It's not their point as stated in the description, but it is the point inferred from the title alone. "How much of Hacker News is about AI?" would have been better.
Would there be some way to see what percentage of the front page meets those qualifications? I don't browse /new very often, so I have less of an idea of how representative that is. It would be interesting to see if they trend similarly.
Yeah, that's probably the main limitation right now. This is measuring /new, so it's every submission rather than just the stuff that makes the front page.
lcamtuf's survey (linked in the footer) was looking at the front page, and the number there was quite a bit higher. So there's likely some effect from voting, although I haven't dug into that yet.
I think Algolia has a way to filter for front-page stories, which should make it pretty easy to add. I haven't tested that yet though but will certainly look into it!
Multiple times authors have been asked about AI usage on HN, because readers (rightfully) suspected it was AI written. I don't have a problem with AI written articles, or assisted.
But then the authors lie or make ludicrous excuses for why their article "sounds" like AI. Then I check with Pangram and yep. It is completely AI written.
You get a few free checks on Pangram each day or so. Just start checking and encourage people to be honest.
AI has its place. Lying about it and misrepresenting it makes everyone dislike AI more, even as they use it themselves.
> Then I check with Pangram and yep. It is completely AI written.
“AI detectors” do NOT work reliably.
Even humans who are ever-so-certain about being able to identify AI usage constantly embarrass themselves when that claimed ability is put to the test. Look up “AI Monet Instagram” for a social experiment where hundreds of cocksure users pointed out “AI signs” in what was claimed to be an AI-generated painting in the style of Monet, but was actually a genuine Monet.
Every technical topic involves AI now. Even if it's about why the ancient Greeks didn't turn their steam engines to industrial use - the AI's should be researching the details. We might as well ask which topics involve humans or use language.
The Show HN volume went 805/month to 4,199/month over that same time period, so from 3.2% of all submissions to 13.4%, while HN overall only grew 1.3x.
So the flat ~24% held through a 5x expansion in the denominator, which makes it
a lot more interesting honestly.
Not a bad idea. I'd look at creating a separate page for embeddings. Crypto and NFT's would be the next best candidates to add. Something for me to work on over time.
An estimated 2,000 articles per month at 2,000 words per article would result in 4,000,000 words, which costs either $200 or $2000 to process at Pangram [1]. It might make life a lot more enjoyable for many users here, though.
Several people have asked for this already, so I'll add it to my to-do list. I just want to be careful about flagging posts that appear to be written by AI. Perhaps I set certain thresholds of certainty to filter by. What I have currently is easily verified and matched by the words we are filtering for.
It is also interesting to note that I have already received 5 AI written emails to the email I linked in the footer.
Something to maybe track later would be /newcomments, but that would only grab the most obvious comments. People can certainly direct an AI to write in a style that sounds human, but I'm sure it'll still catch a fair amount.
If you read a lot, you start to get a feel for phone typos vs AI hiccups. I can guarantee that is post was made by a brain designed to eat berries in a cave. Almost all the another posts look a bit scary.
I think I misunderstood - how many posts were about AI vs how many posts were written by AI.
The topic of an article being AI and an "AI-Generated Article" are two completely different things.
Agreed. "How much of Hacker News is about AI?" would have been a more accurate title.
I don't believe that is their point.
It's not their point as stated in the description, but it is the point inferred from the title alone. "How much of Hacker News is about AI?" would have been better.
Would there be some way to see what percentage of the front page meets those qualifications? I don't browse /new very often, so I have less of an idea of how representative that is. It would be interesting to see if they trend similarly.
Thanks for the cool project!
Yeah, that's probably the main limitation right now. This is measuring /new, so it's every submission rather than just the stuff that makes the front page.
lcamtuf's survey (linked in the footer) was looking at the front page, and the number there was quite a bit higher. So there's likely some effect from voting, although I haven't dug into that yet.
I think Algolia has a way to filter for front-page stories, which should make it pretty easy to add. I haven't tested that yet though but will certainly look into it!
At first I thought about how many posts were written by AI
Most of the time things make sense in my head until someone points it out haha. I should have written it as "related to AI" or something like that.
Multiple times authors have been asked about AI usage on HN, because readers (rightfully) suspected it was AI written. I don't have a problem with AI written articles, or assisted.
But then the authors lie or make ludicrous excuses for why their article "sounds" like AI. Then I check with Pangram and yep. It is completely AI written.
You get a few free checks on Pangram each day or so. Just start checking and encourage people to be honest.
AI has its place. Lying about it and misrepresenting it makes everyone dislike AI more, even as they use it themselves.
> Then I check with Pangram and yep. It is completely AI written.
“AI detectors” do NOT work reliably.
Even humans who are ever-so-certain about being able to identify AI usage constantly embarrass themselves when that claimed ability is put to the test. Look up “AI Monet Instagram” for a social experiment where hundreds of cocksure users pointed out “AI signs” in what was claimed to be an AI-generated painting in the style of Monet, but was actually a genuine Monet.
Every technical topic involves AI now. Even if it's about why the ancient Greeks didn't turn their steam engines to industrial use - the AI's should be researching the details. We might as well ask which topics involve humans or use language.
It's remarkable how stable 20-30% show HN concentration is after late 2022, for the extended stats.
The Show HN volume went 805/month to 4,199/month over that same time period, so from 3.2% of all submissions to 13.4%, while HN overall only grew 1.3x. So the flat ~24% held through a 5x expansion in the denominator, which makes it a lot more interesting honestly.
I expected this to call out the number of links that are obviously vibe coded, like itself!
If you see:
- ALL CAPS subheaders
- a light beige background
- burnt sienna anchor and/or highlight text
- a "• LIVE" (or really any pulsing • symbol with an ALL CAPS follow-up)
- tons of variation in font sizes (this one has 15 different text sizes!)
- more typically, a lot of overexplaining
...you know it's vibe'd, specifically with Claude!
Interestingly, we can deduce that the actual percentage will always be higher since not everyone uses those labels/words.
Would be curious to see how many comments are AI generated.
I wonder what the longest chain of Ai on Ai comments are.
it would be cool to use embeddings to categorize into like 10-20 categories and get a giant pie chart
Not a bad idea. I'd look at creating a separate page for embeddings. Crypto and NFT's would be the next best candidates to add. Something for me to work on over time.
You don't need a website for this.
you could fetch the text from linked blog posts over time and run 'em through pangram, I bet >20% of blogs linked on HN today are totally LLM written
An estimated 2,000 articles per month at 2,000 words per article would result in 4,000,000 words, which costs either $200 or $2000 to process at Pangram [1]. It might make life a lot more enjoyable for many users here, though.
[1] https://www.pangram.com/pricing?category=developers
Several people have asked for this already, so I'll add it to my to-do list. I just want to be careful about flagging posts that appear to be written by AI. Perhaps I set certain thresholds of certainty to filter by. What I have currently is easily verified and matched by the words we are filtering for.
It is also interesting to note that I have already received 5 AI written emails to the email I linked in the footer.
Related:
How Much of HN is AI? https://news.ycombinator.com/item?id=49435728
Yes! This post was the inspiration and is credited in the footer.
Who would do this project and not use an LLM to classify what counts as "AI"?
I tried to click on the link at work and it got blocked as "pornography".
Brilliant
Good. Now do what % of posts link to AI generated articles.
That would be a much more depressing statistic.
And so comes the AI comments as many as the AI posts just saying for fun
Something to maybe track later would be /newcomments, but that would only grab the most obvious comments. People can certainly direct an AI to write in a style that sounds human, but I'm sure it'll still catch a fair amount.
If you read a lot, you start to get a feel for phone typos vs AI hiccups. I can guarantee that is post was made by a brain designed to eat berries in a cave. Almost all the another posts look a bit scary.
I think I misunderstood - how many posts were about AI vs how many posts were written by AI.
I obviously am suffering from berry fog.
Exactly yep. Maybe even more when time machine has been invented by that AI bot in far future I meant the number of comments by AI(-bots)
Recursive given that this is AI generated from another post on how much AI was on HN.