Hacker News new | ask | show | jobs
GPT-3 generated blog post reach #1 on Hacker News (technologyreview.com)
173 points by hamsterbooster 2134 days ago
15 comments

HN submission from the blog post author, where the comments from dang reveal the author engaged in voting manipulation: https://news.ycombinator.com/item?id=24062702
Sorry to pop the balloon, but the story is bogus.

It's false that the post was generated by GPT-3. The author admitted to writing the title and editing the intro. He also described the article body this way: "as unedited as possible"—in other words, edited. It's false that (as he originally claimed) only one commenter called the post out as GPT-3, and false that all such comments were downvoted.

All that is just what he publicly admitted. How much of the rest is also fake? People who try to game HN like he did, including with bogus accounts and fake votes, are not known for scruples. It seems that, having got busted in dishonest attempts to get attention on HN, he decided to get attention from journalists instead, and found one who didn't bother to check the other side of the story.

I was trying to be neutral but your response is much more accurate. :P
Upon reading things like this, I feel... so innocent and naive.
Regardless of who or what wrote it, the article is insubstantial, pseudo-intellectual blather. It reads like a fourth grader mocking grown-up problems by throwing around words he doesn't understand.

The following is completely nonsensical, and occurs very early in the article, right after the general introduction:

"Over-Thinking (OT) is the act of trying to come up with ideas that have already been thought through by someone else. OT usually results in ideas that are impractical, impossible, or even stupid."

The only reason to keep reading after this is just to see what other bullshit has been heaped on, like not being able to take your eyes away from a train wreck.

I didn't visit the story when it appeared (regardless of its upvotes) because the very title smelled of self-helpy twaddle. I would for sure have flagged it.

That behavior may be a clue as to what happened. The submission title was crafted in such a way as to deter "nonsense-averse" users from clicking on it, and that may have helped it evade flagging. If a submission evades flags, the only other points it can get are upvotes.

With that there is no credibility to the blog being actually generated by GPT-3, or at least being heavily cherry picked.
I think GPT-3 could be blessing for Reddit and HN type forums.

Add some GPT-3 content and links into the feed. Decrease the vote weight for users who upvote them and increase vote weight for those who downvote them.

This is actually a great idea.

I've noticed a trend recently in a subreddit I frequent that obviously wrong statements are suddenly getting upvoted. I'm not talking about things that conflict with my opinions - I'm talking about statements that are demonstrably and objectively incorrect.

I think the issue is a critical mass of people very new to the topic at hand who upvote things that sound reasonable without having the knowledge to be able to engage with the statements even slightly critically. This forms a feedback loop as upvoted comments are assumed to be reasonable. It means you can say something blatantly wrong, but stated in a manner such that it assumes the form of a sensible insight, and be upvoted for it.

Something like the system you propose would do well to promote genuine insight above platitudes, and platitudes above superficially plausible misinformation.

I don't see this happen often in areas where I'm an expert but I do see it, and it's weird.

There's a post somewhere, maybe one of the Stack Overflow sites or Reddit where somebody asks about how SSH works, and the answer given and upvoted is horribly wrong, it's like how somebody who half-understood an explanation of PGP might think SSH could work.

So I down-voted that and I wrote an explanation based on my understanding but referring to the RFC as I went, and, whenever I was surprised by the RFC, also checking the OpenSSH source code (the RFC is correct, but, you know, always worth checking).

It got downvoted. Zero comments. So clearly people are looking at these two explanations that are quite different and they are down-voting the one they... don't like?

Or even better, for repetitively-discussed topics, just generate the entire comment thread to save us all time. Examples: nuclear power, housing/transit policy, Tesla/Elon Musk.
Standard comments can be automated even without GPT-3:

For all scientific studies the correlation is not causality comment.

For articles related to economy the mandatory money printing rant. Sometimes the whole fractional reserve banking spiel.

For astronomy the "blows my mind how insignificant blah blah."

Anything related to nutrition should have keto anecdote. "... and I have never felt so good."

Any Show HN gets a top comment saying it's a nice pet project but pretty useless seeing as there's some other tool out there that does the same thing except better
GTP-3 is able to output true statements (like sentences from wikipedia).

Sometimes true sentences contribute to discussion.

Was this written by GTP-3?

I agree with everyone: this is an intriguing idea.

I feel like if we did it, it would be sporting to reveal which comments they were, after a suitable amount of time.

One tricky bit is which accounts would post the comments. If they were all new accounts that hadn't posted anything before, that would lessen the value of the test. I'm not sure it would make sense to have dedicated accounts for this. Perhaps we'd have to sprinkle such comments among established accounts? With permission from the account holder, of course, plus swearing them to secrecy? That starts to sound complicated.

You can sprinkle into my comment stream. It's already generated mock confusion: https://news.ycombinator.com/item?id=23980312

(The thread at https://news.ycombinator.com/item?id=24006393 shows that GPT-3 can already pass my screening turing hurdle, one which many actual people arguing on the internet fail.)

====

stolen from https://arachnoid.com/jokes/index.html :

Two academics, Albert and Bill, are sitting in a bar waiting for their friend Charlie.

Albert: "Charlie thinks women don't know any math, he might be right, but I want to play a trick." Albert calls the waitress over.

Albert: "Delia, my friend Charlie is going to arrive in a bit, and I want to play a trick. When he arrives, I'm going to ask you a question, and I want you to answer, 'X cubed divided by three.' Can you remember that?"

Delia: "Sure, no problem, I can remember that."

Charlie arrives and Albert raises his favorite topic.

Albert: "I think you're wrong, I think women can learn math. Just as a test, let's ask the waitress a math question." Albert calls Delia over.

Albert: "What's the integral of x squared, derived with respect to x?"

Delia: "Umm, that would be ... x cubed divided by three"

Delia: "... plus a constant."

You can hide everyone's names for a day or so.
That would change too many variables at once.
I'd like to see a more dynamic interpretation of this.

1) A form has seed users, extrapolate how they would vote based on people who vote like them but see more content. Use the extrapolated prediction of seed behavior to rank.

2) Let the viewer change who their seeds are manually. Let the viewer rank the posts, and see which seeds work best for them. Make this process and equilibrium building dynamic.

excellent idea!
So, you’re experimenting on users without their knowledge? There’s a thing called ethics. You might want to look into that.
You can avoid ethical concerns by ensuring that users are technically informed this might be happening. In fact you might get great results by informing users this might be happening even if it you rarely or never actually do it.

You can also phrase it quite broadly so as to ensure that participants were technically informed of why it's being done and yet nonetheless unaware of what exactly you're actually doing. That's how psychologists design experiments that want to measure something subjects would prefer to conceal because of low social desirability.

For example suppose we're wondering if people are secretly biased against rectangular shapes in video games but are feeling a social pressure not to admit this bias. We tell subjects we want to test for bias against rectangles, they're going to play a video game, they are to collide with the red objects (regardless of shape) and avoid blue objects, we will show how many rectangles they hit on the screen.

But we don't actually care about the count, we actually use eye-tracking technology to measure which objects on the screen the subjects look at, when and for how long and we use this fact, that the subjects don't realise we care about, to check for bias.

Isn't every single ad ever created experimenting on users without their knowledge? I.e., will the user click on this ad or buy the product if we show them this content? That has much greater probability of adverse consequences for the user than the proposed experiment.
Sure, I have no problem with calling all advertising unethical. A lot of software A/B tests also land in unethical territory too.
So if a restaurant sells today pizza only while yesterday sold only pasta to see what sells better, is that unethical?

Or is it unethical if half the people get one waiter and the other half a different one, to see who can sell better?

Who said users can't be told about the system?
Good argument against A/B tests and using any form of analytics to improve your site!
I remember skimming this article. My initial feeling was that it was written by someone who wasn't a native English speaker. It seems like I'll have to pay more close attention to that feeling in the future.
On the other hand, you could have been right. It could have been a non-native speaker.

I'd hate to end up in a situation where non-native speakers are accused of being bots...

I feel like so many of you are just utterly too kind. Without getting into the generated article, we have a lot of Orwellian new speak everywhere. Your standard corporate/management/hr speak is very pervasive. It sounds inhuman. This is a language many people adopt to fit in and make it in this world. It shadows itself in blogs they write, particularly signaling blogs.

The generated article follows in this vein. That’s what gpt will replicate, not the simplicity of a non native speaker (that would be easier to spot). It will follow the amorphous blob shape of saying something, but nothing, with the ominous undertone of ‘you know what’s going on, but you wouldn’t dare speak up’.

How many of you read something from a company and instantly think ‘this sounds like horseshit?’. How long did we let that go on? Forever right? We lost this fight before it even happened.

I think I agree with most of what you said, I just think you need to be careful. Don't confuse "horseshit" with bad English. The latter may contain something interesting, and the former never will.
You don't need to be a non-native speaker for that. If the other side is convinced you're a bot, it will be very difficult to make them believe otherwise. Which in a funny way has echoes of conspiracy theory believers who take every denial as a further proof of their existing belief.

The first I heard of this was 15 years ago[0]. The original article is no longer online, so the link takes to a PDF rendition. It cites the source.

0: https://msu.edu/~pennock5/courses/ALife/Striegel_Failed_Turi...

I am impressed by the result there.

I read the article with a motivation trying to find fault in it. Albeit being somewhat repetitive, it doesn't appear to be a piece that is inhuman.

The model even knows to demonstrate its idea with examples, which is pretty wow worthy.

Yeah I skimmed through and was like "there is no content here"
I have hopes this will actually improve the average quality of blog posts that get upvoted and shared on platforms.

Because people may realize that if their blog entry is going to be so bad it could be generated by GPT-3, they should probably be doing something else. And everyone else who is upvoting may just become a bit more aware what constitutes something of substance.

Things GPT-3 can't do:

    - Research
    - Technical Documentation
    - Investigative Journalism
    - Write useful software
GPT-3 may be able to fake the first three, but that would be glaringly obvious (because it'd be lying if it isn't just copying and also each of those are generally more than just text content).
The HN article that this is about: https://news.ycombinator.com/item?id=23893817

Congratulations to the authors of these comments, who correctly guessed that it was written by GPT-3:

https://news.ycombinator.com/item?id=23894742

https://news.ycombinator.com/item?id=23894000

Ah ah ah, the answer "your comment punches below the belt and isn't acceptable in a community like this.".
is this the largest scale Turing test ever performed on an audience acutely aware of what a Turing test is?
I think fewer people know what the Imitation game actually consists of than think they do - namely that a questioner has to guess whether a machine is a respondent to their questions, rather than another player. That certainly hasn't been tried with GPT-3 that I'm aware of.

However I would agree that it is an oblique version - that is, can a machine fool humans into thinking that they are human.

In which case I think it's probably safe to assume that GPT-3 has passed.

Turing actually asked a slightly different question, that I think is a lot more interesting. From Computing Machinery and Intelligence by A. M. Turing:

> The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game he says either "X is A and Y is B" or "X is B and Y is A." The interrogator is allowed to put questions to A and B

> What will happen when a machine takes the part of A in this game?" Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman?

Somewhat politically incorrect, assuming men and women should ever be distinguished, but much more revealing about how exactly people see themselves.

Wow, good observation. I wouldn't be surprised if this is actually true.
Any evidence to suggest that HN crowd distribution doesn't lean towards Pareto?
how does pareto distribution apply in considering hn comments? you're saying a small % are aware of Turing tests and the rest are gullible? perhaps.
> how does pareto distribution apply in considering hn comments? you're saying a small % are aware of Turing tests and the rest are gullible? perhaps.

Exactly.

The problem is that it's also true. Comments like that decrease the quality of the discussion by attacking the "author" rather than the content.

(If the commenter had written even a couple sentences on why they suspected GPT-3, I would feel differently.)

I mean, there's not much to elaborate on beyond what they said: "Zero substantive content, pure regurgitation."
It's not true, though. There really is substantive content in that blog post (although it's trite and probably useless advice, it's certainly making a real argument).
I dunno. "Trite" and "substantiative" usually don't go together.
Then don't say anything. Click the hide button.
I agree, and this is my instinct, but I worry about the effort becoming very asymmetric. In a world where a large amount of content CAN be generated (my time to read becomes much larger than time to generate / get my attention), and I may or may not be reading sincerely generated content, at what point do I get to attack the content creator for that, especially since it's very difficult for me to understand any given content creators intentions without some level of social proof or trust.

I think we're likely to get to a place where we either consolidate to trusted sources of information, or accept machine generated content as valuable on its own, even if for novelty and just move on with our miserable lives.

This is exactly the reason why academics like their ivory tower, and value credentials more than egalitarian hackers do.
Because they can trust the work of their peers as having met some agreed upon standard, or are you saying academics are bad and hackers are actually egalitarian as a general rule?
Exactly. I can't tell if that poster actually suspects the content was generated, or if they are simply using the comparison as an insult.
As a criticism, i don't see how that matters. An article written by a human at the quality level of gpt-3 has the same value as an article actually written by gpt-3. Knowing that gpt-3 actually wrote the article doesn't change its level vapidity. Either the article is vapid or it isn't.
But there's no need to accuse the author of being a robot. They could have written:

> This article is vapid. Zero substantive content, pure regurgitation

Now, I would still have downvoted such a comment. If I'm reading a thread, I clearly like at least the article's topic, and I want to read more on that topic, not dismissive comments telling me that it's stupid. But, at least the modified version focuses on the content, rather than the author.

It's possible to point out issues with a submission without insulting the submitter or author. Calling what is presumably a person the equivalent of GPT-3 is needlessly rude. Even insulting the work, but not the (presumed) person behind it is better.
OP should apply to be a Blade Runner. Hottest new job.
Totally sound and what we need more of anyways.
This is only the tip of the iceberg.

I'm certain GPT-3 has been commenting on HN threads for a while now. In some cases, its presence has been disclosed (see, for example: https://news.ycombinator.com/item?id=23886503) In other cases, GPT-3's presence has not been disclosed; the machine has been pretending to be a human being, largely unnoticed. Consider only how easy it is for it to write short, punchy comments -- say, one to three sentences long.

By implication, there's a high probability that we -- you, me, and everyone else on HN -- have been upvoting and downvoting GPT-3 comments for a while without realizing it.

And the technology is only going to get better.

I'm reminded of https://xkcd.com/810/.

The example comments you linked (Where GPT-3's presence was disclosed) were believably human, particularly if you were skimming, but they were not good comments. If not for the note at the end about GPT-3, I'm pretty confident they would have been downvoted.

And if I'm wrong, and GPT-3 is actually capable of writing thoughtful and substantive comments... well, in the words of XKCD, "mission fucking accomplished."

The longer the comment, the easier to detect GPT-3. The more rigorous the subject, the easier to detect GPT-3. But GPT-3's presence is harder to detect for short comments on less rigorous subjects generated using as input context actual headlines and top-voted comments on HN.
Arrgh. I can't believe I missed that.

To be fair though: it's quite clear that all the discussion in that post is commenting on the headline and not the content. Bland self-improvement "life hacks" are one of the metaphorical crack pipes of this site. We all have way too much to say on the subject of our own productivity.

So I think it's less likely that HN was fooled by GPT-3 but that GPT-3 was good enough at filling a plausible article out around a tempting headline.

I didn't see the article at the time but from skimming it I'm sure I would have been fooled.

GPT-3 fits snugly into the pattern of "AI does intelligent thing -> we decide/realize thing doesn't reflect meaningful intelligence". Maybe in this case the devaluing of blog "crack" is a positive thing.

The article if written by a human was bad -- poorly argued and trite.

However, my mind is blown if it really was written by an AI. As bad as it might seem by human standards, it's almost impossible for me to accept that this was created by an entity without consciousness or at least understanding.

Edit: it seems this might be a fraud; i.e., it indeed was produced by an entity with consciousness. I almost hope that’s true, as it’s much less unsettling.

It is stringing together phrases and sentences that appear in the training data. After digesting gigabytes of text, it has built up structures that represent grammar and semantics to some degree, so you'll find very coherent sentences, because those sentences were originally written by humans or pasted together from sentences written by humans. It's amazing that it works as well as it does.
I disagree. The essay in question develops a coherent argument over several paragraphs. It’s not just disjointed individual sentences.

The argument about overthinking vs creative thinking isn’t particularly _great_, but it’s certainly intelligible.

Saw this coming a mile away.

A lot of articles that make it to the front page of HN are formulaic. We open ourselves up to this. Every blog post with shallow observations, every tutorial showcasing the first few pages of documentation, every biography on how to make money fast, every lucky shit that pontificates on how to manage teams and companies, every one selling an ebook, and everyone selling an ebook about selling an ebook after having sold 50 lifetime ebooks (topic being about success of course), and it was only a matter of time.

I count 3-4 posts about depression and existentialism per week on HN, and few ever reference the depth in which many great writers dig deep into the subject. Exercise more I guess.

Time to add ‘did a novice or a robot or a sociopathic narcissist write this?’ to our critical thinking toolbox.

Edit: I can’t tell if I fell for a gpt article about a gpt article, for what it’s worth. This is going to be a disaster when it hits the masses.

I couldn't read this article. I was convinced it was another AI generated text. I just scrolled to the bottom to see if I was right. Then didn't read anyway because it could still probably be an ai.

You guys saw the last ai generated text about ai generated text-right?

Looks like he wrote the title by hand, how many upvoters read the actual article?
And another article from the author's blog, that describes the whole game:

"What I would do with GPT-3 if I had no ethics" [8/3/20]

https://adolos.substack.com/p/what-i-would-do-with-gpt-3-if-...

> Ever since COVID hit, everyone and their mother started writing online. One of the most interesting ways people have been playing with this technology is in feeding it article headlines and introductions.

> While the output is not perfect, you can easily curate it to something that's convincing. This will make it so easy for people to just pump out clickbait articles to drive traffic.

> It would be pretty simple to do actually.

> First thing you would need to do is come up with a name. If it were me, I’d name it after the Greek god of deception or something like that just to be clever. Then I’d just stick an “A” in front so nobody gets suspicious.

> After that, I’d make a substack because it takes no time to set up. Once thats done you have to come up with some content. GPT-3 isn’t great with logic, so inspirational posts would probably be best, maybe some pieces on productivity too.

> Once you have your name, your website, and your content, its time to promote. Just start posting your articles on a website like Hacker News and a couple are bound to get popular.

Interesting. Has anyone done any analysis on what the tells are for GPT-3 written articles?

To me they feel like they're slightly off grammatically. Not in an ESL way, but more like someone really anxious who wants to explain a conspiracy theory to you. However i can't entirely put my finger on it. They do seem to overuse self-reflective statements (I think X) and transitionsal phrases. Maybe.

The best part is the very serious people shushing those who saw through the prank.
I've seen this criticism a few times, specifically about the example in the article, and it's bonkers to me. What I see is somebody calling the blog post garbage, and another person saying that person was being hostile. In what possible interpretation is that a "very serious" person "shushing" somebody who "saw through the prank"? It's totally reasonable to interpret "this looks like an AI wrote it - regurgitated garbage" as primarily an insult. That it turned out to be factually true is unrelated to that.
The comment sin question were

This is either something written by GPT-3, or the human equivalent. Zero substantive content, pure regurgitation. and

I think this was written by GPT-3.

I think you've misrepresented the tone of those comments, and saying that the correctness of their matter-of-fact opinions is unrelated to their validity is strange to me.

It's not just that these commenters said 'this blog post is no good' but that they correctly identified its artificial nature. It's like the difference between dismissing a photo or social media profile as fake and correctly pointing out that it uses an image from thispersondoesnotexist.com.

Hold on, there's some inaccuracy here. Only one of those comments got pushback, and that comment wasn't simply matter-of-fact; the problem with it (from my point of view anyhow) was that it added a gratuitous insult ("or the human equivalent"). That made the whole thing read more like snark than straightforwardly raising a question. The other comment was more matter-of-fact about calling GPT-3 and didn't get any pushback.

The problem is that the cases legitimately overlap. That is, "sounds like GPT-3" gets used as an internet insult (example: https://news.ycombinator.com/item?id=23687199) just like "sounds like this was written by a Markov chain" used to be (example: https://news.ycombinator.com/item?id=19614166). It's not surprising that someone interpreted the first comment that way, because it contained extra markers of rudeness. That may have been a losing bet but it wasn't a bad one. Perhaps the other comment didn't get interpreted that way because it didn't throw in any extra cues of rudeness—or perhaps it was just random. Impossible to tell from a sample size of 2.

Not to take away from the glory of lukev for calling it correctly. I just don't think the reply deserves to be jumped on so harshly.

Sure, Dan.
Those serious people keep this place sane/productive
To be fair: it sounds like a GPT-3 post was considered an insult back when. Not sure it passes as an insult anymore.
25 days ago?
Way back in the day.

Covid will be remembered in the future as when GTP-3 arose from the ashes.

Time isn’t the issue here, it’s the fact that people defended an AI as if it were a person.
“I think the value of online content is going to be reduced a lot.”

Or raised...

Indeed, it's not like every community out there get HN-like level content.
Was this article written by GPT-3?

It's so meta even this article ;)

Was this comment left by GPT-3?
It's GPT-3 comments all the way down, I'm afraid.

Anyway, for those who missed the joke in the grandparent post:

https://xkcd.com/917/

Are we GPT-3?
Not sure if you meant this in jest, but seeing what GPT-3 can do and extrapolating from there - I'm left with an uneasy feeling about this very idea..

What does it really mean to understand something? Is it all just an illusion generated by a fancier biological generative transformer? :P

The first GPT-3 academic paper will be accepted before the end of the year.