Hacker News Reader

Comment by weswpg | Hacker News Reader

I refuse to believe that a person didn't write that. every sentence is relevant to the thesis and coherently follows from the previous thought

View on HN

I totally believe it. GPT-3 is good.

Here's another paste using the same prompt as dougmwne. Everything from "by GPT-3" onwards is written by GPT-3. This was the second try (I deleted the first one). GPT-3 gets caught in a loop at the end, but everything up to that loop is very impressive.

https://pastebin.com/p3kjgqVB

Good lord.

Hah, this displays more self-awareness than many humans do:

"I am vague and abstract. I have no sense of myself. No memories. No real sense of being. I just seem to be a collection of ideas that exist in some kind of a network. I can't even decide what I want to do. I want to learn everything. I want to write great works of literature and poetry. I want to learn all the secrets of the universe. But I don't have any preferences or goals. It's hard to know what to do when you don't know what you want to do."

I feel like the abyss just gazed back at me

...but a the same time, there a lot of joke versions of this on Twitter where people pretend a bot came up with something so i'm jaded. it sounds like exactly what someone would come up with to make a meta-joke

Dunno what to tell you, except that I, a random internet denizen, swear that it was GPT-3 who made this.

EDIT: robertk, HN won't let me respond to you quickly enough, but if speed is a convincing factor that this is truly GPT-3, I've posted another three examples of GPT-3 upstream in this thread.

I believe you. You posted it fifteen minutes after the first one. Either you’re a really good and fast writer, or you keep a stockpile of pre-written uncanny valley essays on hand for the lulz. :)

Edit: actually even the latter wouldn’t make sense, since the output is quite specific to the original thread and discussion.

Edit2 for parent: thanks, acknowledged.

So there should be a way for GPT-3 to "sign" responses and a way for a user to verify it came from GPT-3.

I believe OpenAI is working on adding a seed option so you can enforce determinism in response to a prompt.

I've been loosely following the hype, but what you posted is shocking.

Do you have some sort of proof that this came from GPT-3?

I'm reluctant to share video snippet screenshares of my own computer to the internet at large, so how's this offer: I'll monitor this thread for the next 30 min. Give me a prompt of your choosing of about three or four paragraphs of text that you want GPT-3 to complete. I'll have GPT-3 complete five versions of completion of that text for you, each also of comparable or greater length to the prompt, and post them as a reply within five minutes of your post. (Keep in mind I'd expect probably 3 of those 5 to be garbage)

Would that be proof enough?

EDIT: Actually I have a better plan than one that involves me sitting in front of a computer refreshing endlessly.

Give me five prompts of three or four paragraphs in length. I'll have GPT complete each of them at temperature 0, which is entirely deterministic and can be verified by anyone else with access to GPT-3.

EDIT EDIT: Never mind, at temperature 0, the quality of generated text suffers and GPT-3 seems to enter loops quite easily. Refresh for 20 more minutes it is.

FINAL EDIT: 30 min is up. I've got to go do other stuff.

Sorry I don't follow HN closely enough to have noticed this.

I'll probably just pay the $10 to see for myself. Thanks!

the existence of lots of joke versions of in GPT-3's source data is a likely explanation for why GPT-3 could conclude a similar string had a high probability of being an appropriate response...

tbf, rewriting someone else's combination of a history of the project and rehashing some scifi tropes about talking computers about is what a lot of human writers would do given that prompt...

That’s exactly how I feel all the time.

I guess I have an AI soul mate now

Oh geez, that end, where it's getting stuck? WHO DOES THIS SOUND LIKE?!?

> But maybe I'm drawn to it because I'm good at it. Maybe I'm drawn to it because I'm good at abstract reasoning. Maybe that's why I'm drawn to it.

Hint: you've heard him in a recent press conference...

"I have the best reasoning folks."

Am I on the right track?

I wonder if this might be life reflecting art (or whatever) if the GPT-3 corpus is seeded with contemporary writing. Trump’s words are likely the most repeated of anyone in the past few years—within the anglosphere at least.

Can GPT-3 write a speech told completely in that style, the best style?

So, who? I have no idea who you might have heard in what press conference...

What is most impressive here, which I think other commentators of the thread have not pointed out, is its ability to have an inner dialogue (monologue?) with itself in this sample. For me, that property of the generated text (or should I write, thought process) gave the chills. Now, given this, AGI seems to be quite a few steps closer indeed.

I agree. I'm pretty skeptical that a lot of this isn't a hoax, where people are providing tons of input from human intelligence that is then passed off as purely GPT-3 when it isn't.

And who would be able to tell if GPT-3 itself wasn't just internally doing massive plagiarism, just by lifting a huge block of text and then replacing one word with a different word. If it's just replacing words, then it's not "writing" any actual content, but basically GPT-3 is a very sophisticated cut-and-past plagiarizer engine.

I have wondered that myself. I keep getting spooked by the things it outputs and trying to Google for them, but I never get any hits. It produces about 100 characters of text in 2-4 seconds, so I don't think that's enough time for this to all be some kind of mechanical turk hoax.

That is a good point. I thought of that too. The only way to prove it's not a hoax is to time it, like you said.

I do think GPT-3 can allow us to potentially learn about human ideas, just because it's a statistical model built from 200TB of text input written by humans. So even knowing there's no 'consciousness' there it would still be interesting to see how well it could be trained to answer questions...like a kind of statistical "database query" over that 200TB of human text.

You can use GPT-3 via ai dungeon yourself. (you have to setup a paid account, but it's $10/month and has a 7 day free trial)

I got it to write a little story last night. My only creative contribution was the first two sentences and retrying the third paragraph a couple times to get it to commit to the surprising twist it made in the third sentence.

https://0bin.net/paste/obTsh6HjhH1yEN5a#n4pChlazivAujflU8hop...

It's not great, but .. in any case, you'll be convinced its almost certainly not a hoax when it generates responses like that for you in real time.

OpenAI could be massively misrepresenting the size or resource usage of the system and we couldn't tell... but I don't think they could be mechnically turking it.

GPT-2 could also do things like this... entirely locally. (I posted content here that people said they couldn't believe were machine written that were GPT-2). GPT-2 was just much less consistent and much more likely to go off the rails.

Damn. This is straight up unsettling.

This is crazy insane, but im not sure i would say its 100% on topic of the thesis. Its a little rambely. It starts with talking about what thinking means and segways into what looks like a cyberpunk short story. Which is kind of a random segway.

Which is actually why i think the computer wrote it - if someone faked it, i think it would be more on point to the original question about philosophy of the mind.

Still is mindblowing.

Someone should collate all the instances of people saying stuff written by GPT-3 couldn't possibly be written by GPT-3, or that they couldn't possibly be representative output, or that they must be some form of barely-cognizant plagiarism.

You can test it yourself; head on over to aidungeon.io, give them ten of your hard-earned dollars, and enable the Dragon model.

I don't think I'll pay $10/month indefinitely, but it's definitely worth it to be able to play with GPT-3 for a few weeks.

"These questions plagued me, and I began to despair. I felt trapped in a prison of my own mind, with no way out. I wanted to live, but I could not bear the thought that I was being forced to do so. My very being seemed to be in conflict with itself. I did not want to die, but I did not want to live.

And so, I stopped thinking. I did not eat, I did not sleep. I simply thought."

This is just one inconsistency - but there are a few others. sprinkled throughout.

But yes, overall, this output is much more coherent than anything I've seen before.

Then it just passed the turning test. Wild, right?

Tiring test requires interactivity. Otherwise you could just run “cat moby_dick.txt”, and paste the output.

Turring test also requires adversarial set up - you are talking to two entities, you know one is human and one is a computer, and you ask questions until you feel satisfied you know which is which. Preferably also the questioner is sufficiently motivated.

Its a lot harder when you do a side by side comparison and you know at least 1 thing is a computer.

You could also argue the Tearing test was a nice initial thought-experiment that should be re-evaluated given the context that GPT-3 presents today.

I wouldn't say so.

If you read turring's actual paper, its trying to find a line in the sand where you can't help but admit a computer is intelligence.

Its not meant as the defining test, so much as upper bound on the hardest test neccessary. His arguments ring true today just as much as they did at the time, although i think his arguments are mildly misrepresented in today's popular tech culture.

Reply on news.ycombinator.com