> 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 still believe that there is never going to be meaningful NLP without a model/knowledge base about the real physical world.
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.
Good catch! The coherent structure was what got me excited. If the structure turns out to be the product of human selection -- and now that you point out the plural on "responses" I think that's likely -- then these results are much more in line with my expectations.
I wish I could see the original of this - this quote “As with previous posts, I picked the best responses, but everything after the bolded prompt is by GPT-3“ could mean anything from a minor improvement to this being essentially human written, and there’s no way to tell.
With GPT-3, I guess lots of upcoming stories of various discussion forums getting the "Sokal Affair"[1] treatment. We'll keep amusing each other by trolling everybody with more fake GPT-3 stories.
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.
> 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.
If you wanted to generate a longer essay by GPT3 but make it really hard to tell, you could just get it started with a 1-2 sentences. Then let it generate N-1 tokens (where N is the window size, I guess 2048), and write a few tokens yourself (and repeat). Then I don't think there's any way to reverse engineer the internal state of the model at any point in the text. It would be like trying to find a string that hashes to a given value.
There is one fringe ironic and probably unrealistic failure case with that approach - what if GPT3 matches expectations too well with actual people and results in good examples being tossed in preference for posts which don't fit with most like say blatant spambots?
But only the best examples of GPT-3 output will be worth putting on the web, so even without any code improvements in future GPT-x models you can still expect an improvement in quality. A writer (human or AI) should be able to improve by writing a million articles and then be told which ones are good (i.e. worth publishing).
I’m continually amazed - flabbergasted - by GPT-3. I’ve read stories, articles and HTML written by it and each time I am shocked at how good the output is. This essay made me laugh!
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.
If you pay just a little attention, you absolutely can: GPT-3 is not saying anything. Even the 'lowest' humans are usually trying to communicate something when they are telling a story, or teaching you how to do a basic skill, or giving you directions.
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.
GPT's text as quoted in the article has a clear thesis stated upfront, later expands on it with examples, summarizes its arguments at the end, and does all this with gusto.
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.
My comment was not an essay, it was a response to someone else's comment. The GP was explicitly saying that GPT-3 could probably produce many of the news content they read, and I was replying to that.
Please also note that it's not very clear to what extent the text of the article is edited - the non-bold text is written by GPT-3, but I don't think it was produced as a single block of text. Instead different parts (sentences? Paragraphs?) were produced individually, selected by a human from many other responses, and assembled together in the shape we are shown. The train of thought among the paragraphs is most likely entirely human work, not AI work, and only the best sounding paragraphs out of a lot of gibberish were likely selected.
It would also be interesting to see how close those paragraphs are to something in GOT-3's training corpus, in terms of structure if not explicit language.
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.)
I do wonder though how close the articles that GPT-3 produced were to the articles it had been trained on. For example, the Methodist Church Split article that it produced has a lot of very specific facts about the Methodist Church and about the split, which shows that it had texts about that spefic event in the training set.
It also has a sentence which contains a pretty obvious non-sequitur, but it's easy to miss it or assume that it's a mistake that a human made.
So overall, I'm guessing GPT-3 may actually be pretty decent at re-telling a story with different words, which sometimes is very hard hard to distinguish from a human doing the same thing.
They also don't describe the way they programmatically selected the output, though I am willing to believe that they more or less randomly sampled the output from each model.
>> 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.
I disagree with tsimionescu that this is a good counterpoint. The comment you reply to says that "GPT-3 is not saying anything". The figure you refer to shows that human judges could not tell the difference between human-generated and GPT-3 generated text. That's apples and oranges. That some humans weren't able to detect autogenerated text doesn't say anything about whether the autogenerated text said anything.
It may or may not be smoke and mirrors, but there's a clear structure to the essay, as well as a logical structure to the arguments in the essay. For example, "humans think they're intelligent", "humans are wrong about everything", and therefore "intelligence isn't about being right".
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.
As I stated elsewhere, the essay is almost certainly constructed by the human writing the article, out of cherry-picked output by GPT-3 stitched together (I'm guessing at the paragraph level).
Also, again almost certainly, the information about what it takes to pass the Turing test is taken from some text in the corpus it was trained on. What GPT-3 did do is recognize that that is relevant to the topic of AI vs human intelligence, but not much more.
> If you pay just a little attention, you absolutely can: GPT-3 is not saying anything. Even the 'lowest' humans are usually trying to communicate something when they are telling a story, or teaching you how to do a basic skill, or giving you directions.
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.
> 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'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.
> 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.
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 ….
I feel like I can understand it (GPT-3 generated text) better than meme-oriented Reddit callback threads. Which may have more to do with burying meaning more than explicating it.
I think the hurdle is going to be getting it to write accurately. For example, if you want it to write a news article, ideally you want to give it some facts (this event happened, x people died) and it pumps out some accurate prose about it. I'm not sure you can trust GPT-3 to do that. It could produce something that sounds good, but there is not really a guarantee it is not spewing falsities that came from it's training data.
I can't tell these examples from human either. I half suspect that Arram Sabeti is running a genuine Turing Test and that many of these outputs were actually written by humans.
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)?
This one seems a bit to good to me to be GPT-3, but it does have some of the general GPT-3 signs.
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.
Yeah I thought it was a rhetorically framed as GPT-3 but really written by a human writing in character as an offended strong AI calling humanity on their bullshit and tautologies. They emphaisze successes but failures dominate and the only consistency is chauvanism in only defining in likeness to success instead of anything relevant like being correct. The simplicity and weakness of the failures listing gave more of a "the AI is outright trolling humans by using patronizingly sophmoric examples" akin to a peeved climatologist starting off with the basics of the water cycle when explaining how global warming can lead to both more flooding and droughts.
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.
That sounds interesting, could you please expand on that? Do you have in mind any particular facet of general intelligence that a milestone cannot be defined for?
Ah apologies for having been unclear. No, not like that - I was thinking about more meta aspects of it:
- 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.
Impressive? Absolutely. Monetizable? Unclear to me, but probably somewhere within the vast ad/chatbot/garbage text generation service-scape.
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.
IMO it's (human control, at high levels) all over as soon as a major corporation or state finds a way to gain significant competitive advantage by handing over the reigns to machines. Absent sustained, vigorous opposition and punishment by an alliance of basically everyone else, we'll compete ourselves right into having no choice but to let AI run things, or else become subservient to and marginalized by those who do.
In other words by the time it's a choice, I doubt it'll even be a choice.
Think about that logic for a bit about distribution of expertise vs control. Reducto ad absurdum wouldn't that say we should get rid of doctors? They have an insurmountable competitive advantage if we personally aren't at least a canidate for medical school. Never in our lives could we do better than them.
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.
I'm not at all following what tall poppies and getting rid of doctors has to do with my view that we're not likely to choose whether machines start giving us orders, but rather to just find ourselves in that situation and with few other options in the nearish future. I'd liken it to the invention of the state, more than anything else, and similarly practically-unavoidable once the advantages are made clear. It's just something that'll happen to us as a result of competitive forces soon after it's made possible, not some deliberate choice we're all going to make collectively, let alone individually.
... either that or all this machine-learning and "AI" development stalls out and we never get much better at it or anything like it. I guess that could happen, too. Doubt it, but you never know.
What "stupid fear"? I'm not even sure it's a dark future, and, as I expect is now clear, I don't really think my or anyone else's opinion on that would affect whether it happens, anyway. It's either possible and so (damn near) inevitable, or not, so won't happen. What I don't think it is, is any kind of choice we're going to make.
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?
I think that one of the neat new narrative art forms that GPT-3 like algorithms will make possible is a prose cinema where the author is more like a director, prompting agents to react to one another. Like lining up billiards shots. I think a bunch of people are doing stuff kinda like that now at least in regard to poetry generation.
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.
Ditto, this seems much more coherent than previous GPT-3 output. Have people gotten better at prompting / selecting output? Or is this a fake GPT-3 output written by a human?
I have been trying out different prompts for a week now and this seems plausibly written by GPT-3. There's quite a bit of luck involved since each time you submit the same prompt you can get a different response. Some prompts produce garbage output and some good prompts produce a range of outputs from lame to amazing. There's quite of bit of cherry picking and selection bias involved. No one publishes the uninteresting responses and you don't read or comment on them. Still, I think this is all quite amazing and it seems like it's close to ready for commercialization.
Quick, someone train an AI on labeled GPT-3 outputs. They can be labeled as "good" "ok" "bad" "bad grammar" "convincing argument" etc...
There's a website, scribophile.com, for crowdsourced literary criticism. We should make something similar but for AI training. The key difference is that the critique of AI output would be more structured (in the form of labels you can apply to sentences, words, paragraphs, etc...that you highlight) than unstructured.
Another website that is more structured than scribophile but is used for crowdsourced photo critiques instead is photofeeler.com. An interesting thing they do is that if someone is consistently "harsh" or "lenient" on photos, as measured by their standard deviation from the average rating, their feedback is adjusted accordingly. My email is in my bio for collabs.
Note: No one should dare submit GPT-3 content to scribophile. It is a beautiful, sacred, and fragile place for humans only.
Whenever people talk about how GPT-3 can’t do a lot of things that humans can do, I always think back to the “Bitter Lesson”. I don’t want to believe that general AI will just come from a stupid amount of compute, but it might.
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.
here's a parragraph in which I replaced "brain" with "scientist" and mouth with "experimental data"
> 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.
One potential area for the future is "augmented writing" where writers aren't as we think of today but more editors who feed in prompts and possibly rearrange and tweak to get better results than just their meat brains could come up with. There would be a diversity of styles and approaches of course.
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.
Related question that I don't know how to quickly find on Internet: imagine that IQ is a good measure for intelligence. I read Ainan Celeste Cawley[1][2] has an IQ of 263 (again, believe this number is accurate for a moment). How do you measure an IQ of 500, 1000, or 5000? I mean, not the actual test but how the test structure would change from measuring normal and outlier IQs?
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?
There's a few measures of "IQ". The most common one today centers at 100, and defines a standard deviation of separation on the task of being "generally intelligent" as 15 points. An IQ of 263 is a claim that that person is nearly in the 11th standard deviation above average, which corresponds to a claim that they are roughly 1 out of 6 × 10^26. This is not a plausible claim. The number is meaningless.
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.
Please forget one moment the definition of intelligence or if IQ is the right measure: how do you imagine a superintelligence in comparison with an intelligent human? I mean it is obvious that superintelligence is not, just, about processing information faster but has some structural changes and the ability to connect dots in different abstraction layers. Just guessing. I imagine that one achievement will be solving math theorems starting with math principles in the way AlphaZero achieves chess or go mastery.
My third paragraph covered that, although less intentionally than you may have wanted. I don't think you just get things that are monotonically better on the one and single "intelligence" axis. I think you get an ever-increasing number of ways to be intelligent. I can easily imagine am intelligence that is great at solving physics but rubbish at dansaeltenzing. I don't know what dansaeltenzing is... I'm not intelligent enough.
IQ can only really be calibrated relative to a specific test for it. I don't think there are tests for which even a perfect score would yield a value above 160, since the scoring is basically what percentile of the population you are scoring higher than. It's not like the tests are even modeled as approximations to some "ideal" test; they are fundamentally about trying to measure deviation against the population at large.
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.
Another annoying GPT piece where there is no ability for a regular member of the public to verify it. I guess in applying to the beta I should have said under 'what do you plan to do with this' - "post on social media cherry-picked examples that hype up GPT-3".
Highly debated argument though; to me it's like saying you CPU doesn't understand HTML and your browser is running on a CPU, hence it can't understand HTML either. Scott Aaronson explained it nicely too: https://scottaaronson.com/democritus/lec4.html#:~:text=Searl... . Even the wikipedia page mentions many reasonable counter-arguments.
Did they train it on Stephen Fry novels? If we deepfaked this text onto his voice and image, I think we might have something better than how Martin Amis turned out.
I mean, it's not going to win any philosophical debates, but these kinds of results are WORLDS better than they would have been just a couple of years ago. I have to wonder what would it take to impress you.
Could you elaborate on why you aren't impressed? This particular example doesn't suffer from either of the main GPT-3 "tells" I've noticed up to this point (weak backbone and nonsensical assertions).
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?
wonder if one day a forum like hackernews appears where gpt-x bots are posting comments on cool articles and blogs created by the very same bots. very deep complicated topics are discussed no human ever understand. If our progression in these fields will not come to a halt then this must happen one day.
There is pretty much one thing advances in AI tells us. Most of humanity is nothing more than a statistical approximation algorithm. But that doesn't mean human intelligence is. What is fundamentally lacking from modern AI is the ability to "invent". They can perfectly (at least very soon) approximate the behavior of "Joe". But they get nowhere close to even touching anything like the Einstein's of humanity.
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.
> They can perfectly (at least very soon) approximate the behavior of "Joe". But they get nowhere close to even touching anything like the Einstein's of humanity.
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.
> lacking from modern AI is the ability to "invent"
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.
> 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.
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.
I would argue that we would indeed be extinct. The phenomenon of apparent subjective experience is perhaps the most critical property of our existence, and based on the foundations of computer science as we understand them today, it is not necessary for computers to have a subjective experience in order to produce output that exceeds human capabilities.
I'm not sure I understand what you're getting at, so let me ask a clarifying question.
Lets suppose that we encounter a civilization that evolves independently outside of our solar system. This alien civilization is based on radically different chemistry - whatever it has that's analogous to a brain doesn't resemble any evolved natural brain on earth, it might not even be carbon based.
Nonetheless this alien civilization presents with language, culture, history, art and science. Would you argue that this civilization also lacks consciousness? How would you know these aliens do or do not have subjective experience, compared to human beings?
I'm wondering how particular you think the experience of consciousness is. Do you need to have carbon based, biological neurons to possess it? How do you determine if something has "subjective experience" or not?
“Homo sapiens” would be extinct, but our legacy and lineage would live on through whatever entities evolve from the first intelligent machines that we build.
I don't think this is very far from innovating. The foundation is building up an accurate mental model of the world, especially in the area you want to innovate in, and then running thought experiments, following through them step by step, and testing in the real world to make sure you're not straying too far. I think being able to write a new, coherent essay involves a lot of the same skills.
But the one building the coherent essay is the author of the article, not GPT-3. The author of the article is taking GPT-3 output that sounds good to them, and stitching it together into an essay.
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.
You may be right, I honestly don't know how this particular algorithm works. We may be closer to true AI intelligence than I think, but I am not sure anyone should wish for that. I am certain that we are still decades away from true human intelligence (Einstein level), and maybe a few years to a decade away from fully replacing the cognitive abilities of an average human. But it will come. And I am not sure the human race will be able to survive it.
I have no more evidence than you do, so take it with a grain of salt; but I think you're mistaken on the intelligence axis. As in, I think the distance from GPT-3 right now to average human, is way larger than from avg human to Einstein. And I agree we're not that far from replacing the work of an average human.
I think I agree with those points. However, the most interesting thing about GPT-3 in my mind is how it was largely a naive scaling up of GPT-2.
I would be very surprised to not see a GPT-4-like system built in the next several years by simply scaling up by 100x again (perhaps with some algorithmic improvements thrown in as well). I don’t know how it will perform but I expect it to be significantly more impressive than GPT-3. Which would be a very interesting thing... the world would look very different. Potentially even scary, especially with immediate follow-up reapplications of the techniques in other areas.
How about consciousness? On the surface GPT3 may appear to act like a human, but by almost any definition of you choose - it still lacks consciousness.
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".
Despite the supposed "mystery" of consciousness, it won't be hard to reproduce in an AI. Take a general intelligence system, give it the ability to sense its own mental state (e.g. by feeding the output of its intermediate layers back into its inputs). Pretty soon it will start building a conceptual representation of itself and the evolution of that representation over time will be conscious self-awareness.
I honestly don't think consciousness is necessary for AI. Consciousness is more like an "invention" of humans to make themselves feel special. If an AI can understand consciousness and simulate it, that will be sufficient. It doesn't have to be conscious to be intelligent.
> I don't think anyone would throw you a murder charge if you dumped the servers hosting GPT3 into the ocean.
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.
> 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.
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.
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.
That sounds like why old broadcast media is so frothing mad at the new - they have to try to please large bodies with lowest common denominator models to try to get the "average" to max yield.
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 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.