It’s practically indistinguishable from a human. Not a creative, insightful and unique human. But an average human? Yes, I cannot tell the difference.
I must repeat that - I cannot tell the difference!
This can probably completely replace or supplement most online content that I see including news, certainly on the vacuous side of things of which I think there is a lot of content.
Those online recipes with irrelevant life stories before them? Replaced. Those opinion pieces in news? Replaced. Basic guides to tasks? Probably replaceable.
I know I probably only see the best output, and it would be nice if I had more context, but the peak performance is amazing
The twitter video showing GPT-3 generate HTML based on your request? I think there’s a lot of potential. I don’t knew whether it can, in general, live up to these specific examples though.
GPT-3 can't do any of that. It can pick up clues from the text to produce incredibly realistic sentences that are related to the broad topics of some text, but it's all smoke and mirrors in the end - there is no model of the world getting expressed in communication, it is just mindless aping of similar speech.
And yes, this basic skill that GPT-3 has is enough to replace some human tasks, like inventing plausible sounding stories for a recipe or perhaps even taking news from one site and writing them on another with slight alterations. Perhaps it will even be able to take some facts and weave them into a speech about that topic.
But it is not even close to doing something like real journalism, even at the level of a car mechanic telling you what happened down at the mall.
You cap your comment by a non-sequitur that GPT is not going to replace journalism.
IMO the piece of text generated by GPT offers more insight and is wittier than yours.
Check out Figure 3.13 in the GPT-3 paper, https://arxiv.org/pdf/2005.14165.pdf.
The authors experimented with 200-word news articles, to see whether 80 human judges could tell the difference between human-generated and GPT-3-generated ones.
It turns out they could not: the human judges correctly identified GPT-3-generated content only 52% of the time, essentially as good as random guessing. (And no, the machine-generated articles were not cherry-picked for the experiment.)
It even concludes with reasonably good advice: to pass a Turing test, AIs should say things that are true, and tap into human emotions. To me, this disproves the idea that it's "not saying anything". It's definitely saying something, and that something is both true and not commonly understood by the general public. It is therefore capable of "teaching a basic skill".
I find that very impressive, and I'm surprised that so many others here don't.
you're right; this AI is no substitute for a journalist. however, I do think the "essay" compares favorably with some papers I peer-reviewed for my college writing seminar. sometimes humans really aren't trying to communicate anything; they are just trying to hit the minimum word count and get a passing grade.
> But it is not even close to doing something like real journalism, even at the level of a car mechanic telling you what happened down at the mall.
You're just talking about layers of abstraction. GPT-3 works with chunks of 3 letters. It can now.
Combine chunks to form valid words. Combine words to form syntactic sentences. Combine words to form semantic sentences. Combine sentences to form consistent paragraphs. Combine paragraphs to form trains of thought.
It can't combine trains of thought to produce a consistent point.
But it's already working successfully at 5 or 6 levels of abstraction up. I don't think it's that much harder to get one to two levels higher in abstraction. It doesn't know level 7 is any different from level 6. It just needs the data and compute to start modeling at that level.
In the past you could've convinced me a different architecture is needed to accomplish that. But I also wouldve doubted it could get this far - why would something stop it now?
It just needs to study more long form work and how to reach a conclusion in an essay starting from paragraph 1.
You know this inevitably tempts the cynical question of how different what you describe is from (to put it optimistically) clickbait generators and (to put it still more cynically) much of the content generated today ….
Oh man. Machines replacing humans writing stories that humans were only writing to satisfy machines in the first place.
I haven't done a survey of GPT-3 output from other sources, though, so I don't know how typical these outputs are. To those who have: what is your opinion? Do you think they're real (in the sense of being written by GPT-3) or fake (in the sense of being written by a human)?
The tone, the use of voice, the straying abstract line of thinking. At the same time these flaws could all be faked.
Overall I give the author the benefit of the doubt that they are legit. But if I had to pick an essay that's fake that I've read (knowing one had to be fake), it'd be this one.
But safest best is that it's real, and we're just amazed by how good gpt-3 is.
If I read this without the context, I would only assume it was written by a human given that is the implication when I read something like this, but I would immediately think it's idiotic and move on.
There is simply nothing being said here. It is just like a desperate student going online to copy sentences, phrases, and words from articles and paste them into an essay, which at its core, is basically what this "AI" is doing.
Frankly, with a little bit of framing, it'd make a great sci-fi short story.
The main problem I see with AI is that it is very easy to approximate "general human intelligence", which is essentially equal to "being indistinguishable from the Joe next to you". But it is a completely different league to actually advance the human race. For that, statistical approximation will never work.
The next step is to create AI that innovates. As long as that isn't done, all we have is a demonstration of how "unintelligent" most human beings really are (i.e. nothing more than a statistical approximation + pattern matching... Instagram and social media essentially is like an AI forcing function for human beings, to make them become average).
And yes, we can couple AI with things like a Go-Engine, SAT solver, theorem provers, etc. to give them abilities beyond what humans can do in these categories, but who builds that? Humans... As long as AI can't build an AI for a category it knows nothing about and has had no training for, that AI remains "as unintelligent as a brick". All it can do is reproduce what its creator taught it.
That isn't necessarily a bad thing at all. This could still be extremely useful for society and put a new evolutionary pressure on the human race to become "above" average. Something that has been utterly lacking in the past century. With general, yet stupid AI becoming a reality soon, >90% of humanity is rendered obsolete. This will cause a significant pressure to improve on an unforeseen scale, which is probably a good thing overall.
Truly intelligent AI on the other hand, might as well lead to our immediate extinction, since it renders the entirety of the human race irrelevant.
This is profoundly underestimating the intelligence of 99.999% of the human race. GPT-3 is doing nothing even remotely close to 'approximating the behavior of "Joe"'. It's not even close to approximating the behavior of a rabbit. It doesn't have goals that it tries to achieve, it doesn't understand its environment and formulate plans for achieving those goals. It can't look at its past and distill some events into a life lesson, or teach you how to perform a basic skill.
The only thing that GPT-3 is good at is producing output that looks like what humans produce when they close down their minds. Yes, GPT-3 is as good as a human at making up stories that don't mean anything, or arguments that go nowhere. It cod probably even produce decent elevator music. But that only shows how mind numbing and devoid of creativity some tasks that we force humans to do really are.
I remember a story about a research group who created an AI to prove Euclidean geometry theorems. The researchers were surprised by its inventiveness almost at the very start, when the AI came up with a new concise proof about the equality of the base angles in an isoceles triangle, which none of the researchers were familiar with.
What the AI had done was demonstrate a side-angle-side congruence of the triangle with itself (BAC ~= CAB) and then immediately deriving the equality of the base angles. I for one find this kind of outside-the-box thinking that an AI can perform to be extremely inventive.
Are we really extinct if we are outlived by AI that we created that emulates our thought patterns, speech and minds, that is a continuation of our art, science, history and culture? I'm not personally that attached to my DNA, if my mind can exist in a form free of DNA, I could care less.
Also note, the one giving the words meaning in the first place is you. GPT-3 is simply repeating patterns that it discovered, but it doesn't have any notion of meaning, any model of the world outside "this is what text I expect to see following this text".
I am not merely being philosophical here. Two GPT-3 instances couldn't "teach" each other even the slightest bit of new information, or prime each other to get some specific kind of response, perhaps trying a few different things to see which prompts are more likely to produce the desired outcome in the other instance - because there is no "desired outcome" - it's all just trying (and usually succeeding exceedingly well!) to produce text that sounds like text it has seen before, while matching the prompt.
I don't think anyone would throw you a murder charge if you dumped the servers hosting GPT3 into the ocean.
Or maybe they would and statistical approximation is all humans are under the hood despite all our insistence we are more "sophisticated".
This argument is fascinating. What is it supposed to prove? Human beings have dumped boatloads of other human beings into the ocean without facing prosecution as well.
I am sorry but when has irrelevance itself ever stopped life? The radiotrophic fungi of Chernobyl essentially spawned as a sick joke evolutionarily speaking. They have one anonalous environment for generations that once they head outside of the scope are essentially on an unviable alien planet.
Irrelevance is never listed as the cause of death on by a coroner. Oh humans and others certainly have rendered species extinct from carelessness, for getting in their way, being too tempting for short term greed, and other reasons but irrelevance is actually protective. Does anybody feel a need to try to render sulfur vent tube worms extinct? There is nothing we want there or really want from the extremeophile creatures to justify genocide.
Even if any offspring AI cared nothing for us they would be more likely to just fuck off to space to find more favorable environments.
https://www.epsilontheory.com/the-elton-hootie-line/
TL;DR: In order to be "heard" or to achieve enough of a following to be economically/politically relevant, you must "dynamically range compress" everything to sound the same and appeal to the largest audience, which removes diversity.
New media can engage in discovery and targetting to a homunculus generated from the individual based upon their models and what they think about them. The model may be a twisted imaginary construct but it is a better fit than the also imaginary "average" human used to target. Now the old media did have their own demographics ranges but they dealt with it backwards and more proxies - asking are people 20-40 interested in foo instead of "is individual person X probably interested in foo?".
I think GPT-3 is very convincing for "soft" topics like the other HN thread "Feeling Unproductive?"[2], and philosophical questions like "What is intelligence?" where debaters can just toss word salad at each other.
It's less convincing for "hard" concrete science topics. E.g. Rust/Go articles of programming to improve performance.
An interesting question is what happens when the input to the future GPT-4 is inadvertently fed by lots of generated GPT-3 output. And in turn, GPT-5 is fed by GPT-4 (which already ingested GPT-3). A lot of the corpus feeding GPT-3 was web scraping and now that source is tainted for future GPT-x models.
It might be possible to filter out GPT-3 generated text from future training data. Simply feed part of the text to GPT-3, and if it is way too good at predicting what follows you throw it out. The same trick could be used to detect students writing essays with it.
This trick will stop working though as more variants of good text prediction algorithms appear, unless we can do the same test against each one.
So, mostly just Internet forum philosophy.
A half-joking prediction:
at some point we'll solve all arbitrarily hard milestones for AIs and will still find ourselves 'nowhere near having real general intelligence'.
At that point we might start questioning our assumptions about intelligence.
it'll be the singularity, superintelligence will be optimizing human affairs and launching probes to other stars as it disassembles the solar system for material to build a Dyson swarm, and people will still claim it's not "really" intelligent.
- shared context (shared biological & cultural heritage)
- recognition & willingness to ascribe intelligence
Basically trying to imply that the general AI problem is of similar nature to the 10x programmer problem - communication & recognition, or lack of thereof.
A 10x programmer makes a hard problem look easy. A general AI makes people around it feel very smart & productive.
Which is also why I prefaced it with, The terrain is not the map - we humans know surprisingly little about our intelligence.
[edit]
Case in point, an older discussion about intelligence in animals: https://news.ycombinator.com/item?id=21772648
Scary? Not GPT-3, but when GPT-6 or 7 gets involved in the political realm, that’s when people will take notice. This essay has a glimmer of “humans can’t be trusted to govern themselves” - and it’s not entirely unconvincing.
In other words by the time it's a choice, I doubt it'll even be a choice.
I think the sheer stupid fear leading to a tall poppy competence treatment is far more dangerous than the dark future ever could be from actually competent entities in charge that we could never do better than.
Disclaimer: I am not an avid science fiction reader but interested in sources talking about superintelligence [3]. Is superintelligence more of the same or it is more about having different layers interconnected?
[1] https://en.wikipedia.org/wiki/Ainan_Celeste_Cawley
IQ tests would have to be constructed to measure somewhat similar intelligences of a population large enough to have a meaningful "population". The scales could then potentially be roughly calibrated to each other, but they wouldn't really be translatable. The task of constructing them would be up to the intelligences in question.
It is possible no such measure could exist; as the "size" of the intelligence increases, the number of degrees of freedom of "intelligence" almost certainly increases, just as we can be good at bugle but terrible at piano as humans (and that's already a fairly microscopic focus in the grand scheme of human activities), but those statements are almost meaningless to ask about a raccoon, even if we give them raccoon versions of the instruments. Even at human scales, while IQ seems to measure something, we can see the measure is getting fairly strained. You probably need an increasingly multidimensional "number" as the intelligence continues to scale up.
As for what intelligence is, all we have are other hypotheses that are on the one hand clearly related to the question at hand, yet on the other, not the answer. Arguably, AI like GPT-3 is also a measure of our best definition of "intelligence". If we could completely clearly define it, we could probably implement whatever it is we defined.
Talking about IQs in excess of human intelligence is not really interesting because of that. It's just not a designed for measuring intelligence as an actual number.
If I had to guess, I would say this essay was written by a human pretending to be GPT-3. Do you agree (except with higher confidence) and that's why you aren't impressed?
> I picked the best responses, but everything after the bolded prompt is by GPT-3.
Based on this, I am pretty sure that the order of paragraphs and the general structure (introduction, arguments, conclusion, PS) are entirely the product of the editor, not of GPT-3. I'm assuming that this is at the level paragraphs and not individual sentences, which does leave some pretty good paragraphs.
Another question that I don't know how to answer is how different these paragraphs are to text that is in the training corpus. I would love to see what is the closest bit of text from the whole corpus to each output paragraph.
And finally, human communication and thought is not organized neatly in a uniform level of difficulty from letters to words to sentences to paragraphs to chapters to novels or anything like that, and an AI that can sometimes produce nice-sounding paragraphs is not necessarily any part of the way to actually communicating a single real fact about the world.
I still believe that there is never going to be meaningful NLP without a model/knowledge base about the real physical world. I don't think human written text has enough information to deduce a model of the world from it without assuming some model ahead of time.
I think this article is quality enough to constitute meaningful NLP. But, your questions about the amount of human intervention are key. If it takes several hours to a day to produce one of these, then it's not really that meaningful. If one person can produce 100 of these in a day, that's pretty meaningful.
I suppose it's a gradual slope here, with spell checkers on the one side, a grammar checker a bit further ahead, then Google Doc's autocomplete, and then close to the very end of the slope you have this system that just needs a few sentences and completes the entire essay for you.
One question then is - at which point on the slope does the comment stop being a direct product of your mind? Another question is - in what ways does this distinction matter?
At what point does the billiard ball stop being an object of your will?
A few weeks ago, GPT-3 generated content looked like nonsensical content farm's content to me. Today, this article makes points and follows an argumentative line.
There are still a few oddities, but this time, it looks like thinking and not just putting related words next to one another with proper grammar.
https://arr.am/2020/07/22/why-gpt-3-is-good-for-comedy-or-re...
> Peter sits in his office, staring out the window, muttering: “I’m telling you it’s a bad idea. A bad idea.”
> He gets a text from Larry, that says “Oops.”
Ever since I had my first encounters with AI, somewhere in the early 90s, one thing has remained rather constant: the extraordinary successes of AI originate mostly from when it is used for tasks we humans perform naturally poorly at, and when confined to a narrow applications.
Most of the often touted amazing general purpose AI is largely just a lot of smoke and mirrors. The main reason for that is another thing that also hardly changed since the early 90s: a lot of profit is made from convincing (/fooling) investors into believing that AI is far more powerful/useful than it actually is.
I think this AI is onto something
As brilliant as it is, I think this speaks more to how we as humanity think about ourselves than it does about AI.
> The point of this is to form a hypothesis. If the scientist and the experimental data say the same thing, then the scientist will think it has a hypothesis that is correct. But if the experimental data and the scientist say different things, then the scientist will think it has a hypothesis that is wrong. The scientist will think the experimental data is right, and it will change its hypothesis.
Imagine say some trying to maintaining training sets per individual character and finding that they would not only provide better lines but choose different actions.
Edit: anecdotes -> analogies
