81 comments

[ 1.7 ms ] story [ 122 ms ] thread
This is good, and reminds me of the "Fake Paper Generator" that some MIT students made using context free grammars in 2005: https://pdos.csail.mit.edu/archive/scigen/

Unfortunately it doesn't work anymore, but the titles it generated for the papers were all plausible but also ridiculous CS paper titles, e.g. Rooter: A Methodology for the Typical Unification of Access Points and Redundancy

(comment deleted)
"Glassdoor: Employees have the right to be anonymous, but the CEO doesn't"

Hard to argue with that logic!

“Progress on addressing the massive crisis in electrical chargers”

I’m glad to hear it!

Where’s “Tiny X written in Y”?
> Show HN: Rhapsody.fm – Send and receive messages using Discord

What an invention!

Personal favourite:

Mark Zuckerberg: “Haters are supporters; hate is a solution”

I came across this one: "Why I Love Vapor (2018)"
Mine was “Elon Musk’s armies are about to begin the slaughter”
It’s not GPT3 it’s just posting from the future!
Interested in the process for fine tuning gpt3 for this
Super easy, just took 10k titles/comments/points from the Angolia API, formatted them as JSON Lines like the following with jq, and fed them to the very well built and documented openai CLI.

{"prompt": "A plausible Hacker News title:", "completion": " The Feynman Lectures on Physics (1964) (280 points, 62 comments) END"}

The space at the beginning of the completion is for tokenizing, and the END token is for use as a stop token in the generations.

Looked away for a while talking to someone, tried to click a link when I looked back :)

Also unrelated but good one "Tether issue: under $1B and all of the tether tokens suspended". This would definitely get my click :)

I didn’t need to look away to click a link. Spoiler: it doesn’t work, nor do the comments
I absolutely want to read some of these. There needs to be a way to for people to post and rank the best matching articles that actually exist. I dream of the day that generative AI is repurposed to act as a more effective search engine.
Well now I need to know more about the "Mystery of the frozen frog latin speaker (2012)"
Ask HN: Is there another HN beginning with a “h”?$

Show HN: Daily XLSX to-do list with attached spreadsheet as material*

Banning JavaScript from web pages is bad for the user

Has Google become too social?

SQLite development from scratch from scratch

Armor-piercing lasers are not shooting lasers but missiles

We made a public blockchain off-chain

Apple sued for pricing user data against provider who did not provide refunds

100% Embarrassing Haskell Builds

Heroku Compose is not fit for purpose

Pain Enhancer

The distance between reality and fantasy is grows ever smaller.

It averages around 0.0005$ per request according to the footer. Could this end up costing the author quite a bit due to HN traffic? Also, whats stopping bad actors from writing a script to continuously fetch the page?

I wonder if some sort of caching might help lower costs.

Nothing. And in fact that’s exactly what happened to me. Some fella from HN spawned like 45 simultaneous wget’s in a loop to cause maximum financial damage. All of a sudden we see Firebase’s cost graph go vertical.

It happened after I mentioned “just be kind, please! Theoretically this could cost a lot of money.”

So there’s at least one person who will do exactly this just for fun.

Firebase customer support was super cool about it, but it still knocked us off the paid tier.

>All of a sudden we see Firebase’s cost graph go vertical.

think of all the money you have saved by not having to hire a hundred engineers to maintain your website's infrastructure!

It's part of the bit.

There's intentionally no caching, every batch is warm from the AI oven.

HN usually drives around 10k visits, so organically it's going to be well within my Saturday night budget. If someone decides to hammer it, well, the OpenAI account has a hard limit of $20/month. It will live until it's killed I guess.

> HN usually drives around 10k visits

You have to account for this not being a regular visit, but rather (I guess) 5-50 "visits" per visit, as people mash F5 to get more titles.

> OpenAI account has a hard limit of $20/month

That makes me feel better about the couple refreshes I did myself :).

“The experience of a young black man in a white prison (2020)“
Oh god, I sense this thing is trying to make us write essays to back up the titles.

My favorite so far: "Why isn't Linux perfect?"

I got "Why encryption doesn’t work", which does feels like an essay which should exist, if it doesn't already.
Is there a dataset of HN titles? This made me want to fiddle with this, but step one is to get the data, and I don't want to crawl HN if the data has already been collected.
There's an API[0] but it's frustratingly limited in capabilities (albeit not rate-limited.) You'll have to iterate all post IDs, download each post as JSON and get the titles that way.

There's also a Google dataset but I don't know the URL for it or if it's up to date.

[0]https://github.com/HackerNews/API

> frustratingly limited in capabilities

What's missing precisely? Seems to be good enough for every use I could think of.

One time I even downloaded every single item from it, with a threaded fetcher of I think 16 threads, iterating from 1 up to latest ID and it was done in some like 2 hours I think.

No ability to directly download threads with a single request, for one, or query it like a database to sort or filter results, exclude unwanted fields, etc.
Those things should be trivial to achieve with most general purpose languages, as the API is so simple. No need for pagination or other things, just request things by ID recursively and you get the full thing, then after than filter/select whatever you want.

Pseudo-code to show how simple it would be:

    function get_thread(id) {
      let item = http.get(`{api}/?id={id}`).body
      if item.childs {
        item.fetched_children = item.childs.map((id) => {
          return get_thread(id)
        })
      }
      return item
    }
(untested, but you don't really need more than that, besides checking if the item was deleted)
Yes, and i've done it. But you still wind up having to make a separate request for each item, which makes building threads incredibly slow. It's also a waste of time if you're filtering out anything, because you still have to make the request and download the item to filter it out.

Which is why it would be preferable for the API itself to support these features.

Can you somehow fine-tune GPT3 on a dataset? I just assumed the OP generated them using a prompt like "top hacker news threads" or somehing like that.
Ask HN: How do you limit your son's software interest?

Show HN: A tool to quickly improve your skills

Amazing stuff.

This has the disconcerting effect of making the real HN titles look like they're generated by GPT-3.
Eventually we will get AI based title analysers that simulate the conditions to predict and optimise where your title will land on the front-page listing order.
I think a case could be made that ML is automated growth hacking.
When there is no new content on HN, I'm gonna refresh this one.
"Ask HN: Is HN worth caring about?" is by far the best one I've gotten.

"James Webb Telescope Recall" was pretty amusing as well.

> The Impossible History of Soil Temperature in the Earth’s Atmosphere

This one seems too good to be AI

Superb and timely. Although "Silicon Valley’s single biggest problem is that it’s the clearest way of doing things" is just garbled.
"Ask HN: I'm working full time on my blog and still have not found a sustainable way to make money"
I refuse to believe this one wasn’t snuck in to the outputs by a human:

> Tesla Co-CEO says the autonomous vehicle ‘is for chancers’

got another one about this Tesla Co-CEO

> Tesla Co-CEO is trying to raise $800M for T4, but details are sketchy

"Pornhub was around for 18 years before people started thinking about it [video]"