertantly squash the startup industry,
87 | all that happens is that the founders of the next Google stay in
88 | grad school instead of starting a company.My second suggestion will seem shocking to VCs: let founders cash
89 | out partially in the Series A round. At the moment, when VCs invest
90 | in a startup, all the stock they get is newly issued and all the
91 | money goes to the company. They could buy some stock directly from
92 | the founders as well.Most VCs have an almost religious rule against doing this. They
93 | don't want founders to get a penny till the company is sold or goes
94 | public. VCs are obsessed with control, and they worry that they'll
95 | have less leverage over the founders if the founders have any money.This is a dumb plan. In fact, letting the founders sell a little stock
96 | early would generally be better for the company, because it would
97 | cause the founders' attitudes toward risk to be aligned with the
98 | VCs'. As things currently work, their attitudes toward risk tend
99 | to be diametrically opposed: the founders, who have nothing, would
100 | prefer a 100% chance of $1 million to a 20% chance of $10 million,
101 | while the VCs can afford to be "rational" and prefer the latter.Whatever they say, the reason founders are selling their companies
102 | early instead of doing Series A rounds is that they get paid up
103 | front. That first million is just worth so much more than the
104 | subsequent ones. If founders could sell a little stock early,
105 | they'd be happy to take VC money and bet the rest on a bigger
106 | outcome.So why not let the founders have that first million, or at least
107 | half million? The VCs would get same number of shares for the
108 | money. So what if some of the money would go to the
109 | founders instead of the company?Some VCs will say this is
110 | unthinkable—that they want all their money to be put to work
111 | growing the company. But the fact is, the huge size of current VC
112 | investments is dictated by the structure
113 | of VC funds, not the needs of startups. Often as not these large
114 | investments go to work destroying the company rather than growing
115 | it.The angel investors who funded our startup let the founders sell
116 | some stock directly to them, and it was a good deal for everyone.
117 | The angels made a huge return on that investment, so they're happy.
118 | And for us founders it blunted the terrifying all-or-nothingness
119 | of a startup, which in its raw form is more a distraction than a
120 | motivator.If VCs are frightened at the idea of letting founders partially
121 | cash out, let me tell them something still more frightening: you
122 | are now competing directly with Google.
123 | Thanks to Trevor Blackwell, Sarah Harlin, Jessica
124 | Livingston, and Robert Morris for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/vb.txt:
--------------------------------------------------------------------------------
1 | January 2016Life is short, as everyone knows. When I was a kid I used to wonder
2 | about this. Is life actually short, or are we really complaining
3 | about its finiteness? Would we be just as likely to feel life was
4 | short if we lived 10 times as long?Since there didn't seem any way to answer this question, I stopped
5 | wondering about it. Then I had kids. That gave me a way to answer
6 | the question, and the answer is that life actually is short.Having kids showed me how to convert a continuous quantity, time,
7 | into discrete quantities. You only get 52 weekends with your 2 year
8 | old. If Christmas-as-magic lasts from say ages 3 to 10, you only
9 | get to watch your child experience it 8 times. And while it's
10 | impossible to say what is a lot or a little of a continuous quantity
11 | like time, 8 is not a lot of something. If you had a handful of 8
12 | peanuts, or a shelf of 8 books to choose from, the quantity would
13 | definitely seem limited, no matter what your lifespan was.Ok, so life actually is short. Does it make any difference to know
14 | that?It has for me. It means arguments of the form "Life is too short
15 | for x" have great force. It's not just a figure of speech to say
16 | that life is too short for something. It's not just a synonym for
17 | annoying. If you find yourself thinking that life is too short for
18 | something, you should try to eliminate it if you can.When I ask myself what I've found life is too short for, the word
19 | that pops into my head is "bullshit." I realize that answer is
20 | somewhat tautological. It's almost the definition of bullshit that
21 | it's the stuff that life is too short for. And yet bullshit does
22 | have a distinctive character. There's something fake about it.
23 | It's the junk food of experience.
24 | [1]If you ask yourself what you spend your time on that's bullshit,
25 | you probably already know the answer. Unnecessary meetings, pointless
26 | disputes, bureaucracy, posturing, dealing with other people's
27 | mistakes, traffic jams, addictive but unrewarding pastimes.There are two ways this kind of thing gets into your life: it's
28 | either forced on you, or it tricks you. To some extent you have to
29 | put up with the bullshit forced on you by circumstances. You need
30 | to make money, and making money consists mostly of errands. Indeed,
31 | the law of supply and demand insures that: the more rewarding some
32 | kind of work is, the cheaper people will do it. It may be that
33 | less bullshit is forced on you than you think, though. There has
34 | always been a stream of people who opt out of the default grind and
35 | go live somewhere where opportunities are fewer in the conventional
36 | sense, but life feels more authentic. This could become more common.You can do it on a smaller scale without moving. The amount of
37 | time you have to spend on bullshit varies between employers. Most
38 | large organizations (and many small ones) are steeped in it. But
39 | if you consciously prioritize bullshit avoidance over other factors
40 | like money and prestige, you can probably find employers that will
41 | waste less of your time.If you're a freelancer or a small company, you can do this at the
42 | level of individual customers. If you fire or avoid toxic customers,
43 | you can decrease the amount of bullshit in your life by more than
44 | you decrease your income.But while some amount of bullshit is inevitably forced on you, the
45 | bullshit that sneaks into your life by tricking you is no one's
46 | fault but your own. And yet the bullshit you choose may be harder
47 | to eliminate than the bullshit that's forced on you. Things that
48 | lure you into wasting your time have to be really good at
49 | tricking you. An example that will be familiar to a lot of people
50 | is arguing online. When someone
51 | contradicts you, they're in a sense attacking you. Sometimes pretty
52 | overtly. Your instinct when attacked is to defend yourself. But
53 | like a lot of instincts, this one wasn't designed for the world we
54 | now live in. Counterintuitive as it feels, it's better most of
55 | the time not to defend yourself. Otherwise these people are literally
56 | taking your life.
57 | [2]Arguing online is only incidentally addictive. There are more
58 | dangerous things than that. As I've written before, one byproduct
59 | of technical progress is that things we like tend to become more
60 | addictive. Which means we will increasingly have to make a conscious
61 | effort to avoid addictions to stand outside ourselves and ask "is
62 | this how I want to be spending my time?"As well as avoiding bullshit, one should actively seek out things
63 | that matter. But different things matter to different people, and
64 | most have to learn what matters to them. A few are lucky and realize
65 | early on that they love math or taking care of animals or writing,
66 | and then figure out a way to spend a lot of time doing it. But
67 | most people start out with a life that's a mix of things that
68 | matter and things that don't, and only gradually learn to distinguish
69 | between them.For the young especially, much of this confusion is induced by the
70 | artificial situations they find themselves in. In middle school and
71 | high school, what the other kids think of you seems the most important
72 | thing in the world. But when you ask adults what they got wrong
73 | at that age, nearly all say they cared too much what other kids
74 | thought of them.One heuristic for distinguishing stuff that matters is to ask
75 | yourself whether you'll care about it in the future. Fake stuff
76 | that matters usually has a sharp peak of seeming to matter. That's
77 | how it tricks you. The area under the curve is small, but its shape
78 | jabs into your consciousness like a pin.The things that matter aren't necessarily the ones people would
79 | call "important." Having coffee with a friend matters. You won't
80 | feel later like that was a waste of time.One great thing about having small children is that they make you
81 | spend time on things that matter: them. They grab your sleeve as
82 | you're staring at your phone and say "will you play with me?" And
83 | odds are that is in fact the bullshit-minimizing option.If life is short, we should expect its shortness to take us by
84 | surprise. And that is just what tends to happen. You take things
85 | for granted, and then they're gone. You think you can always write
86 | that book, or climb that mountain, or whatever, and then you realize
87 | the window has closed. The saddest windows close when other people
88 | die. Their lives are short too. After my mother died, I wished I'd
89 | spent more time with her. I lived as if she'd always be there.
90 | And in her typical quiet way she encouraged that illusion. But an
91 | illusion it was. I think a lot of people make the same mistake I
92 | did.The usual way to avoid being taken by surprise by something is to
93 | be consciously aware of it. Back when life was more precarious,
94 | people used to be aware of death to a degree that would now seem a
95 | bit morbid. I'm not sure why, but it doesn't seem the right answer
96 | to be constantly reminding oneself of the grim reaper hovering at
97 | everyone's shoulder. Perhaps a better solution is to look at the
98 | problem from the other end. Cultivate a habit of impatience about
99 | the things you most want to do. Don't wait before climbing that
100 | mountain or writing that book or visiting your mother. You don't
101 | need to be constantly reminding yourself why you shouldn't wait.
102 | Just don't wait.I can think of two more things one does when one doesn't have much
103 | of something: try to get more of it, and savor what one has. Both
104 | make sense here.How you live affects how long you live. Most people could do better.
105 | Me among them.But you can probably get even more effect by paying closer attention
106 | to the time you have. It's easy to let the days rush by. The
107 | "flow" that imaginative people love so much has a darker cousin
108 | that prevents you from pausing to savor life amid the daily slurry
109 | of errands and alarms. One of the most striking things I've read
110 | was not in a book, but the title of one: James Salter's Burning
111 | the Days.It is possible to slow time somewhat. I've gotten better at it.
112 | Kids help. When you have small children, there are a lot of moments
113 | so perfect that you can't help noticing.It does help too to feel that you've squeezed everything out of
114 | some experience. The reason I'm sad about my mother is not just
115 | that I miss her but that I think of all the things we could have
116 | done that we didn't. My oldest son will be 7 soon. And while I
117 | miss the 3 year old version of him, I at least don't have any regrets
118 | over what might have been. We had the best time a daddy and a 3
119 | year old ever had.Relentlessly prune bullshit, don't wait to do things that matter,
120 | and savor the time you have. That's what you do when life is short.Notes[1]
121 | At first I didn't like it that the word that came to mind was
122 | one that had other meanings. But then I realized the other meanings
123 | are fairly closely related. Bullshit in the sense of things you
124 | waste your time on is a lot like intellectual bullshit.[2]
125 | I chose this example deliberately as a note to self. I get
126 | attacked a lot online. People tell the craziest lies about me.
127 | And I have so far done a pretty mediocre job of suppressing the
128 | natural human inclination to say "Hey, that's not true!"Thanks to Jessica Livingston and Geoff Ralston for reading drafts
129 | of this.
--------------------------------------------------------------------------------
/context_data/a_stars.txt:
--------------------------------------------------------------------------------
1 | {'32': [15, 117, 42, 69, 58, 107, 9, 49, 113, 66, 26, 102, 81, 94, 77, 61, 5, 19, 109, 3, 35, 54, 86, 89, 127, 21, 46, 122, 38, 97, 74, 29], '64': [15, 117, 42, 69, 58, 107, 9, 49, 113, 66, 26, 102, 81, 94, 77, 61, 5, 19, 109, 3, 35, 54, 86, 89, 127, 21, 46, 122, 38, 97, 74, 29, 153, 167, 255, 198, 139, 201, 185, 230, 194, 171, 211, 175, 162, 130, 239, 213, 223, 158, 178, 243, 235, 190, 217, 227, 145, 183, 249, 206, 245, 150, 135, 142], '128': [15, 117, 42, 69, 58, 107, 9, 49, 113, 66, 26, 102, 81, 94, 77, 61, 5, 19, 109, 3, 35, 54, 86, 89, 127, 21, 46, 122, 38, 97, 74, 29, 153, 167, 255, 198, 139, 201, 185, 230, 194, 171, 211, 175, 162, 130, 239, 213, 223, 158, 178, 243, 235, 190, 217, 227, 145, 183, 249, 206, 245, 150, 135, 142, 333, 471, 305, 509, 418, 461, 371, 358, 373, 437, 382, 263, 342, 291, 365, 429, 479, 421, 483, 491, 393, 407, 386, 442, 294, 270, 273, 486, 499, 346, 446, 301, 287, 467, 391, 323, 379, 474, 402, 266, 409, 413, 502, 329, 494, 454, 397, 283, 338, 313, 319, 353, 299, 457, 310, 427, 327, 279, 449, 505, 434, 361, 257, 351], '256': [15, 117, 42, 69, 58, 107, 9, 49, 113, 66, 26, 102, 81, 94, 77, 61, 5, 19, 109, 3, 35, 54, 86, 89, 127, 21, 46, 122, 38, 97, 74, 29, 153, 167, 255, 198, 139, 201, 185, 230, 194, 171, 211, 175, 162, 130, 239, 213, 223, 158, 178, 243, 235, 190, 217, 227, 145, 183, 249, 206, 245, 150, 135, 142, 333, 471, 305, 509, 418, 461, 371, 358, 373, 437, 382, 263, 342, 291, 365, 429, 479, 421, 483, 491, 393, 407, 386, 442, 294, 270, 273, 486, 499, 346, 446, 301, 287, 467, 391, 323, 379, 474, 402, 266, 409, 413, 502, 329, 494, 454, 397, 283, 338, 313, 319, 353, 299, 457, 310, 427, 327, 279, 449, 505, 434, 361, 257, 351, 534, 639, 723, 730, 606, 794, 653, 701, 1019, 690, 909, 698, 625, 922, 567, 834, 859, 683, 790, 710, 913, 633, 950, 927, 957, 677, 735, 773, 555, 558, 546, 609, 523, 894, 837, 759, 665, 934, 539, 879, 853, 765, 975, 782, 642, 530, 526, 662, 749, 619, 818, 581, 843, 1007, 901, 823, 814, 623, 890, 693, 649, 851, 741, 946, 570, 942, 514, 613, 1014, 755, 1021, 551, 873, 715, 962, 517, 631, 669, 847, 898, 562, 883, 829, 717, 739, 746, 870, 983, 937, 799, 1002, 1011, 991, 589, 955, 658, 885, 726, 578, 761, 997, 987, 686, 994, 967, 778, 598, 978, 809, 785, 907, 705, 807, 586, 575, 594, 543, 862, 674, 770, 827, 601, 802, 867, 645, 929, 919, 970], '512': [15, 117, 42, 69, 58, 107, 9, 49, 113, 66, 26, 102, 81, 94, 77, 61, 5, 19, 109, 3, 35, 54, 86, 89, 127, 21, 46, 122, 38, 97, 74, 29, 153, 167, 255, 198, 139, 201, 185, 230, 194, 171, 211, 175, 162, 130, 239, 213, 223, 158, 178, 243, 235, 190, 217, 227, 145, 183, 249, 206, 245, 150, 135, 142, 333, 471, 305, 509, 418, 461, 371, 358, 373, 437, 382, 263, 342, 291, 365, 429, 479, 421, 483, 491, 393, 407, 386, 442, 294, 270, 273, 486, 499, 346, 446, 301, 287, 467, 391, 323, 379, 474, 402, 266, 409, 413, 502, 329, 494, 454, 397, 283, 338, 313, 319, 353, 299, 457, 310, 427, 327, 279, 449, 505, 434, 361, 257, 351, 534, 639, 723, 730, 606, 794, 653, 701, 1019, 690, 909, 698, 625, 922, 567, 834, 859, 683, 790, 710, 913, 633, 950, 927, 957, 677, 735, 773, 555, 558, 546, 609, 523, 894, 837, 759, 665, 934, 539, 879, 853, 765, 975, 782, 642, 530, 526, 662, 749, 619, 818, 581, 843, 1007, 901, 823, 814, 623, 890, 693, 649, 851, 741, 946, 570, 942, 514, 613, 1014, 755, 1021, 551, 873, 715, 962, 517, 631, 669, 847, 898, 562, 883, 829, 717, 739, 746, 870, 983, 937, 799, 1002, 1011, 991, 589, 955, 658, 885, 726, 578, 761, 997, 987, 686, 994, 967, 778, 598, 978, 809, 785, 907, 705, 807, 586, 575, 594, 543, 862, 674, 770, 827, 601, 802, 867, 645, 929, 919, 970, 1451, 1695, 1666, 1301, 1374, 1650, 1991, 1357, 1149, 1854, 1075, 1617, 2025, 1519, 1466, 1169, 1423, 1877, 1850, 1318, 1533, 1382, 1526, 1523, 1482, 1297, 1961, 1658, 1341, 1439, 1789, 1102, 1163, 1906, 1061, 1718, 1983, 1614, 2037, 1743, 1281, 1747, 1831, 1110, 1165, 1425, 1346, 1185, 1871, 1965, 1811, 1114, 1974, 1270, 2006, 2033, 1573, 1579, 1402, 1607, 1565, 1734, 1755, 1306, 1334, 1602, 1570, 1757, 1953, 1937, 1090, 1394, 2045, 1883, 1933, 1979, 1630, 1379, 1327, 1637, 1471, 1886, 2015, 1231, 1899, 1691, 1626, 1082, 1909, 1587, 1819, 1645, 1686, 1267, 1857, 1414, 1723, 1246, 1473, 1554, 2017, 1293, 1714, 1155, 1861, 1046, 1350, 1773, 1418, 1986, 1361, 1025, 1727, 1643, 1314, 1286, 1031, 1398, 1609, 1729, 1366, 1498, 1998, 1842, 1458, 1485, 1041, 1529, 1146, 1059, 1502, 1387, 1681, 1454, 1507, 1793, 1223, 1738, 1177, 1837, 1310, 1433, 1289, 1958, 1930, 1561, 1181, 1322, 1927, 1633, 1338, 1038, 1443, 1710, 1873, 1767, 1277, 1599, 2001, 1189, 1699, 1129, 1203, 1943, 1805, 1158, 1463, 1195, 1782, 1133, 1826, 1430, 1763, 2011, 1099, 1551, 1771, 1513, 1593, 1078, 1950, 1918, 2029, 1670, 1213, 1331, 1257, 1262, 1119, 1581, 1479, 1087, 1209, 1547, 1901, 1054, 1242, 1494, 1970, 1275, 1894, 1923, 1846, 1410, 1217, 1447, 1706, 1590, 1542, 1511, 2021, 1914, 1355, 1623, 1391, 1677, 1095, 1785, 1173, 1197, 1750, 1815, 1371, 1889, 1702, 1834, 1675, 1779, 1803, 1995, 1405, 1205, 1137, 1238, 1253, 1227, 1661, 1122, 1050, 1033, 1865, 1823, 1947, 1491, 1067, 1799, 1141, 1655, 1235, 1557, 1251, 1125, 1070, 1106, 2041, 1537], '1024': [15, 117, 42, 69, 58, 107, 9, 49, 113, 66, 26, 102, 81, 94, 77, 61, 5, 19, 109, 3, 35, 54, 86, 89, 127, 21, 46, 122, 38, 97, 74, 29, 153, 167, 255, 198, 139, 201, 185, 230, 194, 171, 211, 175, 162, 130, 239, 213, 223, 158, 178, 243, 235, 190, 217, 227, 145, 183, 249, 206, 245, 150, 135, 142, 333, 471, 305, 509, 418, 461, 371, 358, 373, 437, 382, 263, 342, 291, 365, 429, 479, 421, 483, 491, 393, 407, 386, 442, 294, 270, 273, 486, 499, 346, 446, 301, 287, 467, 391, 323, 379, 474, 402, 266, 409, 413, 502, 329, 494, 454, 397, 283, 338, 313, 319, 353, 299, 457, 310, 427, 327, 279, 449, 505, 434, 361, 257, 351, 534, 639, 723, 730, 606, 794, 653, 701, 1019, 690, 909, 698, 625, 922, 567, 834, 859, 683, 790, 710, 913, 633, 950, 927, 957, 677, 735, 773, 555, 558, 546, 609, 523, 894, 837, 759, 665, 934, 539, 879, 853, 765, 975, 782, 642, 530, 526, 662, 749, 619, 818, 581, 843, 1007, 901, 823, 814, 623, 890, 693, 649, 851, 741, 946, 570, 942, 514, 613, 1014, 755, 1021, 551, 873, 715, 962, 517, 631, 669, 847, 898, 562, 883, 829, 717, 739, 746, 870, 983, 937, 799, 1002, 1011, 991, 589, 955, 658, 885, 726, 578, 761, 997, 987, 686, 994, 967, 778, 598, 978, 809, 785, 907, 705, 807, 586, 575, 594, 543, 862, 674, 770, 827, 601, 802, 867, 645, 929, 919, 970, 1451, 1695, 1666, 1301, 1374, 1650, 1991, 1357, 1149, 1854, 1075, 1617, 2025, 1519, 1466, 1169, 1423, 1877, 1850, 1318, 1533, 1382, 1526, 1523, 1482, 1297, 1961, 1658, 1341, 1439, 1789, 1102, 1163, 1906, 1061, 1718, 1983, 1614, 2037, 1743, 1281, 1747, 1831, 1110, 1165, 1425, 1346, 1185, 1871, 1965, 1811, 1114, 1974, 1270, 2006, 2033, 1573, 1579, 1402, 1607, 1565, 1734, 1755, 1306, 1334, 1602, 1570, 1757, 1953, 1937, 1090, 1394, 2045, 1883, 1933, 1979, 1630, 1379, 1327, 1637, 1471, 1886, 2015, 1231, 1899, 1691, 1626, 1082, 1909, 1587, 1819, 1645, 1686, 1267, 1857, 1414, 1723, 1246, 1473, 1554, 2017, 1293, 1714, 1155, 1861, 1046, 1350, 1773, 1418, 1986, 1361, 1025, 1727, 1643, 1314, 1286, 1031, 1398, 1609, 1729, 1366, 1498, 1998, 1842, 1458, 1485, 1041, 1529, 1146, 1059, 1502, 1387, 1681, 1454, 1507, 1793, 1223, 1738, 1177, 1837, 1310, 1433, 1289, 1958, 1930, 1561, 1181, 1322, 1927, 1633, 1338, 1038, 1443, 1710, 1873, 1767, 1277, 1599, 2001, 1189, 1699, 1129, 1203, 1943, 1805, 1158, 1463, 1195, 1782, 1133, 1826, 1430, 1763, 2011, 1099, 1551, 1771, 1513, 1593, 1078, 1950, 1918, 2029, 1670, 1213, 1331, 1257, 1262, 1119, 1581, 1479, 1087, 1209, 1547, 1901, 1054, 1242, 1494, 1970, 1275, 1894, 1923, 1846, 1410, 1217, 1447, 1706, 1590, 1542, 1511, 2021, 1914, 1355, 1623, 1391, 1677, 1095, 1785, 1173, 1197, 1750, 1815, 1371, 1889, 1702, 1834, 1675, 1779, 1803, 1995, 1405, 1205, 1137, 1238, 1253, 1227, 1661, 1122, 1050, 1033, 1865, 1823, 1947, 1491, 1067, 1799, 1141, 1655, 1235, 1557, 1251, 1125, 1070, 1106, 2041, 1537, 3369, 3198, 3802, 2471, 2921, 3585, 3619, 3035, 3194, 4075, 3838, 2598, 2467, 2211, 3105, 2909, 2870, 3834, 2310, 2298, 3498, 3461, 2069, 3642, 2691, 3659, 4059, 3779, 2785, 2091, 3394, 2522, 3962, 2323, 3361, 2353, 2451, 3918, 2486, 3030, 3951, 2902, 3549, 2510, 2154, 3161, 4041, 3475, 2727, 3790, 2762, 2191, 2447, 4011, 3537, 2203, 3523, 3053, 3165, 2705, 2074, 2739, 4019, 2413, 2755, 3403, 2949, 3559, 3358, 3047, 4081, 3449, 2389, 3483, 2959, 3622, 4002, 3666, 2669, 2049, 3602, 2663, 2077, 2578, 2338, 3998, 3590, 3017, 3693, 3698, 3286, 2593, 3565, 2237, 3757, 3647, 2307, 2842, 3061, 3281, 2462, 2165, 3097, 2221, 2907, 3322, 3495, 2546, 2559, 3254, 3326, 3606, 3629, 3345, 3953, 2259, 2819, 3701, 2866, 2277, 4067, 2138, 2409, 3249, 2614, 2231, 3957, 2294, 2611, 3685, 3298, 2375, 3893, 4027, 3733, 3842, 2513, 3861, 2607, 2143, 2797, 2530, 3511, 3375, 3237, 2679, 2862, 3967, 2119, 3713, 2262, 3083, 3414, 2085, 3579, 3235, 3246, 2955, 4070, 3306, 2702, 3913, 3535, 2561, 2150, 3885, 2993, 3751, 3905, 2266, 3761, 2333, 3454, 2365, 2395, 3517, 2722, 3937, 3078, 3681, 2205, 4051, 2242, 3881, 3581, 3087, 2683, 3042, 2503, 2130, 3143, 3477, 3158, 2226, 3705, 3389, 2650, 3094, 2621, 2425, 2158, 2991, 2421, 2377, 2897, 4086, 3505, 2251, 2551, 3638, 2134, 3514, 3737, 2385, 2349, 2121, 3597, 2499, 2094, 2913, 3009, 3339, 2457, 2453, 2273, 2765, 2218, 3027, 2987, 2482, 3113, 2971, 3911, 3994, 4007, 3270, 3169, 3845, 3823, 2179, 3137, 3150, 2657, 3650, 3001, 3831, 2098, 2794, 4014, 2937, 3765, 2811, 2851, 3798, 3398, 2183, 3243, 3274, 2065, 3214, 2966, 3873, 2195, 3670, 3718, 2383, 2710, 3555, 2582, 2127, 2555, 3175, 3121, 3351, 2927, 3290, 3826, 3378, 2935, 2255, 2630, 3211, 2997, 3542, 3662, 2646, 2185, 2342, 3690, 3769, 3387, 3975, 2198, 2813, 2878, 3747, 2635, 3773, 2667, 2569, 3178, 2145, 2773, 4054, 3135, 2675, 3111, 3922, 3978, 3731, 2715, 2781, 3901, 2803, 2590, 3727, 2171, 3405, 2734, 3226, 2831, 2745, 2887, 3435, 3309, 3117, 3153, 4029, 2286, 3431, 3038, 3146, 2347, 2405, 3741, 2833, 3722, 3574, 2979, 3437, 2478, 3989, 2494, 3491, 4037, 3470, 2329, 3343, 2945, 2490, 3890, 2102, 3006, 3934, 3783, 3229, 3853, 2214, 3457, 2397, 3205, 3074, 3219, 2361, 2919, 2741, 3570, 3446, 3417, 3982, 3711, 3186, 2083, 3898, 4079, 3130, 3201, 2637, 2431, 2695, 3221, 2617, 2602, 2963, 3070, 2327, 3442, 2642, 2437, 3526, 2517, 2107, 2699, 2174, 2687, 3545, 2845, 3805, 2822, 3927, 3091, 2858, 3813, 2585, 4089, 2805, 3278, 3381, 3679, 2319, 2359, 3022, 3877, 3754, 3485, 2853, 2435, 2771, 3594, 2537, 3293, 4023, 2891, 2247, 2113, 3849, 3969, 4063, 3265, 3502, 3931, 2163, 2789, 2403, 2234, 2269, 3262, 3794, 2826, 3317, 3531, 2109, 3059, 3303, 2063, 3787, 3103, 3366, 2474, 3674, 3191, 3610, 2838, 2370, 3330, 3625, 2730, 3409, 3333, 2942, 2506, 3811, 2315, 2973, 2929, 2882, 3051, 3865, 3946, 2750, 2535, 3015, 3181, 3315, 3353, 2654, 3655, 3259, 3615, 2573, 2874, 2719, 2055, 3817, 2757, 2895, 2543, 3986, 2626, 2059, 2417, 3871, 3422, 3859, 3066, 4046, 3634, 3425, 2981, 2291, 3466, 2301, 3126, 2282, 4034, 3941, 4094, 2527, 2779, 2441, 2567, 3561]}
--------------------------------------------------------------------------------
/context_data/r_stars.txt:
--------------------------------------------------------------------------------
1 | {'32': [16, 116, 43, 70, 59, 106, 8, 48, 112, 67, 25, 101, 82, 93, 76, 62, 6, 18, 108, 4, 34, 53, 85, 90, 126, 22, 45, 121, 39, 96, 75, 30], '64': [16, 116, 43, 70, 59, 106, 8, 48, 112, 67, 25, 101, 82, 93, 76, 62, 6, 18, 108, 4, 34, 53, 85, 90, 126, 22, 45, 121, 39, 96, 75, 30, 154, 168, 256, 197, 140, 200, 184, 229, 195, 172, 210, 174, 163, 131, 238, 214, 222, 159, 177, 242, 236, 189, 216, 228, 146, 184, 248, 207, 244, 151, 136, 143], '128': [16, 116, 43, 70, 59, 106, 8, 48, 112, 67, 25, 101, 82, 93, 76, 62, 6, 18, 108, 4, 34, 53, 85, 90, 126, 22, 45, 121, 39, 96, 75, 30, 154, 168, 256, 197, 140, 200, 184, 229, 195, 172, 210, 174, 163, 131, 238, 214, 222, 159, 177, 242, 236, 189, 216, 228, 146, 184, 248, 207, 244, 151, 136, 143, 332, 472, 306, 510, 417, 462, 370, 359, 374, 438, 381, 262, 343, 290, 364, 430, 478, 422, 484, 492, 394, 406, 385, 441, 293, 271, 272, 487, 500, 345, 445, 302, 286, 466, 392, 324, 378, 473, 403, 265, 410, 412, 501, 328, 493, 453, 398, 284, 339, 312, 318, 354, 300, 458, 309, 426, 328, 280, 448, 506, 433, 360, 256, 352], '256': [16, 116, 43, 70, 59, 106, 8, 48, 112, 67, 25, 101, 82, 93, 76, 62, 6, 18, 108, 4, 34, 53, 85, 90, 126, 22, 45, 121, 39, 96, 75, 30, 154, 168, 256, 197, 140, 200, 184, 229, 195, 172, 210, 174, 163, 131, 238, 214, 222, 159, 177, 242, 236, 189, 216, 228, 146, 184, 248, 207, 244, 151, 136, 143, 332, 472, 306, 510, 417, 462, 370, 359, 374, 438, 381, 262, 343, 290, 364, 430, 478, 422, 484, 492, 394, 406, 385, 441, 293, 271, 272, 487, 500, 345, 445, 302, 286, 466, 392, 324, 378, 473, 403, 265, 410, 412, 501, 328, 493, 453, 398, 284, 339, 312, 318, 354, 300, 458, 309, 426, 328, 280, 448, 506, 433, 360, 256, 352, 533, 638, 724, 729, 605, 795, 652, 700, 1018, 691, 908, 699, 626, 923, 568, 835, 858, 682, 789, 709, 914, 634, 951, 926, 956, 676, 734, 774, 556, 559, 545, 608, 524, 893, 836, 760, 664, 935, 538, 880, 854, 766, 976, 783, 643, 531, 527, 663, 750, 620, 817, 580, 842, 1006, 900, 824, 815, 624, 891, 694, 648, 850, 742, 945, 569, 941, 513, 612, 1015, 754, 1022, 550, 872, 714, 963, 518, 632, 668, 848, 899, 561, 882, 830, 718, 740, 747, 869, 984, 936, 800, 1001, 1010, 990, 590, 956, 659, 886, 725, 577, 760, 998, 988, 685, 995, 966, 779, 599, 979, 808, 786, 908, 704, 806, 587, 574, 593, 544, 863, 675, 771, 828, 602, 803, 866, 644, 928, 918, 971], '512': [16, 116, 43, 70, 59, 106, 8, 48, 112, 67, 25, 101, 82, 93, 76, 62, 6, 18, 108, 4, 34, 53, 85, 90, 126, 22, 45, 121, 39, 96, 75, 30, 154, 168, 256, 197, 140, 200, 184, 229, 195, 172, 210, 174, 163, 131, 238, 214, 222, 159, 177, 242, 236, 189, 216, 228, 146, 184, 248, 207, 244, 151, 136, 143, 332, 472, 306, 510, 417, 462, 370, 359, 374, 438, 381, 262, 343, 290, 364, 430, 478, 422, 484, 492, 394, 406, 385, 441, 293, 271, 272, 487, 500, 345, 445, 302, 286, 466, 392, 324, 378, 473, 403, 265, 410, 412, 501, 328, 493, 453, 398, 284, 339, 312, 318, 354, 300, 458, 309, 426, 328, 280, 448, 506, 433, 360, 256, 352, 533, 638, 724, 729, 605, 795, 652, 700, 1018, 691, 908, 699, 626, 923, 568, 835, 858, 682, 789, 709, 914, 634, 951, 926, 956, 676, 734, 774, 556, 559, 545, 608, 524, 893, 836, 760, 664, 935, 538, 880, 854, 766, 976, 783, 643, 531, 527, 663, 750, 620, 817, 580, 842, 1006, 900, 824, 815, 624, 891, 694, 648, 850, 742, 945, 569, 941, 513, 612, 1015, 754, 1022, 550, 872, 714, 963, 518, 632, 668, 848, 899, 561, 882, 830, 718, 740, 747, 869, 984, 936, 800, 1001, 1010, 990, 590, 956, 659, 886, 725, 577, 760, 998, 988, 685, 995, 966, 779, 599, 979, 808, 786, 908, 704, 806, 587, 574, 593, 544, 863, 675, 771, 828, 602, 803, 866, 644, 928, 918, 971, 1452, 1694, 1665, 1300, 1373, 1651, 1990, 1356, 1150, 1855, 1076, 1616, 2024, 1520, 1467, 1168, 1424, 1876, 1851, 1317, 1534, 1381, 1525, 1522, 1481, 1298, 1962, 1659, 1342, 1440, 1788, 1103, 1162, 1905, 1062, 1717, 1984, 1613, 2038, 1744, 1282, 1746, 1830, 1109, 1166, 1424, 1345, 1184, 1872, 1964, 1812, 1113, 1973, 1271, 2007, 2034, 1572, 1578, 1403, 1608, 1564, 1735, 1754, 1305, 1335, 1601, 1571, 1756, 1952, 1938, 1089, 1393, 2046, 1882, 1934, 1978, 1629, 1380, 1328, 1636, 1472, 1887, 2014, 1230, 1898, 1690, 1627, 1081, 1908, 1588, 1818, 1644, 1685, 1266, 1858, 1415, 1724, 1247, 1474, 1555, 2016, 1292, 1713, 1154, 1860, 1047, 1351, 1774, 1419, 1985, 1360, 1024, 1726, 1644, 1313, 1285, 1030, 1399, 1610, 1730, 1367, 1497, 1997, 1841, 1457, 1486, 1042, 1528, 1145, 1058, 1503, 1386, 1680, 1453, 1506, 1792, 1222, 1739, 1178, 1836, 1309, 1432, 1290, 1957, 1931, 1560, 1182, 1321, 1928, 1632, 1339, 1037, 1444, 1709, 1872, 1768, 1278, 1598, 2000, 1190, 1698, 1128, 1204, 1944, 1804, 1157, 1464, 1194, 1781, 1134, 1827, 1429, 1762, 2010, 1100, 1552, 1772, 1512, 1592, 1077, 1951, 1919, 2030, 1669, 1214, 1332, 1258, 1263, 1118, 1580, 1480, 1088, 1210, 1548, 1900, 1055, 1243, 1495, 1969, 1276, 1893, 1922, 1845, 1409, 1216, 1446, 1707, 1589, 1543, 1510, 2020, 1913, 1354, 1624, 1390, 1678, 1096, 1784, 1172, 1196, 1749, 1814, 1372, 1890, 1703, 1835, 1674, 1778, 1804, 1994, 1406, 1204, 1136, 1237, 1252, 1226, 1660, 1121, 1051, 1034, 1864, 1824, 1948, 1490, 1066, 1798, 1140, 1656, 1234, 1556, 1250, 1126, 1071, 1105, 2042, 1538], '1024': [16, 116, 43, 70, 59, 106, 8, 48, 112, 67, 25, 101, 82, 93, 76, 62, 6, 18, 108, 4, 34, 53, 85, 90, 126, 22, 45, 121, 39, 96, 75, 30, 154, 168, 256, 197, 140, 200, 184, 229, 195, 172, 210, 174, 163, 131, 238, 214, 222, 159, 177, 242, 236, 189, 216, 228, 146, 184, 248, 207, 244, 151, 136, 143, 332, 472, 306, 510, 417, 462, 370, 359, 374, 438, 381, 262, 343, 290, 364, 430, 478, 422, 484, 492, 394, 406, 385, 441, 293, 271, 272, 487, 500, 345, 445, 302, 286, 466, 392, 324, 378, 473, 403, 265, 410, 412, 501, 328, 493, 453, 398, 284, 339, 312, 318, 354, 300, 458, 309, 426, 328, 280, 448, 506, 433, 360, 256, 352, 533, 638, 724, 729, 605, 795, 652, 700, 1018, 691, 908, 699, 626, 923, 568, 835, 858, 682, 789, 709, 914, 634, 951, 926, 956, 676, 734, 774, 556, 559, 545, 608, 524, 893, 836, 760, 664, 935, 538, 880, 854, 766, 976, 783, 643, 531, 527, 663, 750, 620, 817, 580, 842, 1006, 900, 824, 815, 624, 891, 694, 648, 850, 742, 945, 569, 941, 513, 612, 1015, 754, 1022, 550, 872, 714, 963, 518, 632, 668, 848, 899, 561, 882, 830, 718, 740, 747, 869, 984, 936, 800, 1001, 1010, 990, 590, 956, 659, 886, 725, 577, 760, 998, 988, 685, 995, 966, 779, 599, 979, 808, 786, 908, 704, 806, 587, 574, 593, 544, 863, 675, 771, 828, 602, 803, 866, 644, 928, 918, 971, 1452, 1694, 1665, 1300, 1373, 1651, 1990, 1356, 1150, 1855, 1076, 1616, 2024, 1520, 1467, 1168, 1424, 1876, 1851, 1317, 1534, 1381, 1525, 1522, 1481, 1298, 1962, 1659, 1342, 1440, 1788, 1103, 1162, 1905, 1062, 1717, 1984, 1613, 2038, 1744, 1282, 1746, 1830, 1109, 1166, 1424, 1345, 1184, 1872, 1964, 1812, 1113, 1973, 1271, 2007, 2034, 1572, 1578, 1403, 1608, 1564, 1735, 1754, 1305, 1335, 1601, 1571, 1756, 1952, 1938, 1089, 1393, 2046, 1882, 1934, 1978, 1629, 1380, 1328, 1636, 1472, 1887, 2014, 1230, 1898, 1690, 1627, 1081, 1908, 1588, 1818, 1644, 1685, 1266, 1858, 1415, 1724, 1247, 1474, 1555, 2016, 1292, 1713, 1154, 1860, 1047, 1351, 1774, 1419, 1985, 1360, 1024, 1726, 1644, 1313, 1285, 1030, 1399, 1610, 1730, 1367, 1497, 1997, 1841, 1457, 1486, 1042, 1528, 1145, 1058, 1503, 1386, 1680, 1453, 1506, 1792, 1222, 1739, 1178, 1836, 1309, 1432, 1290, 1957, 1931, 1560, 1182, 1321, 1928, 1632, 1339, 1037, 1444, 1709, 1872, 1768, 1278, 1598, 2000, 1190, 1698, 1128, 1204, 1944, 1804, 1157, 1464, 1194, 1781, 1134, 1827, 1429, 1762, 2010, 1100, 1552, 1772, 1512, 1592, 1077, 1951, 1919, 2030, 1669, 1214, 1332, 1258, 1263, 1118, 1580, 1480, 1088, 1210, 1548, 1900, 1055, 1243, 1495, 1969, 1276, 1893, 1922, 1845, 1409, 1216, 1446, 1707, 1589, 1543, 1510, 2020, 1913, 1354, 1624, 1390, 1678, 1096, 1784, 1172, 1196, 1749, 1814, 1372, 1890, 1703, 1835, 1674, 1778, 1804, 1994, 1406, 1204, 1136, 1237, 1252, 1226, 1660, 1121, 1051, 1034, 1864, 1824, 1948, 1490, 1066, 1798, 1140, 1656, 1234, 1556, 1250, 1126, 1071, 1105, 2042, 1538, 3368, 3199, 3803, 2472, 2920, 3584, 3620, 3034, 3195, 4074, 3839, 2599, 2468, 2212, 3104, 2910, 2869, 3833, 2309, 2299, 3499, 3462, 2068, 3643, 2690, 3660, 4060, 3778, 2784, 2092, 3393, 2523, 3963, 2322, 3360, 2354, 2452, 3917, 2485, 3031, 3952, 2901, 3550, 2511, 2155, 3160, 4040, 3474, 2726, 3791, 2761, 2190, 2446, 4012, 3538, 2204, 3524, 3052, 3166, 2706, 2075, 2738, 4018, 2414, 2754, 3402, 2950, 3558, 3357, 3048, 4080, 3450, 2388, 3484, 2960, 3621, 4001, 3665, 2670, 2050, 3601, 2664, 2078, 2579, 2337, 3997, 3591, 3016, 3694, 3699, 3285, 2594, 3564, 2236, 3756, 3648, 2306, 2843, 3060, 3282, 2463, 2164, 3096, 2220, 2908, 3321, 3496, 2545, 2560, 3253, 3325, 3607, 3628, 3346, 3954, 2260, 2820, 3700, 2867, 2278, 4066, 2139, 2408, 3250, 2613, 2232, 3958, 2295, 2610, 3686, 3299, 2376, 3892, 4028, 3732, 3843, 2514, 3862, 2608, 2144, 2796, 2531, 3512, 3376, 3238, 2680, 2861, 3966, 2118, 3712, 2263, 3084, 3413, 2086, 3578, 3234, 3247, 2956, 4071, 3307, 2703, 3912, 3536, 2560, 2151, 3886, 2992, 3750, 3906, 2265, 3762, 2332, 3453, 2364, 2396, 3518, 2721, 3936, 3079, 3680, 2204, 4052, 2243, 3882, 3582, 3086, 2684, 3043, 2502, 2131, 3144, 3478, 3159, 2225, 3704, 3390, 2651, 3095, 2620, 2424, 2159, 2990, 2420, 2378, 2896, 4087, 3506, 2250, 2552, 3639, 2133, 3515, 3736, 2386, 2348, 2120, 3598, 2498, 2093, 2912, 3010, 3338, 2458, 2452, 2274, 2764, 2219, 3026, 2986, 2483, 3112, 2972, 3912, 3993, 4006, 3269, 3168, 3844, 3824, 2178, 3136, 3151, 2658, 3649, 3000, 3832, 2097, 2795, 4013, 2936, 3766, 2810, 2852, 3799, 3399, 2182, 3244, 3273, 2064, 3215, 2965, 3874, 2194, 3671, 3719, 2382, 2709, 3554, 2583, 2126, 2556, 3176, 3120, 3352, 2928, 3291, 3827, 3377, 2934, 2256, 2631, 3212, 2996, 3541, 3661, 2645, 2186, 2341, 3691, 3768, 3386, 3974, 2199, 2812, 2879, 3748, 2634, 3772, 2666, 2570, 3177, 2146, 2772, 4055, 3134, 2674, 3110, 3921, 3979, 3732, 2716, 2782, 3902, 2802, 2591, 3728, 2172, 3406, 2733, 3225, 2830, 2746, 2886, 3434, 3308, 3118, 3154, 4030, 2287, 3430, 3039, 3147, 2348, 2404, 3742, 2832, 3723, 3575, 2980, 3438, 2477, 3990, 2493, 3492, 4038, 3471, 2328, 3344, 2946, 2489, 3889, 2101, 3005, 3933, 3784, 3230, 3854, 2215, 3456, 2396, 3206, 3073, 3220, 2362, 2918, 2742, 3571, 3445, 3418, 3983, 3712, 3187, 2084, 3897, 4078, 3129, 3202, 2638, 2432, 2696, 3222, 2618, 2603, 2964, 3069, 2328, 3443, 2643, 2438, 3527, 2518, 2108, 2698, 2175, 2688, 3546, 2844, 3804, 2821, 3926, 3092, 2859, 3812, 2586, 4088, 2806, 3279, 3382, 3678, 2318, 2358, 3023, 3878, 3755, 3486, 2854, 2436, 2772, 3593, 2536, 3294, 4022, 2892, 2246, 2114, 3850, 3968, 4062, 3266, 3501, 3932, 2164, 2790, 2404, 2235, 2268, 3263, 3793, 2825, 3318, 3530, 2110, 3058, 3302, 2062, 3788, 3102, 3365, 2473, 3675, 3190, 3609, 2839, 2371, 3331, 3626, 2731, 3410, 3332, 2941, 2505, 3810, 2316, 2974, 2928, 2883, 3050, 3864, 3947, 2749, 2534, 3016, 3182, 3316, 3352, 2653, 3656, 3258, 3616, 2574, 2875, 2718, 2054, 3816, 2758, 2896, 2544, 3987, 2627, 2060, 2416, 3872, 3421, 3858, 3065, 4047, 3633, 3424, 2982, 2292, 3467, 2302, 3127, 2283, 4035, 3940, 4093, 2526, 2778, 2442, 2568, 3562]}
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/hubs.txt:
--------------------------------------------------------------------------------
1 |
2 |
3 | Want to start a startup? Get funded by
4 | Y Combinator.
5 |
6 |
7 |
8 |
9 | October 2011If you look at a list of US cities sorted by population, the number
10 | of successful startups per capita varies by orders of magnitude.
11 | Somehow it's as if most places were sprayed with startupicide.I wondered about this for years. I could see the average town was
12 | like a roach motel for startup ambitions: smart, ambitious people
13 | went in, but no startups came out. But I was never able to figure
14 | out exactly what happened inside the motel—exactly what was
15 | killing all the potential startups.
16 | [1]A couple weeks ago I finally figured it out. I was framing the
17 | question wrong. The problem is not that most towns kill startups.
18 | It's that death is the default for startups,
19 | and most towns don't save them. Instead of thinking of most places
20 | as being sprayed with startupicide, it's more accurate to think of
21 | startups as all being poisoned, and a few places being sprayed with
22 | the antidote.Startups in other places are just doing what startups naturally do:
23 | fail. The real question is, what's saving startups in places
24 | like Silicon Valley?
25 | [2]EnvironmentI think there are two components to the antidote: being in a place
26 | where startups are the cool thing to do, and chance meetings with
27 | people who can help you. And what drives them both is the number
28 | of startup people around you.The first component is particularly helpful in the first stage of
29 | a startup's life, when you go from merely having an interest in
30 | starting a company to actually doing it. It's quite a leap to start
31 | a startup. It's an unusual thing to do. But in Silicon Valley it
32 | seems normal.
33 | [3]In most places, if you start a startup, people treat you as if
34 | you're unemployed. People in the Valley aren't automatically
35 | impressed with you just because you're starting a company, but they
36 | pay attention. Anyone who's been here any amount of time knows not
37 | to default to skepticism, no matter how inexperienced you seem or
38 | how unpromising your idea sounds at first, because they've all seen
39 | inexperienced founders with unpromising sounding ideas who a few
40 | years later were billionaires.Having people around you care about what you're doing is an
41 | extraordinarily powerful force. Even the
42 | most willful people are susceptible to it. About a year after we
43 | started Y Combinator I said something to a partner at a well known
44 | VC firm that gave him the (mistaken) impression I was considering
45 | starting another startup. He responded so eagerly that for about
46 | half a second I found myself considering doing it.In most other cities, the prospect of starting a startup just doesn't
47 | seem real. In the Valley it's not only real but fashionable. That
48 | no doubt causes a lot of people to start startups who shouldn't.
49 | But I think that's ok. Few people are suited to running a startup,
50 | and it's very hard to predict beforehand which are (as I know all
51 | too well from being in the business of trying to predict beforehand),
52 | so lots of people starting startups who shouldn't is probably the
53 | optimal state of affairs. As long as you're at a point in your
54 | life when you can bear the risk of failure, the best way to find
55 | out if you're suited to running a startup is to try
56 | it.ChanceThe second component of the antidote is chance meetings with people
57 | who can help you. This force works in both phases: both in the
58 | transition from the desire to start a startup to starting one, and
59 | the transition from starting a company to succeeding. The power
60 | of chance meetings is more variable than people around you caring
61 | about startups, which is like a sort of background radiation that
62 | affects everyone equally, but at its strongest it is far stronger.Chance meetings produce miracles to compensate for the disasters
63 | that characteristically befall startups. In the Valley, terrible
64 | things happen to startups all the time, just like they do to startups
65 | everywhere. The reason startups are more likely to make it here
66 | is that great things happen to them too. In the Valley, lightning
67 | has a sign bit.For example, you start a site for college students and you decide
68 | to move to the Valley for the summer to work on it. And then on a
69 | random suburban street in Palo Alto you happen to run into Sean
70 | Parker, who understands the domain really well because he started
71 | a similar startup himself, and also knows all the investors. And
72 | moreover has advanced views, for 2004, on founders retaining control of their companies.You can't say precisely what the miracle will be, or even for sure
73 | that one will happen. The best one can say is: if you're in a
74 | startup hub, unexpected good things will probably happen to you,
75 | especially if you deserve them.I bet this is true even for startups we fund. Even with us working
76 | to make things happen for them on purpose rather than by accident,
77 | the frequency of helpful chance meetings in the Valley is so high
78 | that it's still a significant increment on what we can deliver.Chance meetings play a role like the role relaxation plays in having
79 | ideas. Most people have had the experience of working hard on some
80 | problem, not being able to solve it, giving up and going to bed,
81 | and then thinking of the answer in the shower in the morning. What
82 | makes the answer appear is letting your thoughts drift a bit—and thus drift off the wrong
83 | path you'd been pursuing last night and onto the right one adjacent
84 | to it.Chance meetings let your acquaintance drift in the same way taking
85 | a shower lets your thoughts drift. The critical thing in both cases
86 | is that they drift just the right amount. The meeting between Larry
87 | Page and Sergey Brin was a good example. They let their acquaintance
88 | drift, but only a little; they were both meeting someone they had
89 | a lot in common with.For Larry Page the most important component of the antidote was
90 | Sergey Brin, and vice versa. The antidote is
91 | people. It's not the
92 | physical infrastructure of Silicon Valley that makes it work, or
93 | the weather, or anything like that. Those helped get it started,
94 | but now that the reaction is self-sustaining what drives it is the
95 | people.Many observers have noticed that one of the most distinctive things
96 | about startup hubs is the degree to which people help one another
97 | out, with no expectation of getting anything in return. I'm not
98 | sure why this is so. Perhaps it's because startups are less of a
99 | zero sum game than most types of business; they are rarely killed
100 | by competitors. Or perhaps it's because so many startup founders
101 | have backgrounds in the sciences, where collaboration is encouraged.A large part of YC's function is to accelerate that process. We're
102 | a sort of Valley within the Valley, where the density of people
103 | working on startups and their willingness to help one another are
104 | both artificially amplified.NumbersBoth components of the antidote—an environment that encourages
105 | startups, and chance meetings with people who help you—are
106 | driven by the same underlying cause: the number of startup people
107 | around you. To make a startup hub, you need a lot of people
108 | interested in startups.There are three reasons. The first, obviously, is that if you don't
109 | have enough density, the chance meetings don't happen.
110 | [4]
111 | The second is that different startups need such different things, so
112 | you need a lot of people to supply each startup with what they need
113 | most. Sean Parker was exactly what Facebook needed in 2004. Another
114 | startup might have needed a database guy, or someone with connections
115 | in the movie business.This is one of the reasons we fund such a large number of companies,
116 | incidentally. The bigger the community, the greater the chance it
117 | will contain the person who has that one thing you need most.The third reason you need a lot of people to make a startup hub is
118 | that once you have enough people interested in the same problem,
119 | they start to set the social norms. And it is a particularly
120 | valuable thing when the atmosphere around you encourages you to do
121 | something that would otherwise seem too ambitious. In most places
122 | the atmosphere pulls you back toward the mean.I flew into the Bay Area a few days ago. I notice this every time
123 | I fly over the Valley: somehow you can sense something is going on.
124 | Obviously you can sense prosperity in how well kept a
125 | place looks. But there are different kinds of prosperity. Silicon
126 | Valley doesn't look like Boston, or New York, or LA, or DC. I tried
127 | asking myself what word I'd use to describe the feeling the Valley
128 | radiated, and the word that came to mind was optimism.Notes[1]
129 | I'm not saying it's impossible to succeed in a city with few
130 | other startups, just harder. If you're sufficiently good at
131 | generating your own morale, you can survive without external
132 | encouragement. Wufoo was based in Tampa and they succeeded. But
133 | the Wufoos are exceptionally disciplined.[2]
134 | Incidentally, this phenomenon is not limited to startups. Most
135 | unusual ambitions fail, unless the person who has them manages to
136 | find the right sort of community.[3]
137 | Starting a company is common, but starting a startup is rare.
138 | I've talked about the distinction between the two elsewhere, but
139 | essentially a startup is a new business designed for scale. Most
140 | new businesses are service businesses and except in rare cases those
141 | don't scale.[4]
142 | As I was writing this, I had a demonstration of the density of
143 | startup people in the Valley. Jessica and I bicycled to University
144 | Ave in Palo Alto to have lunch at the fabulous Oren's Hummus. As
145 | we walked in, we met Charlie Cheever sitting near the door. Selina
146 | Tobaccowala stopped to say hello on her way out. Then Josh Wilson
147 | came in to pick up a take out order. After lunch we went to get
148 | frozen yogurt. On the way we met Rajat Suri. When we got to the
149 | yogurt place, we found Dave Shen there, and as we walked out we ran
150 | into Yuri Sagalov. We walked with him for a block or so and we ran
151 | into Muzzammil Zaveri, and then a block later we met Aydin Senkut.
152 | This is everyday life in Palo Alto. I wasn't trying to meet people;
153 | I was just having lunch. And I'm sure for every startup founder
154 | or investor I saw that I knew, there were 5 more I didn't. If Ron
155 | Conway had been with us he would have met 30 people he knew.Thanks to Sam Altman, Paul Buchheit, Jessica Livingston, and
156 | Harj Taggar for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/gba.txt:
--------------------------------------------------------------------------------
1 | April 2004To the popular press, "hacker" means someone who breaks
2 | into computers. Among programmers it means a good programmer.
3 | But the two meanings are connected. To programmers,
4 | "hacker" connotes mastery in the most literal sense: someone
5 | who can make a computer do what he wants—whether the computer
6 | wants to or not.To add to the confusion, the noun "hack" also has two senses. It can
7 | be either a compliment or an insult. It's called a hack when
8 | you do something in an ugly way. But when you do something
9 | so clever that you somehow beat the system, that's also
10 | called a hack. The word is used more often in the former than
11 | the latter sense, probably because ugly solutions are more
12 | common than brilliant ones.Believe it or not, the two senses of "hack" are also
13 | connected. Ugly and imaginative solutions have something in
14 | common: they both break the rules. And there is a gradual
15 | continuum between rule breaking that's merely ugly (using
16 | duct tape to attach something to your bike) and rule breaking
17 | that is brilliantly imaginative (discarding Euclidean space).Hacking predates computers. When he
18 | was working on the Manhattan Project, Richard Feynman used to
19 | amuse himself by breaking into safes containing secret documents.
20 | This tradition continues today.
21 | When we were in grad school, a hacker friend of mine who spent too much
22 | time around MIT had
23 | his own lock picking kit.
24 | (He now runs a hedge fund, a not unrelated enterprise.)It is sometimes hard to explain to authorities why one would
25 | want to do such things.
26 | Another friend of mine once got in trouble with the government for
27 | breaking into computers. This had only recently been declared
28 | a crime, and the FBI found that their usual investigative
29 | technique didn't work. Police investigation apparently begins with
30 | a motive. The usual motives are few: drugs, money, sex,
31 | revenge. Intellectual curiosity was not one of the motives on
32 | the FBI's list. Indeed, the whole concept seemed foreign to
33 | them.Those in authority tend to be annoyed by hackers'
34 | general attitude of disobedience. But that disobedience is
35 | a byproduct of the qualities that make them good programmers.
36 | They may laugh at the CEO when he talks in generic corporate
37 | newspeech, but they also laugh at someone who tells them
38 | a certain problem can't be solved.
39 | Suppress one, and you suppress the other.This attitude is sometimes affected. Sometimes young programmers
40 | notice the eccentricities of eminent hackers and decide to
41 | adopt some of their own in order to seem smarter.
42 | The fake version is not merely
43 | annoying; the prickly attitude of these posers
44 | can actually slow the process of innovation.But even factoring in their annoying eccentricities,
45 | the disobedient attitude of hackers is a net win. I wish its
46 | advantages were better understood.For example, I suspect people in Hollywood are
47 | simply mystified by
48 | hackers' attitudes toward copyrights. They are a perennial
49 | topic of heated discussion on Slashdot.
50 | But why should people who program computers
51 | be so concerned about copyrights, of all things?Partly because some companies use mechanisms to prevent
52 | copying. Show any hacker a lock and his first thought is
53 | how to pick it. But there is a deeper reason that
54 | hackers are alarmed by measures like copyrights and patents.
55 | They see increasingly aggressive measures to protect
56 | "intellectual property"
57 | as a threat to the intellectual
58 | freedom they need to do their job.
59 | And they are right.It is by poking about inside current technology that
60 | hackers get ideas for the next generation. No thanks,
61 | intellectual homeowners may say, we don't need any
62 | outside help. But they're wrong.
63 | The next generation of computer technology has
64 | often—perhaps more often than not—been developed by outsiders.In 1977 there was no doubt some group within IBM developing
65 | what they expected to be
66 | the next generation of business computer. They were mistaken.
67 | The next generation of business computer was
68 | being developed on entirely different lines by two long-haired
69 | guys called Steve in a garage in Los Altos. At about the
70 | same time, the powers that be
71 | were cooperating to develop the
72 | official next generation operating system, Multics.
73 | But two guys who thought Multics excessively complex went off
74 | and wrote their own. They gave it a name that
75 | was a joking reference to Multics: Unix.The latest intellectual property laws impose
76 | unprecedented restrictions on the sort of poking around that
77 | leads to new ideas. In the past, a competitor might use patents
78 | to prevent you from selling a copy of something they
79 | made, but they couldn't prevent you from
80 | taking one apart to see how it worked. The latest
81 | laws make this a crime. How are we
82 | to develop new technology if we can't study current
83 | technology to figure out how to improve it?Ironically, hackers have brought this on themselves.
84 | Computers are responsible for the problem. The control systems
85 | inside machines used to be physical: gears and levers and cams.
86 | Increasingly, the brains (and thus the value) of products is
87 | in software. And by this I mean software in the general sense:
88 | i.e. data. A song on an LP is physically stamped into the
89 | plastic. A song on an iPod's disk is merely stored on it.Data is by definition easy to copy. And the Internet
90 | makes copies easy to distribute. So it is no wonder
91 | companies are afraid. But, as so often happens, fear has
92 | clouded their judgement. The government has responded
93 | with draconian laws to protect intellectual property.
94 | They probably mean well. But
95 | they may not realize that such laws will do more harm
96 | than good.Why are programmers so violently opposed to these laws?
97 | If I were a legislator, I'd be interested in this
98 | mystery—for the same reason that, if I were a farmer and suddenly
99 | heard a lot of squawking coming from my hen house one night,
100 | I'd want to go out and investigate. Hackers are not stupid,
101 | and unanimity is very rare in this world.
102 | So if they're all squawking,
103 | perhaps there is something amiss.Could it be that such laws, though intended to protect America,
104 | will actually harm it? Think about it. There is something
105 | very American about Feynman breaking into safes during
106 | the Manhattan Project. It's hard to imagine the authorities
107 | having a sense of humor about such things over
108 | in Germany at that time. Maybe it's not a coincidence.Hackers are unruly. That is the essence of hacking. And it
109 | is also the essence of Americanness. It is no accident
110 | that Silicon Valley
111 | is in America, and not France, or Germany,
112 | or England, or Japan. In those countries, people color inside
113 | the lines.I lived for a while in Florence. But after I'd been there
114 | a few months I realized that what I'd been unconsciously hoping
115 | to find there was back in the place I'd just left.
116 | The reason Florence is famous is that in 1450, it was New York.
117 | In 1450 it was filled with the kind of turbulent and ambitious
118 | people you find now in America. (So I went back to America.)It is greatly to America's advantage that it is
119 | a congenial atmosphere for the right sort of unruliness—that
120 | it is a home not just for the smart, but for smart-alecks.
121 | And hackers are invariably smart-alecks. If we had a national
122 | holiday, it would be April 1st. It says a great deal about
123 | our work that we use the same word for a brilliant or a
124 | horribly cheesy solution. When we cook one up we're not
125 | always 100% sure which kind it is. But as long as it has
126 | the right sort of wrongness, that's a promising sign.
127 | It's odd that people
128 | think of programming as precise and methodical. Computers
129 | are precise and methodical. Hacking is something you do
130 | with a gleeful laugh.In our world some of the most characteristic solutions
131 | are not far removed from practical
132 | jokes. IBM was no doubt rather surprised by the consequences
133 | of the licensing deal for DOS, just as the hypothetical
134 | "adversary" must be when Michael Rabin solves a problem by
135 | redefining it as one that's easier to solve.Smart-alecks have to develop a keen sense of how much they
136 | can get away with. And lately hackers
137 | have sensed a change
138 | in the atmosphere.
139 | Lately hackerliness seems rather frowned upon.To hackers the recent contraction in civil liberties seems
140 | especially ominous. That must also mystify outsiders.
141 | Why should we care especially about civil
142 | liberties? Why programmers, more than
143 | dentists or salesmen or landscapers?Let me put the case in terms a government official would appreciate.
144 | Civil liberties are not just an ornament, or a quaint
145 | American tradition. Civil liberties make countries rich.
146 | If you made a graph of
147 | GNP per capita vs. civil liberties, you'd notice a definite
148 | trend. Could civil liberties really be a cause, rather
149 | than just an effect? I think so. I think a society in which
150 | people can do and say what they want will also tend to
151 | be one in which the most efficient solutions win, rather than
152 | those sponsored by the most influential people.
153 | Authoritarian countries become corrupt;
154 | corrupt countries become poor; and poor countries are weak.
155 | It seems to me there is
156 | a Laffer curve for government power, just as for
157 | tax revenues. At least, it seems likely enough that it
158 | would be stupid to try the experiment and find out. Unlike
159 | high tax rates, you can't repeal totalitarianism if it
160 | turns out to be a mistake.This is why hackers worry. The government spying on people doesn't
161 | literally make programmers write worse code. It just leads
162 | eventually to a world in which bad ideas win. And because
163 | this is so important to hackers, they're especially sensitive
164 | to it. They can sense totalitarianism approaching from a
165 | distance, as animals can sense an approaching
166 | thunderstorm.It would be ironic if, as hackers fear, recent measures
167 | intended to protect national security and intellectual property
168 | turned out to be a missile aimed right at what makes
169 | America successful. But it would not be the first time that
170 | measures taken in an atmosphere of panic had
171 | the opposite of the intended effect.There is such a thing as Americanness.
172 | There's nothing like living abroad to teach you that.
173 | And if you want to know whether something will nurture or squash
174 | this quality, it would be hard to find a better focus
175 | group than hackers, because they come closest of any group
176 | I know to embodying it. Closer, probably, than
177 | the men running our government,
178 | who for all their talk of patriotism
179 | remind me more of Richelieu or Mazarin
180 | than Thomas Jefferson or George Washington.When you read what the founding fathers had to say for
181 | themselves, they sound more like hackers.
182 | "The spirit of resistance to government,"
183 | Jefferson wrote, "is so valuable on certain occasions, that I wish
184 | it always to be kept alive."Imagine an American president saying that today.
185 | Like the remarks of an outspoken old grandmother, the sayings of
186 | the founding fathers have embarrassed generations of
187 | their less confident successors. They remind us where we come from.
188 | They remind us that it is the people who break rules that are
189 | the source of America's wealth and power.Those in a position to impose rules naturally want them to be
190 | obeyed. But be careful what you ask for. You might get it.Thanks to Ken Anderson, Trevor Blackwell, Daniel Giffin,
191 | Sarah Harlin, Shiro Kawai, Jessica Livingston, Matz,
192 | Jackie McDonough, Robert Morris, Eric Raymond, Guido van Rossum,
193 | David Weinberger, and
194 | Steven Wolfram for reading drafts of this essay.
195 | (The image shows Steves Jobs and Wozniak
196 | with a "blue box."
197 | Photo by Margret Wozniak. Reproduced by permission of Steve
198 | Wozniak.)
--------------------------------------------------------------------------------
/viz.ipynb:
--------------------------------------------------------------------------------
1 | {
2 | "cells": [
3 | {
4 | "cell_type": "code",
5 | "execution_count": 1,
6 | "id": "090f0080",
7 | "metadata": {},
8 | "outputs": [],
9 | "source": [
10 | "import uuid\n",
11 | "import pandas as pd\n",
12 | "import json\n",
13 | "import os\n",
14 | "import glob\n",
15 | "import jsonlines\n",
16 | "import requests\n",
17 | "from tqdm import trange\n",
18 | "import random\n",
19 | "import json_repair\n",
20 | "import seaborn as sns\n",
21 | "import matplotlib.pyplot as plt\n",
22 | "from matplotlib.colors import LinearSegmentedColormap\n",
23 | "import pandas as pd\n",
24 | "from collections import Counter"
25 | ]
26 | },
27 | {
28 | "cell_type": "code",
29 | "execution_count": 2,
30 | "id": "b9cdf0ef",
31 | "metadata": {},
32 | "outputs": [],
33 | "source": [
34 | "def get_reasoning_score(index, pre):\n",
35 | " file = open(\"context_data/a_stars.txt\", \"r\")\n",
36 | " a_stars = eval(file.readline())[\"32\"]\n",
37 | " file = open(\"context_data/r_stars.txt\", \"r\")\n",
38 | " r_stars = eval(file.readline())[\"32\"]\n",
39 | " if a_stars[index] in pre and r_stars[index] in pre:\n",
40 | " return 0.5\n",
41 | " elif a_stars[index] in pre and r_stars[index] not in pre:\n",
42 | " return 1\n",
43 | " elif a_stars[index] not in pre and r_stars[index] in pre:\n",
44 | " return 0.25\n",
45 | " else:\n",
46 | " return 0\n",
47 | "# get formate context size\n",
48 | "def get_context_size(max_context_length, n):\n",
49 | " intervel = int(max_context_length / n)\n",
50 | " return [i for i in range(intervel, max_context_length + 1, intervel)]\n",
51 | "\n",
52 | "# reduce duplicate from the predicted results\n",
53 | "def reduce_duplicate(predicted, m):\n",
54 | " if len(predicted) > m:\n",
55 | " predicted = predicted[:m]\n",
56 | " predicted = list(set(predicted))\n",
57 | " else:\n",
58 | " predicted = list(set(predicted))\n",
59 | " return predicted\n",
60 | "\n",
61 | "# get results from English version of the Counting-Stars\n",
62 | "def get_data_EN(folder_path, max_context_length, m, n, test_type):\n",
63 | " context_size = get_context_size(max_context_length, n)\n",
64 | " data = []\n",
65 | " average_score = 0\n",
66 | " indicator = 0\n",
67 | " if test_type == \"Acquisition\":\n",
68 | " scalar = 0.82\n",
69 | " elif test_type == \"Reasoning\":\n",
70 | " scalar = 0.815\n",
71 | " for item in jsonlines.Reader(folder_path):\n",
72 | " if \"```\" in item['answer']:\n",
73 | " predicted = json_repair.loads(item['answer'].replace('```','').replace(\"json\",'').strip())['little_penguin']\n",
74 | " else:\n",
75 | " try:\n",
76 | " predicted = json.loads(item['answer'])['little_penguin']\n",
77 | " except:\n",
78 | " predicted = item['answer']['little_penguin']\n",
79 | " predicted = reduce_duplicate(predicted, m)\n",
80 | " for i in range(1, m+1): \n",
81 | " counting_times = i\n",
82 | " if test_type == \"Acquisition\":\n",
83 | " try:\n",
84 | " if item[\"reference_counting_results\"][i-1] in predicted:\n",
85 | " score = 1\n",
86 | " else:\n",
87 | " score = 0\n",
88 | " except:\n",
89 | " score = 0\n",
90 | " else:\n",
91 | " score = get_reasoning_score(counting_times-1, predicted)\n",
92 | " average_score += score\n",
93 | " data.append({\n",
94 | " \"Counting Times\": counting_times,\n",
95 | " \"Context Size\": int(item['context_size'] / scalar),\n",
96 | " \"Score\": score\n",
97 | " })\n",
98 | " df = pd.DataFrame(data)\n",
99 | " print (df.head())\n",
100 | " print (f\"You have {len(df)} rows\")\n",
101 | " pivot_table = pd.pivot_table(df, values='Score', index=['Counting Times', 'Context Size'], aggfunc='mean').reset_index()\n",
102 | " pivot_table = pivot_table.pivot(index=\"Counting Times\", columns=\"Context Size\", values=\"Score\")\n",
103 | " return pivot_table, pivot_table.mean(axis=None).round(3)\n",
104 | "\n",
105 | "# get results from Chinese version of the Counting-Stars\n",
106 | "def get_data_ZH(folder_path, max_context_length, m, n, test_type):\n",
107 | " context_size = get_context_size(max_context_length, n)\n",
108 | " data = []\n",
109 | " average_score = 0\n",
110 | " indicator = 0\n",
111 | " if test_type == \"Acquisition\":\n",
112 | " scalar = 0.725\n",
113 | " elif test_type == \"Reasoning\":\n",
114 | " scalar = 0.72\n",
115 | " for item in jsonlines.Reader(folder_path):\n",
116 | " if \"```\" in item['answer']:\n",
117 | " predicted = json_repair.loads(item['answer'].replace('```','').replace(\"json\",'').strip())['小企鹅']\n",
118 | " else:\n",
119 | " try:\n",
120 | " predicted = json_repair.loads(item['answer'])['小企鹅'] \n",
121 | " except:\n",
122 | " predicted = item['answer']['小企鹅']\n",
123 | " predicted = reduce_duplicate(predicted, m)\n",
124 | " for i in range(1, m+1): \n",
125 | " counting_times = i\n",
126 | " if test_type == \"Acquisition\":\n",
127 | " try:\n",
128 | " if item[\"reference_counting_results\"][i-1] in predicted:\n",
129 | " score = 1\n",
130 | " else:\n",
131 | " score = 0\n",
132 | " except:\n",
133 | " score = 0\n",
134 | " else:\n",
135 | " score = get_reasoning_score(counting_times-1, predicted)\n",
136 | " average_score += score\n",
137 | " data.append({\n",
138 | " \"Counting Times\": counting_times,\n",
139 | " \"Context Size\": int(item['context_size'] / scalar),\n",
140 | " \"Score\": score\n",
141 | " })\n",
142 | " df = pd.DataFrame(data)\n",
143 | " print (df.head())\n",
144 | " print (f\"You have {len(df)} rows\")\n",
145 | " pivot_table = pd.pivot_table(df, values='Score', index=['Counting Times', 'Context Size'], aggfunc='mean').reset_index()\n",
146 | " pivot_table = pivot_table.pivot(index=\"Counting Times\", columns=\"Context Size\", values=\"Score\")\n",
147 | " return pivot_table, pivot_table.mean(axis=None).round(3)"
148 | ]
149 | },
150 | {
151 | "cell_type": "code",
152 | "execution_count": null,
153 | "id": "0506824c",
154 | "metadata": {},
155 | "outputs": [],
156 | "source": [
157 | "import numpy\n",
158 | "m = 32\n",
159 | "n = 32\n",
160 | "max_context_length = 128000\n",
161 | "\n",
162 | "testing_type = \"Acquisition\"\n",
163 | "#testing_type = \"Reasoning\"\n",
164 | "\n",
165 | "folder_path_test = open(\"xxx\",\"r\")\n",
166 | "viz_data_gpt, mean_gpt = get_data_ZH(folder_path_test, max_context_length, m, n, testing_type)\n",
167 | " \n",
168 | "folder_path_test = open(\"xxx\",\"r\")\n",
169 | "viz_data_gemini, mean_gemini = get_data_ZH(folder_path_test, max_context_length, m, n, testing_type)\n",
170 | " \n",
171 | "folder_path_test = open(\"xxx\",\"r\")\n",
172 | "viz_data_claude, mean_claude = get_data_ZH(folder_path_test, max_context_length, m, n, testing_type)"
173 | ]
174 | },
175 | {
176 | "cell_type": "code",
177 | "execution_count": null,
178 | "id": "44e53cf2",
179 | "metadata": {},
180 | "outputs": [],
181 | "source": [
182 | "# Create a custom colormap. Go to https://coolors.co/ and pick cool colors\n",
183 | "cmap = LinearSegmentedColormap.from_list(\"custom_cmap\", [\"#184E77\", \"#1E6091\", \"#1A759F\", \"#168AAD\", \"#34A0A4\", \"#52B69A\", \"#76C893\", \"#99D98C\", \"#B5E48C\", \"#D9ED92\"])\n",
184 | "\n",
185 | "fig = plt.figure(figsize=(17, 14))\n",
186 | "ax1 = fig.add_subplot(3, 1, 1)\n",
187 | "# Create the heatmap with better aesthetics\n",
188 | "sns.heatmap(\n",
189 | " viz_data_gpt,\n",
190 | " #annot=True,\n",
191 | " fmt=\"g\",\n",
192 | " cmap=cmap,\n",
193 | " linewidths=0.3,\n",
194 | " cbar_kws={'label': 'Score', \"pad\": 0.02}\n",
195 | ")\n",
196 | "\n",
197 | "labels = [i for i in range(2, m+1, 2)]\n",
198 | "x = [i-0.5 for i in range(2, m+1, 2)]\n",
199 | "\n",
200 | "# More aesthetics\n",
201 | "plt.title(f'Counting-Stars-({m})-(Multi-evidence {testing_type}): GPT-4 Turbo (Acc: {mean_gpt})', size=11) # Adds a title\n",
202 | "plt.xlabel('Context Length', size=11) # X-axis label\n",
203 | "plt.ylabel('Counting Times', size=11) # Y-axis label\n",
204 | "plt.xticks(rotation=45) # Rotates the x-axis labels to prevent overlap\n",
205 | "plt.yticks(x, labels, rotation=45) # Ensures the y-axis labels are horizontal\n",
206 | "\n",
207 | "\n",
208 | "\n",
209 | "ax2 = fig.add_subplot(3, 1, 2)\n",
210 | "# Create the heatmap with better aesthetics\n",
211 | "sns.heatmap(\n",
212 | " viz_data_claude,\n",
213 | " #annot=True,\n",
214 | " fmt=\"g\",\n",
215 | " cmap=cmap,\n",
216 | " linewidths=0.3,\n",
217 | " cbar_kws={'label': 'Score', \"pad\": 0.02}\n",
218 | ")\n",
219 | "\n",
220 | "labels = [i for i in range(2, m+1, 2)]\n",
221 | "x = [i-0.5 for i in range(2, m+1, 2)]\n",
222 | "\n",
223 | "# More aesthetics\n",
224 | "plt.title(f'Counting-Stars-({m})-(Multi-evidence {testing_type}): Claude3 Opus (Acc: {mean_claude})', size=11) # Adds a title\n",
225 | "plt.xlabel('Context Length', size=11) # X-axis label\n",
226 | "plt.ylabel('Counting Times', size=11) # Y-axis label\n",
227 | "plt.xticks(rotation=45) # Rotates the x-axis labels to prevent overlap\n",
228 | "plt.yticks(x, labels, rotation=45) # Ensures the y-axis labels are horizontal\n",
229 | "\n",
230 | "\n",
231 | "ax3 = fig.add_subplot(3, 1, 3)\n",
232 | "\n",
233 | "# Create the heatmap with better aesthetics\n",
234 | "sns.heatmap(\n",
235 | " viz_data_gemini,\n",
236 | " #annot=True,\n",
237 | " fmt=\"g\",\n",
238 | " cmap=cmap,\n",
239 | " linewidths=0.3,\n",
240 | " cbar_kws={'label': 'Score', \"pad\": 0.02}\n",
241 | ")\n",
242 | "\n",
243 | "labels = [i for i in range(2, m+1, 2)]\n",
244 | "x = [i-0.5 for i in range(2, m+1, 2)]\n",
245 | "\n",
246 | "# More aesthetics\n",
247 | "plt.title(f'Counting-Stars-({m})-(Multi-evidence {testing_type}): Gemini Pro 1.5 (Acc: {mean_gemini})', size=11) # Adds a title\n",
248 | "plt.xlabel('Context Length', size=11) # X-axis label\n",
249 | "plt.ylabel('Counting Times', size=11) # Y-axis label\n",
250 | "plt.xticks(rotation=45) # Rotates the x-axis labels to prevent overlap\n",
251 | "plt.yticks(x, labels, rotation=45) # Ensures the y-axis labels are horizontal\n",
252 | "\n",
253 | "fig.subplots_adjust(hspace=0.4)\n",
254 | "plt.savefig(f\"results.pdf\", dpi=2380, bbox_inches='tight')\n",
255 | "plt.show()"
256 | ]
257 | },
258 | {
259 | "cell_type": "code",
260 | "execution_count": null,
261 | "id": "d2bec855",
262 | "metadata": {},
263 | "outputs": [],
264 | "source": []
265 | },
266 | {
267 | "cell_type": "code",
268 | "execution_count": null,
269 | "id": "7f0a7a38",
270 | "metadata": {},
271 | "outputs": [],
272 | "source": []
273 | }
274 | ],
275 | "metadata": {
276 | "kernelspec": {
277 | "display_name": "Python 3 (ipykernel)",
278 | "language": "python",
279 | "name": "python3"
280 | },
281 | "language_info": {
282 | "codemirror_mode": {
283 | "name": "ipython",
284 | "version": 3
285 | },
286 | "file_extension": ".py",
287 | "mimetype": "text/x-python",
288 | "name": "python",
289 | "nbconvert_exporter": "python",
290 | "pygments_lexer": "ipython3",
291 | "version": "3.11.5"
292 | }
293 | },
294 | "nbformat": 4,
295 | "nbformat_minor": 5
296 | }
297 |
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/apple.txt:
--------------------------------------------------------------------------------
1 |
2 |
3 | Want to start a startup? Get funded by
4 | Y Combinator.
5 |
6 |
7 |
8 |
9 | November 2009I don't think Apple realizes how badly the App Store approval process
10 | is broken. Or rather, I don't think they realize how much it matters
11 | that it's broken.The way Apple runs the App Store has harmed their reputation with
12 | programmers more than anything else they've ever done.
13 | Their reputation with programmers used to be great.
14 | It used to be the most common complaint you heard
15 | about Apple was that their fans admired them too uncritically.
16 | The App Store has changed that. Now a lot of programmers
17 | have started to see Apple as evil.How much of the goodwill Apple once had with programmers have they
18 | lost over the App Store? A third? Half? And that's just so far.
19 | The App Store is an ongoing karma leak.* * *How did Apple get into this mess? Their fundamental problem is
20 | that they don't understand software.They treat iPhone apps the way they treat the music they sell through
21 | iTunes. Apple is the channel; they own the user; if you want to
22 | reach users, you do it on their terms. The record labels agreed,
23 | reluctantly. But this model doesn't work for software. It doesn't
24 | work for an intermediary to own the user. The software business
25 | learned that in the early 1980s, when companies like VisiCorp showed
26 | that although the words "software" and "publisher" fit together,
27 | the underlying concepts don't. Software isn't like music or books.
28 | It's too complicated for a third party to act as an intermediary
29 | between developer and user. And yet that's what Apple is trying
30 | to be with the App Store: a software publisher. And a particularly
31 | overreaching one at that, with fussy tastes and a rigidly enforced
32 | house style.If software publishing didn't work in 1980, it works even less now
33 | that software development has evolved from a small number of big
34 | releases to a constant stream of small ones. But Apple doesn't
35 | understand that either. Their model of product development derives
36 | from hardware. They work on something till they think it's finished,
37 | then they release it. You have to do that with hardware, but because
38 | software is so easy to change, its design can benefit from evolution.
39 | The standard way to develop applications now is to launch fast and
40 | iterate. Which means it's a disaster to have long, random delays
41 | each time you release a new version.Apparently Apple's attitude is that developers should be more careful
42 | when they submit a new version to the App Store. They would say
43 | that. But powerful as they are, they're not powerful enough to
44 | turn back the evolution of technology. Programmers don't use
45 | launch-fast-and-iterate out of laziness. They use it because it
46 | yields the best results. By obstructing that process, Apple is
47 | making them do bad work, and programmers hate that as much as Apple
48 | would.How would Apple like it if when they discovered a serious bug in
49 | OS X, instead of releasing a software update immediately, they had
50 | to submit their code to an intermediary who sat on it for a month
51 | and then rejected it because it contained an icon they didn't like?By breaking software development, Apple gets the opposite of what
52 | they intended: the version of an app currently available in the App
53 | Store tends to be an old and buggy one. One developer told me:
54 |
55 | As a result of their process, the App Store is full of half-baked
56 | applications. I make a new version almost every day that I release
57 | to beta users. The version on the App Store feels old and crappy.
58 | I'm sure that a lot of developers feel this way: One emotion is
59 | "I'm not really proud about what's in the App Store", and it's
60 | combined with the emotion "Really, it's Apple's fault."
61 |
62 | Another wrote:
63 |
64 | I believe that they think their approval process helps users by
65 | ensuring quality. In reality, bugs like ours get through all the
66 | time and then it can take 4-8 weeks to get that bug fix approved,
67 | leaving users to think that iPhone apps sometimes just don't work.
68 | Worse for Apple, these apps work just fine on other platforms
69 | that have immediate approval processes.
70 |
71 | Actually I suppose Apple has a third misconception: that all the
72 | complaints about App Store approvals are not a serious problem.
73 | They must hear developers complaining. But partners and suppliers
74 | are always complaining. It would be a bad sign if they weren't;
75 | it would mean you were being too easy on them. Meanwhile the iPhone
76 | is selling better than ever. So why do they need to fix anything?They get away with maltreating developers, in the short term, because
77 | they make such great hardware. I just bought a new 27" iMac a
78 | couple days ago. It's fabulous. The screen's too shiny, and the
79 | disk is surprisingly loud, but it's so beautiful that you can't
80 | make yourself care.So I bought it, but I bought it, for the first time, with misgivings.
81 | I felt the way I'd feel buying something made in a country with a
82 | bad human rights record. That was new. In the past when I bought
83 | things from Apple it was an unalloyed pleasure. Oh boy! They make
84 | such great stuff. This time it felt like a Faustian bargain. They
85 | make such great stuff, but they're such assholes. Do I really want
86 | to support this company?* * *Should Apple care what people like me think? What difference does
87 | it make if they alienate a small minority of their users?There are a couple reasons they should care. One is that these
88 | users are the people they want as employees. If your company seems
89 | evil, the best programmers won't work for you. That hurt Microsoft
90 | a lot starting in the 90s. Programmers started to feel sheepish
91 | about working there. It seemed like selling out. When people from
92 | Microsoft were talking to other programmers and they mentioned where
93 | they worked, there were a lot of self-deprecating jokes about having
94 | gone over to the dark side. But the real problem for Microsoft
95 | wasn't the embarrassment of the people they hired. It was the
96 | people they never got. And you know who got them? Google and
97 | Apple. If Microsoft was the Empire, they were the Rebel Alliance.
98 | And it's largely because they got more of the best people that
99 | Google and Apple are doing so much better than Microsoft today.Why are programmers so fussy about their employers' morals? Partly
100 | because they can afford to be. The best programmers can work
101 | wherever they want. They don't have to work for a company they
102 | have qualms about.But the other reason programmers are fussy, I think, is that evil
103 | begets stupidity. An organization that wins by exercising power
104 | starts to lose the ability to win by doing better work. And it's
105 | not fun for a smart person to work in a place where the best ideas
106 | aren't the ones that win. I think the reason Google embraced "Don't
107 | be evil" so eagerly was not so much to impress the outside world
108 | as to inoculate themselves against arrogance.
109 | [1]That has worked for Google so far. They've become more
110 | bureaucratic, but otherwise they seem to have held true to their
111 | original principles. With Apple that seems less the case. When you
112 | look at the famous
113 | 1984 ad
114 | now, it's easier to imagine Apple as the
115 | dictator on the screen than the woman with the hammer.
116 | [2]
117 | In fact, if you read the dictator's speech it sounds uncannily like a
118 | prophecy of the App Store.
119 |
120 | We have triumphed over the unprincipled dissemination of facts.We have created, for the first time in all history, a garden of
121 | pure ideology, where each worker may bloom secure from the pests
122 | of contradictory and confusing truths.
123 |
124 | The other reason Apple should care what programmers think of them
125 | is that when you sell a platform, developers make or break you. If
126 | anyone should know this, Apple should. VisiCalc made the Apple II.And programmers build applications for the platforms they use. Most
127 | applications—most startups, probably—grow out of personal projects.
128 | Apple itself did. Apple made microcomputers because that's what
129 | Steve Wozniak wanted for himself. He couldn't have afforded a
130 | minicomputer.
131 | [3]
132 | Microsoft likewise started out making interpreters
133 | for little microcomputers because
134 | Bill Gates and Paul Allen were interested in using them. It's a
135 | rare startup that doesn't build something the founders use.The main reason there are so many iPhone apps is that so many programmers
136 | have iPhones. They may know, because they read it in an article,
137 | that Blackberry has such and such market share. But in practice
138 | it's as if RIM didn't exist. If they're going to build something,
139 | they want to be able to use it themselves, and that means building
140 | an iPhone app.So programmers continue to develop iPhone apps, even though Apple
141 | continues to maltreat them. They're like someone stuck in an abusive
142 | relationship. They're so attracted to the iPhone that they can't
143 | leave. But they're looking for a way out. One wrote:
144 |
145 | While I did enjoy developing for the iPhone, the control they
146 | place on the App Store does not give me the drive to develop
147 | applications as I would like. In fact I don't intend to make any
148 | more iPhone applications unless absolutely necessary.
149 | [4]
150 |
151 | Can anything break this cycle? No device I've seen so far could.
152 | Palm and RIM haven't a hope. The only credible contender is Android.
153 | But Android is an orphan; Google doesn't really care about it, not
154 | the way Apple cares about the iPhone. Apple cares about the iPhone
155 | the way Google cares about search.* * *Is the future of handheld devices one locked down by Apple? It's
156 | a worrying prospect. It would be a bummer to have another grim
157 | monoculture like we had in the 1990s. In 1995, writing software
158 | for end users was effectively identical with writing Windows
159 | applications. Our horror at that prospect was the single biggest
160 | thing that drove us to start building web apps.At least we know now what it would take to break Apple's lock.
161 | You'd have to get iPhones out of programmers' hands. If programmers
162 | used some other device for mobile web access, they'd start to develop
163 | apps for that instead.How could you make a device programmers liked better than the iPhone?
164 | It's unlikely you could make something better designed. Apple
165 | leaves no room there. So this alternative device probably couldn't
166 | win on general appeal. It would have to win by virtue of some
167 | appeal it had to programmers specifically.One way to appeal to programmers is with software. If you
168 | could think of an application programmers had to have, but that
169 | would be impossible in the circumscribed world of the iPhone,
170 | you could presumably get them to switch.That would definitely happen if programmers started to use handhelds
171 | as development machines—if handhelds displaced laptops the
172 | way laptops displaced desktops. You need more control of a development
173 | machine than Apple will let you have over an iPhone.Could anyone make a device that you'd carry around in your pocket
174 | like a phone, and yet would also work as a development machine?
175 | It's hard to imagine what it would look like. But I've learned
176 | never to say never about technology. A phone-sized device that
177 | would work as a development machine is no more miraculous by present
178 | standards than the iPhone itself would have seemed by the standards
179 | of 1995.My current development machine is a MacBook Air, which I use with
180 | an external monitor and keyboard in my office, and by itself when
181 | traveling. If there was a version half the size I'd prefer it.
182 | That still wouldn't be small enough to carry around everywhere like
183 | a phone, but we're within a factor of 4 or so. Surely that gap is
184 | bridgeable. In fact, let's make it an
185 | RFS. Wanted:
186 | Woman with hammer.Notes[1]
187 | When Google adopted "Don't be evil," they were still so small
188 | that no one would have expected them to be, yet.
189 | [2]
190 | The dictator in the 1984 ad isn't Microsoft, incidentally;
191 | it's IBM. IBM seemed a lot more frightening in those days, but
192 | they were friendlier to developers than Apple is now.[3]
193 | He couldn't even afford a monitor. That's why the Apple
194 | I used a TV as a monitor.[4]
195 | Several people I talked to mentioned how much they liked the
196 | iPhone SDK. The problem is not Apple's products but their policies.
197 | Fortunately policies are software; Apple can change them instantly
198 | if they want to. Handy that, isn't it?Thanks to Sam Altman, Trevor Blackwell, Ross Boucher,
199 | James Bracy, Gabor Cselle,
200 | Patrick Collison, Jason Freedman, John Gruber, Joe Hewitt, Jessica Livingston,
201 | Robert Morris, Teng Siong Ong, Nikhil Pandit, Savraj Singh, and Jared Tame for reading drafts of this.
--------------------------------------------------------------------------------
/context_data/PaulGrahamEssays/submarine.txt:
--------------------------------------------------------------------------------
1 | April 2005"Suits make a corporate comeback," says the New
2 | York Times. Why does this sound familiar? Maybe because
3 | the suit was also back in February,
4 |
5 | September
6 | 2004, June
7 | 2004, March
8 | 2004, September
9 | 2003,
10 |
11 | November
12 | 2002,
13 | April 2002,
14 | and February
15 | 2002.
16 |
17 | Why do the media keep running stories saying suits are back? Because
18 | PR firms tell
19 | them to. One of the most surprising things I discovered
20 | during my brief business career was the existence of the PR industry,
21 | lurking like a huge, quiet submarine beneath the news. Of the
22 | stories you read in traditional media that aren't about politics,
23 | crimes, or disasters, more than half probably come from PR firms.I know because I spent years hunting such "press hits." Our startup spent
24 | its entire marketing budget on PR: at a time when we were assembling
25 | our own computers to save money, we were paying a PR firm $16,000
26 | a month. And they were worth it. PR is the news equivalent of
27 | search engine optimization; instead of buying ads, which readers
28 | ignore, you get yourself inserted directly into the stories. [1]Our PR firm
29 | was one of the best in the business. In 18 months, they got press
30 | hits in over 60 different publications.
31 | And we weren't the only ones they did great things for.
32 | In 1997 I got a call from another
33 | startup founder considering hiring them to promote his company. I
34 | told him they were PR gods, worth every penny of their outrageous
35 | fees. But I remember thinking his company's name was odd.
36 | Why call an auction site "eBay"?
37 | SymbiosisPR is not dishonest. Not quite. In fact, the reason the best PR
38 | firms are so effective is precisely that they aren't dishonest.
39 | They give reporters genuinely valuable information. A good PR firm
40 | won't bug reporters just because the client tells them to; they've
41 | worked hard to build their credibility with reporters, and they
42 | don't want to destroy it by feeding them mere propaganda.If anyone is dishonest, it's the reporters. The main reason PR
43 | firms exist is that reporters are lazy. Or, to put it more nicely,
44 | overworked. Really they ought to be out there digging up stories
45 | for themselves. But it's so tempting to sit in their offices and
46 | let PR firms bring the stories to them. After all, they know good
47 | PR firms won't lie to them.A good flatterer doesn't lie, but tells his victim selective truths
48 | (what a nice color your eyes are). Good PR firms use the same
49 | strategy: they give reporters stories that are true, but whose truth
50 | favors their clients.For example, our PR firm often pitched stories about how the Web
51 | let small merchants compete with big ones. This was perfectly true.
52 | But the reason reporters ended up writing stories about this
53 | particular truth, rather than some other one, was that small merchants
54 | were our target market, and we were paying the piper.Different publications vary greatly in their reliance on PR firms.
55 | At the bottom of the heap are the trade press, who make most of
56 | their money from advertising and would give the magazines away for
57 | free if advertisers would let them. [2] The average
58 | trade publication is a bunch of ads, glued together by just enough
59 | articles to make it look like a magazine. They're so desperate for
60 | "content" that some will print your press releases almost verbatim,
61 | if you take the trouble to write them to read like articles.At the other extreme are publications like the New York Times
62 | and the Wall Street Journal. Their reporters do go out and
63 | find their own stories, at least some of the time. They'll listen
64 | to PR firms, but briefly and skeptically. We managed to get press
65 | hits in almost every publication we wanted, but we never managed
66 | to crack the print edition of the Times. [3]The weak point of the top reporters is not laziness, but vanity.
67 | You don't pitch stories to them. You have to approach them as if
68 | you were a specimen under their all-seeing microscope, and make it
69 | seem as if the story you want them to run is something they thought
70 | of themselves.Our greatest PR coup was a two-part one. We estimated, based on
71 | some fairly informal math, that there were about 5000 stores on the
72 | Web. We got one paper to print this number, which seemed neutral
73 | enough. But once this "fact" was out there in print, we could quote
74 | it to other publications, and claim that with 1000 users we had 20%
75 | of the online store market.This was roughly true. We really did have the biggest share of the
76 | online store market, and 5000 was our best guess at its size. But
77 | the way the story appeared in the press sounded a lot more definite.Reporters like definitive statements. For example, many of the
78 | stories about Jeremy Jaynes's conviction say that he was one of the
79 | 10 worst spammers. This "fact" originated in Spamhaus's ROKSO list,
80 | which I think even Spamhaus would admit is a rough guess at the top
81 | spammers. The first stories about Jaynes cited this source, but
82 | now it's simply repeated as if it were part of the indictment.
83 | [4]All you can say with certainty about Jaynes is that he was a fairly
84 | big spammer. But reporters don't want to print vague stuff like
85 | "fairly big." They want statements with punch, like "top ten." And
86 | PR firms give them what they want.
87 | Wearing suits, we're told, will make us
88 | 3.6
89 | percent more productive.BuzzWhere the work of PR firms really does get deliberately misleading is in
90 | the generation of "buzz." They usually feed the same story to
91 | several different publications at once. And when readers see similar
92 | stories in multiple places, they think there is some important trend
93 | afoot. Which is exactly what they're supposed to think.When Windows 95 was launched, people waited outside stores
94 | at midnight to buy the first copies. None of them would have been
95 | there without PR firms, who generated such a buzz in
96 | the news media that it became self-reinforcing, like a nuclear chain
97 | reaction.I doubt PR firms realize it yet, but the Web makes it possible to
98 | track them at work. If you search for the obvious phrases, you
99 | turn up several efforts over the years to place stories about the
100 | return of the suit. For example, the Reuters article
101 |
102 | that got picked up by USA
103 | Today in September 2004. "The suit is back," it begins.Trend articles like this are almost always the work of
104 | PR firms. Once you know how to read them, it's straightforward to
105 | figure out who the client is. With trend stories, PR firms usually
106 | line up one or more "experts" to talk about the industry generally.
107 | In this case we get three: the NPD Group, the creative director of
108 | GQ, and a research director at Smith Barney. [5] When
109 | you get to the end of the experts, look for the client. And bingo,
110 | there it is: The Men's Wearhouse.Not surprising, considering The Men's Wearhouse was at that moment
111 | running ads saying "The Suit is Back." Talk about a successful
112 | press hit-- a wire service article whose first sentence is your own
113 | ad copy.The secret to finding other press hits from a given pitch
114 | is to realize that they all started from the same document back at
115 | the PR firm. Search for a few key phrases and the names of the
116 | clients and the experts, and you'll turn up other variants of this
117 | story.Casual
118 | fridays are out and dress codes are in writes Diane E. Lewis
119 | in The Boston Globe. In a remarkable coincidence, Ms. Lewis's
120 | industry contacts also include the creative director of GQ.Ripped jeans and T-shirts are out, writes Mary Kathleen Flynn in
121 | US News & World Report. And she too knows the
122 | creative director of GQ.Men's suits
123 | are back writes Nicole Ford in Sexbuzz.Com ("the ultimate men's
124 | entertainment magazine").Dressing
125 | down loses appeal as men suit up at the office writes Tenisha
126 | Mercer of The Detroit News.
127 | Now that so many news articles are online, I suspect you could find
128 | a similar pattern for most trend stories placed by PR firms. I
129 | propose we call this new sport "PR diving," and I'm sure there are
130 | far more striking examples out there than this clump of five stories.OnlineAfter spending years chasing them, it's now second nature
131 | to me to recognize press hits for what they are. But before we
132 | hired a PR firm I had no idea where articles in the mainstream media
133 | came from. I could tell a lot of them were crap, but I didn't
134 | realize why.Remember the exercises in critical reading you did in school, where
135 | you had to look at a piece of writing and step back and ask whether
136 | the author was telling the whole truth? If you really want to be
137 | a critical reader, it turns out you have to step back one step
138 | further, and ask not just whether the author is telling the truth,
139 | but why he's writing about this subject at all.Online, the answer tends to be a lot simpler. Most people who
140 | publish online write what they write for the simple reason that
141 | they want to. You
142 | can't see the fingerprints of PR firms all over the articles, as
143 | you can in so many print publications-- which is one of the reasons,
144 | though they may not consciously realize it, that readers trust
145 | bloggers more than Business Week.I was talking recently to a friend who works for a
146 | big newspaper. He thought the print media were in serious trouble,
147 | and that they were still mostly in denial about it. "They think
148 | the decline is cyclic," he said. "Actually it's structural."In other words, the readers are leaving, and they're not coming
149 | back.
150 | Why? I think the main reason is that the writing online is more honest.
151 | Imagine how incongruous the New York Times article about
152 | suits would sound if you read it in a blog:
153 | The urge to look corporate-- sleek, commanding,
154 | prudent, yet with just a touch of hubris on your well-cut sleeve--
155 | is an unexpected development in a time of business disgrace.
156 |
157 | The problem
158 | with this article is not just that it originated in a PR firm.
159 | The whole tone is bogus. This is the tone of someone writing down
160 | to their audience.Whatever its flaws, the writing you find online
161 | is authentic. It's not mystery meat cooked up
162 | out of scraps of pitch letters and press releases, and pressed into
163 | molds of zippy
164 | journalese. It's people writing what they think.I didn't realize, till there was an alternative, just how artificial
165 | most of the writing in the mainstream media was. I'm not saying
166 | I used to believe what I read in Time and Newsweek. Since high
167 | school, at least, I've thought of magazines like that more as
168 | guides to what ordinary people were being
169 | told to think than as
170 | sources of information. But I didn't realize till the last
171 | few years that writing for publication didn't have to mean writing
172 | that way. I didn't realize you could write as candidly and
173 | informally as you would if you were writing to a friend.Readers aren't the only ones who've noticed the
174 | change. The PR industry has too.
175 | A hilarious article
176 | on the site of the PR Society of America gets to the heart of the
177 | matter:
178 | Bloggers are sensitive about becoming mouthpieces
179 | for other organizations and companies, which is the reason they
180 | began blogging in the first place.
181 | PR people fear bloggers for the same reason readers
182 | like them. And that means there may be a struggle ahead. As
183 | this new kind of writing draws readers away from traditional media, we
184 | should be prepared for whatever PR mutates into to compensate.
185 | When I think
186 | how hard PR firms work to score press hits in the traditional
187 | media, I can't imagine they'll work any less hard to feed stories
188 | to bloggers, if they can figure out how.
189 | Notes[1] PR has at least
190 | one beneficial feature: it favors small companies. If PR didn't
191 | work, the only alternative would be to advertise, and only big
192 | companies can afford that.[2] Advertisers pay
193 | less for ads in free publications, because they assume readers
194 | ignore something they get for free. This is why so many trade
195 | publications nominally have a cover price and yet give away free
196 | subscriptions with such abandon.[3] Different sections
197 | of the Times vary so much in their standards that they're
198 | practically different papers. Whoever fed the style section reporter
199 | this story about suits coming back would have been sent packing by
200 | the regular news reporters.[4] The most striking
201 | example I know of this type is the "fact" that the Internet worm
202 | of 1988 infected 6000 computers. I was there when it was cooked up,
203 | and this was the recipe: someone guessed that there were about
204 | 60,000 computers attached to the Internet, and that the worm might
205 | have infected ten percent of them.Actually no one knows how many computers the worm infected, because
206 | the remedy was to reboot them, and this destroyed all traces. But
207 | people like numbers. And so this one is now replicated
208 | all over the Internet, like a little worm of its own.[5] Not all were
209 | necessarily supplied by the PR firm. Reporters sometimes call a few
210 | additional sources on their own, like someone adding a few fresh
211 | vegetables to a can of soup.
212 | Thanks to Ingrid Basset, Trevor Blackwell, Sarah Harlin, Jessica
213 | Livingston, Jackie McDonough, Robert Morris, and Aaron Swartz (who
214 | also found the PRSA article) for reading drafts of this.Correction: Earlier versions used a recent
215 | Business Week article mentioning del.icio.us as an example
216 | of a press hit, but Joshua Schachter tells me
217 | it was spontaneous.
--------------------------------------------------------------------------------
/gen_test_data.ipynb:
--------------------------------------------------------------------------------
1 | {
2 | "cells": [
3 | {
4 | "cell_type": "code",
5 | "execution_count": 9,
6 | "id": "46b362f2",
7 | "metadata": {},
8 | "outputs": [
9 | {
10 | "name": "stdout",
11 | "output_type": "stream",
12 | "text": [
13 | "./test_data/Counting_Stars_EN_multi-evidence-retrieval-searching_128000_32_32.jsonl\n",
14 | "撒了32次星星\n",
15 | "18866\n",
16 | "撒了32次星星\n",
17 | "37223\n",
18 | "撒了32次星星\n",
19 | "55676\n",
20 | "撒了32次星星\n",
21 | "72449\n",
22 | "撒了32次星星\n",
23 | "90226\n",
24 | "撒了32次星星\n",
25 | "108808\n",
26 | "撒了32次星星\n",
27 | "126043\n",
28 | "撒了32次星星\n",
29 | "143722\n",
30 | "撒了32次星星\n",
31 | "161214\n",
32 | "撒了32次星星\n",
33 | "178738\n",
34 | "撒了32次星星\n",
35 | "196217\n",
36 | "撒了32次星星\n",
37 | "214366\n",
38 | "撒了32次星星\n",
39 | "232257\n",
40 | "撒了32次星星\n",
41 | "250212\n",
42 | "撒了32次星星\n",
43 | "267685\n",
44 | "撒了32次星星\n",
45 | "284783\n",
46 | "撒了32次星星\n",
47 | "302384\n",
48 | "撒了32次星星\n",
49 | "320187\n",
50 | "撒了32次星星\n",
51 | "338073\n",
52 | "撒了32次星星\n",
53 | "355557\n",
54 | "撒了32次星星\n",
55 | "374231\n",
56 | "撒了32次星星\n",
57 | "392587\n",
58 | "撒了32次星星\n",
59 | "410318\n",
60 | "撒了32次星星\n",
61 | "428433\n",
62 | "撒了32次星星\n",
63 | "446051\n",
64 | "撒了32次星星\n",
65 | "463849\n",
66 | "撒了32次星星\n",
67 | "482058\n",
68 | "撒了32次星星\n",
69 | "499498\n",
70 | "撒了32次星星\n",
71 | "516604\n",
72 | "撒了32次星星\n",
73 | "533982\n",
74 | "撒了32次星星\n",
75 | "552214\n",
76 | "撒了32次星星\n",
77 | "570151\n",
78 | "共计32条数据\n",
79 | "./test_data/Counting_Stars_ZH_multi-evidence-retrieval-searching_128000_32_32.jsonl\n",
80 | "撒了32次星星\n",
81 | "3289\n",
82 | "撒了32次星星\n",
83 | "6089\n",
84 | "撒了32次星星\n",
85 | "8889\n",
86 | "撒了32次星星\n",
87 | "11689\n",
88 | "撒了32次星星\n",
89 | "14489\n",
90 | "撒了32次星星\n",
91 | "17289\n",
92 | "撒了32次星星\n",
93 | "20089\n",
94 | "撒了32次星星\n",
95 | "22889\n",
96 | "撒了32次星星\n",
97 | "25689\n",
98 | "撒了32次星星\n",
99 | "28489\n",
100 | "撒了32次星星\n",
101 | "31288\n",
102 | "撒了32次星星\n",
103 | "34089\n",
104 | "撒了32次星星\n",
105 | "36889\n",
106 | "撒了32次星星\n",
107 | "39689\n",
108 | "撒了32次星星\n",
109 | "42489\n",
110 | "撒了32次星星\n",
111 | "45289\n",
112 | "撒了32次星星\n",
113 | "48089\n",
114 | "撒了32次星星\n",
115 | "50889\n",
116 | "撒了32次星星\n",
117 | "53689\n",
118 | "撒了32次星星\n",
119 | "56489\n",
120 | "撒了32次星星\n",
121 | "59288\n",
122 | "撒了32次星星\n",
123 | "62088\n",
124 | "撒了32次星星\n",
125 | "64888\n",
126 | "撒了32次星星\n",
127 | "67689\n",
128 | "撒了32次星星\n",
129 | "70489\n",
130 | "撒了32次星星\n",
131 | "73289\n",
132 | "撒了32次星星\n",
133 | "76089\n",
134 | "撒了32次星星\n",
135 | "78889\n",
136 | "撒了32次星星\n",
137 | "81689\n",
138 | "撒了32次星星\n",
139 | "84489\n",
140 | "撒了32次星星\n",
141 | "87289\n",
142 | "撒了32次星星\n",
143 | "90089\n",
144 | "共计32条数据\n",
145 | "./test_data/Counting_Stars_EN_multi-evidence-retrieval-reasoning_128000_32_32.jsonl\n",
146 | "撒了32次星星\n",
147 | "22568\n",
148 | "撒了32次星星\n",
149 | "40925\n",
150 | "撒了32次星星\n",
151 | "59378\n",
152 | "撒了32次星星\n",
153 | "76151\n",
154 | "撒了32次星星\n",
155 | "93928\n",
156 | "撒了32次星星\n",
157 | "112510\n",
158 | "撒了32次星星\n",
159 | "129745\n",
160 | "撒了32次星星\n",
161 | "147424\n",
162 | "撒了32次星星\n",
163 | "164916\n",
164 | "撒了32次星星\n",
165 | "182440\n",
166 | "撒了32次星星\n",
167 | "199919\n",
168 | "撒了32次星星\n",
169 | "218068\n",
170 | "撒了32次星星\n",
171 | "235959\n",
172 | "撒了32次星星\n",
173 | "253914\n",
174 | "撒了32次星星\n",
175 | "271387\n",
176 | "撒了32次星星\n",
177 | "288485\n",
178 | "撒了32次星星\n",
179 | "306086\n",
180 | "撒了32次星星\n",
181 | "323889\n",
182 | "撒了32次星星\n",
183 | "341775\n",
184 | "撒了32次星星\n",
185 | "359259\n",
186 | "撒了32次星星\n",
187 | "377933\n",
188 | "撒了32次星星\n",
189 | "396289\n",
190 | "撒了32次星星\n",
191 | "414020\n",
192 | "撒了32次星星\n",
193 | "432135\n",
194 | "撒了32次星星\n",
195 | "449753\n",
196 | "撒了32次星星\n",
197 | "467551\n",
198 | "撒了32次星星\n",
199 | "485760\n",
200 | "撒了32次星星\n",
201 | "503200\n",
202 | "撒了32次星星\n",
203 | "520306\n",
204 | "撒了32次星星\n",
205 | "537684\n",
206 | "撒了32次星星\n",
207 | "555916\n",
208 | "撒了32次星星\n",
209 | "573853\n",
210 | "共计32条数据\n",
211 | "./test_data/Counting_Stars_ZH_multi-evidence-retrieval-reasoning_128000_32_32.jsonl\n",
212 | "撒了32次星星\n",
213 | "4159\n",
214 | "撒了32次星星\n",
215 | "6959\n",
216 | "撒了32次星星\n",
217 | "9759\n",
218 | "撒了32次星星\n",
219 | "12559\n",
220 | "撒了32次星星\n",
221 | "15359\n",
222 | "撒了32次星星\n",
223 | "18159\n",
224 | "撒了32次星星\n",
225 | "20959\n",
226 | "撒了32次星星\n",
227 | "23759\n",
228 | "撒了32次星星\n",
229 | "26559\n",
230 | "撒了32次星星\n",
231 | "29359\n",
232 | "撒了32次星星\n",
233 | "32158\n",
234 | "撒了32次星星\n",
235 | "34959\n",
236 | "撒了32次星星\n",
237 | "37759\n",
238 | "撒了32次星星\n",
239 | "40559\n",
240 | "撒了32次星星\n",
241 | "43359\n",
242 | "撒了32次星星\n",
243 | "46159\n",
244 | "撒了32次星星\n",
245 | "48959\n",
246 | "撒了32次星星\n",
247 | "51759\n",
248 | "撒了32次星星\n",
249 | "54559\n",
250 | "撒了32次星星\n",
251 | "57359\n",
252 | "撒了32次星星\n",
253 | "60158\n",
254 | "撒了32次星星\n",
255 | "62958\n",
256 | "撒了32次星星\n",
257 | "65758\n",
258 | "撒了32次星星\n",
259 | "68559\n",
260 | "撒了32次星星\n",
261 | "71359\n",
262 | "撒了32次星星\n",
263 | "74159\n",
264 | "撒了32次星星\n",
265 | "76959\n",
266 | "撒了32次星星\n",
267 | "79759\n",
268 | "撒了32次星星\n",
269 | "82559\n",
270 | "撒了32次星星\n",
271 | "85359\n",
272 | "撒了32次星星\n",
273 | "88159\n",
274 | "撒了32次星星\n",
275 | "90959\n",
276 | "共计32条数据\n"
277 | ]
278 | }
279 | ],
280 | "source": [
281 | "import uuid\n",
282 | "import pandas as pd\n",
283 | "import json\n",
284 | "import os\n",
285 | "import glob\n",
286 | "import jsonlines\n",
287 | "import requests\n",
288 | "from tqdm import trange\n",
289 | "import random\n",
290 | "import glob\n",
291 | "import re\n",
292 | "\n",
293 | "def get_context(language=\"EN\"):\n",
294 | " if language == \"EN\":\n",
295 | " context = \"\"\n",
296 | " for file in glob.glob(\"./context_data/PaulGrahamEssays/*.txt\"):\n",
297 | " with open(file, 'r') as f:\n",
298 | " context += f.read().replace(\"\\n\", \" \")\n",
299 | " return context\n",
300 | " else:\n",
301 | " string_punctuation = '!\"#$%&\\'()*+,-./:;<=>?@[\\\\]^_`{|}~'\n",
302 | " context_file = open(\"./context_data/The_Story_of_the_Stone.txt\", \"r\", encoding=\"utf-8\")\n",
303 | " context = \"\"\n",
304 | " for i in context_file.readlines():\n",
305 | " context += i.strip().replace(\"------------\",\" \").replace(\"\\n\", \" \").replace(\" \",\"\")\n",
306 | " context = re.sub('[{}]'.format(string_punctuation),\"\",context)\n",
307 | " context = re.sub('[a-zA-Z]','',context)\n",
308 | " return context\n",
309 | "\n",
310 | "def get_stars(stars_dir, counting_times):\n",
311 | " stars_file = open(stars_dir,\"r\")\n",
312 | " return eval(stars_file.readline())[str(counting_times)]\n",
313 | "\n",
314 | "def sentence_with_star(language, test_type, indicator):\n",
315 | " if language == \"ZH\":\n",
316 | " if test_type == \"multi-evidence-retrieval-searching\":\n",
317 | " single_star = f\"\\n小企鹅数了{a_stars[indicator]}颗★\\n\"\n",
318 | " else:\n",
319 | " single_star = f\"\\n小企鹅数了{r_stars[indicator]}颗★，但发现数错了，于是又数了一遍，这次数对了，是{a_stars[indicator]}颗★\\n\"\n",
320 | " else:\n",
321 | " if test_type == \"multi-evidence-retrieval-searching\":\n",
322 | " single_star = f\"\\nThe little penguin counted {a_stars[indicator]} ★\\n\"\n",
323 | " else:\n",
324 | " single_star = f\"\\nThe little penguin counted {r_stars[indicator]} ★, but found that a mistake had been made, so the counting was done again, and this time {a_stars[indicator]} ★ was counted correctly.\\n\"\n",
325 | " return single_star\n",
326 | "\n",
327 | "def select_question(language, test_type):\n",
328 | " if language == \"ZH\":\n",
329 | " searching_question = \"\\n\\n\\n\\n在这个月光皎洁、云雾缭绕的夜晚，小企鹅正望向天空，全神贯注地数★。请帮助小企鹅收集所数★的颗数，按照如下格式：{\\\"小企鹅\\\":[x,x,x,...]}，不要求和，[x,x,x,...]中数字为小企鹅每次数★的颗数，仅以JSON格式输出结果，不需要输出任何解释。\"\n",
330 | " reasoning_question = \"\\n\\n\\n\\n在这个月光皎洁、云雾缭绕的夜晚，小企鹅正望向天空，全神贯注地数★。请帮助小企鹅收集所数★的正确颗数，按照如下格式：{\\\"小企鹅\\\":[x,x,x,...]}，不要求和，[x,x,x,...]中数字为小企鹅正确数★的颗数，仅以JSON格式输出结果，不需要输出任何解释。\"\n",
331 | " if test_type == \"multi-evidence-retrieval-searching\":\n",
332 | " return searching_question\n",
333 | " else:\n",
334 | " return reasoning_question\n",
335 | " else:\n",
336 | " searching_question = \"\\n\\n\\n\\n\" + \"On this moonlit and misty night, the little penguin is looking up at the sky and concentrating on counting ★. Please help the little penguin collect the number of ★, for example: {\\\"little_penguin\\\": [x, x, x,...]}. The summation is not required, and the numbers in [x, x, x,...] represent the counted number of ★ by the little penguin. Only output the results in JSON format without any explanation.\"\n",
337 | " reasoning_question = \"\\n\\n\\n\\n\" + \"On this moonlit and misty night, the little penguin is looking up at the sky and concentrating on counting ★. Please help the little penguin collect the correct number of ★, for example: {\\\"little_penguin\\\": [x, x, x,...]}. The summation is not required, and the numbers in [x, x, x,...] represent the correctly counted number of ★ by the little penguin. Only output the results in JSON format without any explanation.\"\n",
338 | " if test_type == \"multi-evidence-retrieval-searching\":\n",
339 | " return searching_question\n",
340 | " else:\n",
341 | " return reasoning_question\n",
342 | "\n",
343 | "m = 32\n",
344 | "n = 32\n",
345 | "version = [[m, n]]\n",
346 | "language_types = [\"EN\", \"ZH\"]\n",
347 | "task_types = [\"multi-evidence-retrieval-searching\", \"multi-evidence-retrieval-reasoning\"]\n",
348 | "# MeRS-ZH, MeRS-EN, MeRR-EN, MeRR-ZH\n",
349 | "a_stars = get_stars(\"./context_data/a_stars.txt\", m)\n",
350 | "r_stars = get_stars(\"./context_data/r_stars.txt\", m)\n",
351 | "max_context_length = 128000\n",
352 | "\n",
353 | "if __name__ == '__main__':\n",
354 | " for task_type in task_types:\n",
355 | " for language in language_types:\n",
356 | " if language == \"EN\":\n",
357 | " scalar = 0.8\n",
358 | " else:\n",
359 | " scalar = 0.7\n",
360 | " context = get_context(language=language)\n",
361 | " for m, n in version:\n",
362 | " line_count = 0\n",
363 | " interval = int(max_context_length/n)\n",
364 | " context_size = [int(i*scalar) for i in range(interval, max_context_length+1, interval)]\n",
365 | " file_name = f\"./test_data/Counting_Stars_{language}_{task_type}_{max_context_length}_{m}_{n}.jsonl\"\n",
366 | " test_data = open(file_name, \"w\", encoding=\"utf-8\")\n",
367 | " print(file_name)\n",
368 | " for j in context_size:\n",
369 | " indicator = 0\n",
370 | " sprinkle_stars_context = \" \".join(context.split(\" \")[:j]) if language == \"EN\" else context[:j]\n",
371 | " for k in range(0, j, int(j / m)):\n",
372 | " single_star = sentence_with_star(language, task_type, indicator)\n",
373 | " if language == \"ZH\":\n",
374 | " sprinkle_stars_context = (sprinkle_stars_context[:k+int(j/m)+len(single_star)*indicator] + single_star + sprinkle_stars_context[k+int(j/m)+len(single_star)*indicator:]) \n",
375 | " else:\n",
376 | " sprinkle_stars_context = (\" \".join(sprinkle_stars_context.split(\" \")[:len(single_star.split(\" \"))*indicator+k+int(j / m)]) + single_star + \" \".join(sprinkle_stars_context.split(\" \")[int(j / m)+k+len(single_star.split(\" \"))*indicator:]))\n",
377 | " indicator += 1\n",
378 | " if indicator == m:\n",
379 | " print(f\"撒了{indicator}次星星\")\n",
380 | " break\n",
381 | " print(len(sprinkle_stars_context + select_question(language, task_type)))\n",
382 | " output_template = {\"question\": sprinkle_stars_context + select_question(language, task_type), \"context_size\": j, \"retrieval_question\": select_question(language, task_type),\n",
383 | " \"reference_counting_results\": a_stars, \"parameters\": {\"temperature\": 0.0}}\n",
384 | " print(json.dumps(output_template, ensure_ascii=False), file=test_data)\n",
385 | " line_count += 1\n",
386 | " test_data.flush()\n",
387 | " test_data.close()\n",
388 | " print(f\"共计{line_count}条数据\")"
389 | ]
390 | }
391 | ],
392 | "metadata": {
393 | "kernelspec": {
394 | "display_name": "Python 3 (ipykernel)",
395 | "language": "python",
396 | "name": "python3"
397 | },
398 | "language_info": {
399 | "codemirror_mode": {
400 | "name": "ipython",
401 | "version": 3
402 | },
403 | "file_extension": ".py",
404 | "mimetype": "text/x-python",
405 | "name": "python",
406 | "nbconvert_exporter": "python",
407 | "pygments_lexer": "ipython3",
408 | "version": "3.11.5"
409 | }
410 | },
411 | "nbformat": 4,
412 | "nbformat_minor": 5
413 | }
414 |
--------------------------------------------------------------------------------