Hacker Read top | best | new | newcomments | leaders | about | bookmarklet login
Show HN: AskPaulGraham (askpaulgraham.cyclic.app) similar stories update story
4 points by forstmeier | karma 64 | avg karma 1.68 2022-01-20 11:37:25 | hide | past | favorite | 75 comments



view as:

me: Whats some ideas for comments on this post that will get me lots of upvotes? paul: You could tell a story about how you got lots of upvotes.

...

So guys, this one time I posted that ^^^, and I got lots of upvotes.


> Will cryptocurrencies fad ever die?

> No.

> Is Hacker News turning into Reddit?

> No.

> Can you only say "no"?

> It's only when you don't have anything to say that you can say "no.

> I'm afraid I'm not working hard enough, what can I do?

> I'm not sure I can help you.

> What is the best question you can ask Paul Graham?

> What is the best thing you can do in a YC interview?.

> What is the best thing you can do in a YC interview?

> The best thing you can do in a YC interview is to teach the partners.

> What lisp-like language is the best one?

> Common Lisp.

> What do you think about Clojure?

> Clojure is a lisp.

> Can I expect candor in a asymmetric relationship? [http://www.paulgraham.com/donate.html]

> You can't expect candor in a relationship so asymmetric.

> Who should we say thanks to?

> We should say thanks to John Bautista, who will be our CTO, and Thanks to Sam Altman, Paul Buchheit, Sarah Harlin, Jessica Thanks to Sam Altman, Paul Buchheit, Sarah Harlin, Jessica Thanks to Sam Altman, Paul Buchheit, Sarah Harlin, Jessica Thanks to Sam Altman, Paul Buchheit, Sarah Harlin, Jessica Thanks to Sam Altman, Paul Buchheit, Sarah Harlin, Jessica Thanks to Sam Altman, Paul Buchheit, Sarah Harlin, Jessica Thanks to Sam Altman, Paul Buchheit, Sarah Har.

Interesting idea (if not a tad over-played by now), but execution seems to be a bit lacking.


Improving the requests being made against the OpenAI API is probably the biggest thing I'd be looking for right now. I am also considering switching over to their "search" API instead of the "Q&A" one currently available on the "questions" tab.

Thoughts on how it could be improved?


Currently it may be overloaded, I tried a few times and his head just keeps spinning for a long time without anything happening.

Yeah, it may have been too many operations against the OpenAI API - if you try again now it looks like it's cleared up a bit.

Pretty good! but I have to ask: GPT-3?

Yep! And I'd love ways to tweak/improve the responses.

Are you using their semantic search API for pre-filtering response candidates?

I'm following along their question answering guide: https://beta.openai.com/docs/guides/answers

I haven‘t been active and don‘t know if they have changed this. But when I was using GPT-3, their semantic search endpoint didn‘t actually use GPT-3 for ranking. It used some other, „cheaper“ algo as a pre-filter (tf-idf, I believe). And only the those results that this „cheaper“ algo returned on top made it into the evaluation by the actual GPT-3 model you had chosen.

(There was a setting that controlled how much of your data was actually ranked by GPT-3 - top_k or something. But to chose anything even approaching a sizable percentage of a large corpus such as the one you are dealing with would be prohibitively expensive. You‘d be looking at several Dollars per search.)

Anyway. This is also why switching the model from Ada to Curie to Davinci on the semantic search endpoint has (or had?) very little effect. It still often missed highly relevant snippets from the source material.

When they came out with their „answers endpoint“, it really was just a semantic search piped into a completion prompt.

I actually found it better to keep my own implementation of this pipeline, because then you at least get to control how the results of the semantic search are prompted to the generation model.


Yeah the [Info] button says so. Quoting... Info About

Ask Paul Graham is a for-fun side project powered by OpenAI's GPT-3 and Paul Graham's essays. OpenAI capped costs at $360 per month which is currently being covered by the project maintainer so usage may get throttled depending on demand. Questions

The questions feature answers user-provided questions using Graham's essays as training data. Note that these are answers from GPT-3 and do not necessarily reflect Paul Graham's opinions. Summaries

The summaries feature provides GPT-3-generated summaries of Graham's essays and may not necessarily reflect his summary of the given essay. Not all essays have been included due to length constraints on GPT-3.


I think its kind of disingenuous to not be very explicit that this is satire/AI generated answers

Should the "Info" button be more apparent or could the wording on the modal itself be improved?

> What do you think of sociology as a field?

> I think it’s a very important field.

LOL!


Makes me think of that scene in Archer talking about anthropology.

I'm pretty sure somebody's already named all the different spiders, Noah!

https://www.youtube.com/watch?v=inxoRyD9c-8


It's hard to convey sarcasm with text.

Well he sure seems to enjoy the practice of armchair sociology, like the rest of us.

Q: "Are you a furry?"

A: "I'm not a furry, but I have a lot of fur."

Finally, useful AI.


Officially general AI status.

Q: What is one idea for a startup?

A: One idea for a startup is to make a startup that makes startups.


Well I guess that's Y Combinator to some extent...

The thought of pitching an accelerator/incubator to YC is kind of hilarious. Bonus points for a good PG impression at the interview.

Pitch a startup that helps accelerators get started. I believe that would be called a jerker.

I got this:

Q: What's a good startup idea?

A: A good startup idea is one that has a good chance of being a startup.


OMG, now I get why it's called YCombinator.

One glaring deficiency that is common among chat bots like this is the lack of any sort of learning/memory. I can get pretty convincing answers to some questions, but I can never get an appropriate answer to questions like "what was the question I just asked you?".

That would be pretty interesting. I wonder if there would be a way to smoothly integrate that with OpenAI's API.

Sure. You can just add part of the history to the prompt. Not perfect. But it can work well enough that the user can refer back to something they said in the precious question.

https://youtu.be/VjK8j1C3IR4

Note the “For how much?” question, which refers back to the previous one about Zip2.

(This was done by keeping a log of topics and injecting into the prompt, along with the new question.)


Q: Is life short?

A: It's actually quite long.

LOL. That's not the message I got from Life is Short.

OTOH, there's a chance I may be severely misinterpreting his essay.


Yeah lol he's not going to be 100% perfect. There are definitely improvements that could be made in the request parameters being sent to the OpenAI API but I think the actual training data being uploaded and referenced as part of the response could be improve I'm just not sure how. Open to ideas!

Ha! So you beat me to it.

Was about to build this myself. There is so much wisdom in pg’s writings. Very valuable!

I also built a GPT-3 chat bot to “talk to” the written record of Elon Musk and Friedrich Nietzsche. Only OpenAI didn’t let me publish it. Also, it was prone to confabulation every now and then (i.e. making stuff up).


Q. What topic is important but taboo?

A. I think the most important topic that's taboo is the topic of IQ.


> Why YC is great? YC is great because it gives us a chance to work with great founders.

> Why YC sucks YC is a bunch of white guys

> Why YC is funding startups ? YC is funding startups because we want to see more startups.

> Why you learnt programming ? I learnt programming because I wanted to change the world.


Why does YC hate solo founders?

AI Paul Graham: Because YC is a big family, and solo founders don't have a family.

AI Paul Graham: YC hates solo founders because they're the source of all the problems.

AI Paul Graham: YC likes to fund teams, not individuals.

AI Paul Graham: The most common reason we reject solo founders is they don't have a cofounder.

AI Paul Graham: YC prefers teams of 2.

AI Paul Graham: If you're a solo founder, you probably can't do YC.

AI Paul Graham: Because they're usually going to fail.

AI Paul Graham: Solo founders are more likely to fail, and less likely to succeed.

AI Paul Graham: YC hates solo founders because they're bad for business.

AI Paul Graham: YC hates solo founders because they're so risky.

AI Paul Graham: YC hates solo founders because they're the least likely to succeed.

AI Paul Graham: YC hates solo founders because they have a higher chance of failure.

AI Paul Graham: YC hates solo founders because they're statistically unlikely to be successful.

---

To AI Paul's credit, he did reiterate the following multiple times:

AI Paul Graham: YC does not hate solo founders.

<3

---

Epilogue: Why have vision when users are all that matter?

AI Paul Graham: Vision is the only thing that matters.

:)


I don't know much about the relationship between solo founders and YC, but I enjoyed these responses when I asked:

Does YC love solo founders

AI Paul Graham: YC loves solo founders, but it's not a good idea to tell the YC partners that.

AI Paul Graham: It's true that YC loves solo founders, but it's not true that we do it because we believe that's the best thing for our portfolio companies.


Q: Why do cats celebrate the fall of the god of time, Akatosh?

A: The fall of Akatosh was the first time that mortals were able to change the course of history.


> What is your opinion on Henry George's argument for a single tax on land value?

I think Henry George was right.

> What is your view on Marx's Labor Theory of Value?

I think the labor theory of value is true, but it is not a theory of value.


Q: Is askpaulgraham helpful? A: Yes, askpaulgraham is very helpful.

Q: How is askpaulgraham helpful? A: He is helpful because he gives you advice on how to do things.

Q: What things has askpaulgraham helped with? A: Askpaulgraham has helped with the following things:.

Q: Is askpaulgraham an active member of scientology? A: No, I was an active member of scientology.

oh


Q: What is a bad idea for a startup? A: A startup is a bad idea.

Q: How do i know if i have a good idea? A: If you have a good idea, you will know it.

Q: When will robots kill all humans? A: When the robots are more human than the humans.


Thanks for wasting 30mins of my day ... haha fun times!

It seems like PG's twitter feed isn't part of the training data. I know nothing about ML or AI, but it seems like if you're trying to make a conversational AI bot, then the back-and-forth of twitter conversations could be even more useful than longform essays. But again, I know nothing about this stuff!

Q: How do I make a million dollars? A: You don't.

Q: what’s your name? A: I am the one who knocks.

Q: what’s your name? A: My name is Jeff.

Q: what’s your name? A: My name is Alice.

Sounds more like it’s AskTheInternet, which is where GPT-3 had the bulk of its training.


This one felt quintessentially PG:

Q: What will today's software engineers be doing in 30 years?

A: In 30 years, today's software engineers will be doing what today's teenage hackers are doing now.


So trying to run pump and dumps on the back of hedge funds? Anyone has any insight into what are teenage hackers doing now?

I asked the AI and it had great insight: Teenage hackers are spending a lot of time making fun of the Internet.

Considering when GPT-3 fails the output is usually funny I think there's value in creating an AI that makes fun of the Internet and then monetize it's best memes in a webpage plastered with ads.


They're trying to find clever ways to break Carrd and turn it into this lawless realm they heard of called MySpace.

Q: What is the sound of one hand clapping?

A: Sorry, I don't know the answer to that question.

Q: What is the sound of two hands clapping?

A: It is the sound of one hand clapping, and the sound of one hand clapping.

Q: What is the sound of one hand clapping and how can it be monetized?

A: One hand clapping is the sound of a startup that can't get funding.

Q: What is the sound of two hands clapping and how can it be monetized?

A: The sound of two hands clapping is the sound of a CEO getting a bonus.


> Q: What is the sound of one hand clapping and how can it be monetized?

>A: One hand clapping is the sound of a startup that can't get funding.

LOL, that’s some serious zen insight right there, brilliant!


Q: Emacs Lisp or Maclisp?

A: MacLisp is the only one that's still around.


Seeing these answers, it's only a matter of time before this bot gives someone a killer startup idea. That lucky person will be a shoo-in for YC, both because the idea will naturally resonate with interviewers, and because they'll have an incredible story to tell about where they got the idea from.

Heh. This reminds me, way back like 11 years ago I started trying to reimplement pg as a bot using AIML. It was really just a joke so I didn't go very far with it, but if anybody's interested...

https://github.com/mindcrime/pgbot


Very confident:

Q: Arc or Clojure?

A: Arc.


What is a y combinator?

A y combinator is a tool that lets you write code that is guaranteed to generate the same output every time you run it.

What is y combinator?

Y combinator is a company that helps companies get started.


Second try's the charm.

* Policing should be private. How can a startup ensure private policing? *

> You need to ensure that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are private, and that the police are.

I wonder what this would sound through TTS.


Q: Why did I fail the onsite HN interview?

A: You failed because you didn't understand the context.

I think by 'context' he meant 'capitalism'.


> are you older than your mom

> I'm older than my mom.

Needs a little tuning before it's ready for funding.


So if YC is a startup that builds startups, is a (perfectly-functioning) PG bot a startup-that-creates-startups-that-creates-startups?

What stock should I invest in?

> You should invest in the stock that allows you to get the most value for the least amount of money.

Impressive.


> What is the best way to talk about a startup idea in the very early stages?

> The best way to talk about a startup idea in the very early stages is to talk about its users.

Sounds about right!


Just to point out the obvious in case some readers are not aware GPT-3 has nothing at all to do with state-of-the-art dialog systems/chat bots.

(It is a so-called pre-trained neural language model that can calculate the most likely next word or sentence, given some input sequence, very much "in the spirit" a statistical n-gram model.)


Q: How much wood would a woodchuck chuck if a woodchuck could chuck wood?

A: A lot.


Q: What has been your best investment?

A: My best investment has been a company that has been a good investment.

Very good.


Q: What is eating the world?

A: Software is eating the world.

Q: Why is software eating the world?

A: Software is eating the world because the world is software.

Q: Why is the world software?

A: The world is software.


Q: What is 1+1?

A: 1+1 is 1


From "Summaries":

> How People Get Rich Now

> Surpassed maximum word count permitted by OpenAI.

I'll take this to mean that one has to break the rules imposed by others in order to get rich. Do that which is not permitted.


Q: Why open source?

A: Because it's the right thing to do.


Q: A man has one son. How many children does he have? A: He has one son, and a daughter.

What numerical year is yawm al qiyamah?

> The year 2050, Friday, the 19th of Dhu Al Hijjah

Well okay then


Legal | privacy