I mean, it's an amalgam of text written by humans - it's not even trying to make a plane that flies, it's trying to make a plane that looks like all the planes designed by humans - and largely succeeding. I'm sorry that's not a task you consider interesting, but the fact that a machine learning model trained in an unsupervised fashion can succeed that much at anything non-trivial is mind-boggling to me, and makes me hopeful that other tasks will also be within reach relatively soon.
> trying to make a plane that looks like all the planes designed by humans [..] I'm sorry that's not a task you consider interesting
Trying to create text that looks like text written by humans? Is that really what GPT-3 is designed to do?
I'm reminded of one of my wife's responses, on being told that a person speaks N languages fluently ... "Does s/he have anything interesting to say in any of them?"
>> trying to make a plane that looks like all the planes designed by humans [..] I'm sorry that's not a task you consider interesting
> Trying to create text that looks like text written by humans? Is that really what GPT-3 is designed to do?
My understanding is, it's designed to attempt to "understand" the inner structure of text written by humans, and to create new text using the same structure based on its "understanding". "Understanding" is only defined as the creation of a probabilistic model - it doesn't "understand" any meaning, only the structure of sentences and paragraphs, Yet, such a mindless machinary can demonstrate a wide range of interesting behaviors.
Using the plane analogy: If you feed the airplanes designed by humans throughout history into GPT-3, GPT-3 is able to produce an original design of convincingly-looking airplanes of all nationalities and types. For example, if you say "1960, turboprop, Britain", it gives you one, and it's also capable of mixing things, such as putting a jet engine on a biplane, or designing a hypothetical WW2 Allies plane using Axis military technology. If you are lucky, some of the planes come with semi-functional systems. Some might be able to ignite its engine before it explodes, and few are even capable of taking off before it explodes (and some are unable to take off, but if you put them in the mid-air, it may be able to travel for one minute before it explodes). And it's all done under unsupervised learning, and without any domain-specific knowledge on Newtonian mechanics, aerodynamics or aviation. All it needs is an astronomical number of blueprints collection.
To me it's an impressive achievement, even if it doesn't understand how to design airplanes.
I read somewhere that unlike GPT-2, GPT-3 "understands" addition. So if you feed it "74 + 129 = " it will come up with 203, even if that specific math problem never occurred in any of the text it was trained on. That to me is the most impressive result so far - not only is it possible, but also more efficient, for the model to actually encode the rules of addition instead of just memorizing the results for specific math problems. This to me seems like the dawn of insight and understanding, and I'm super excited to see what comes next.
It's said that GPT-3 shows rudimentary hints of "general intelligence"-like behaviors, it can really "solve" (not just memorizing answers) many puzzles without fine-tuning it by domain-specific training data, although the performance is limited without doing it.
David Chalmers [0] wrote,
> When I was a graduate student in Douglas Hofstadter’s AI lab, we used letterstring analogy puzzles (if abc goes to abd, what does iijjkk go to?) as a testbed for intelligence. My fellow student Melanie Mitchell devised a program, Copycat, that was quite good at solving these puzzles. Copycat took years to write. Now Mitchell has tested GPT-3 on the same puzzles, and has found that it does a reasonable job on them (e.g. giving the answer iijjll). It is not perfect by any means and not as good as Copycat, but its results are still remarkable in a program with no fine-tuning for this domain.
> What fascinates me about GPT-3 is that it suggests a potential mindless path to artificial general intelligence (or AGI). GPT-3’s training is mindless. It is just analyzing statistics of language. But to do this really well, some capacities of general intelligence are needed, and GPT-3 develops glimmers of them.
I've always considered language just another form of logic (a very fuzzy one of course). To say something sensical in any language requires (almost by definition) some logical consistency. Otherwise you get non-sequiturs and insane ramblings. The greater the coherency of a narrative (e.g. paragraph-to-paragraph like GPT-3, vs. just word-to-word like a Markov chain), the greater degree of consistent logical underpinnings are needed. So that a tool trained to produce coherent human writing by necessity has embedded in itself a tool for logical reasoning is not too surprising (not to say it isn't impressive!).
Specifically it's trained to, when given some text from the internet (presumed to mostly have been written by a human), predict what comes next (there's a lot of details of exactly how you express this mathematically, but that's the basic summary). This is the objective function it is optimised to maximise while it is being trained, how close did the prediction match the actual text which came next.
This is done mostly because it's very easy to get a huge amount of data and score its performance numerically on that data without any manual process of deciding what the correct answer is. It turns out given enough data and a large enough network, it becomes very good at it, even to human eyes.
> the fact that a machine learning model trained in an unsupervised fashion can succeed that much at anything non-trivial is mind-boggling to me, and makes me hopeful that other tasks will also be within reach relatively soon.
To me, on the contrary, GPT-3 is quite a saddening thing. With every example I come across, I get a more and more clear idea of the limits of current mainstream approaches in AI.
When it comes to predicting or achieving human-level general behavior it seems to me that the SOTA is reaching a hard limit that lies far below actual AGI.
I'm not a researcher in this field, so this opinion should be taken with more than a grain of salt, but if I'm right, 10 years from now we're still gonna be stuck at GPT-3-ish levels of clever mimicry.
Back when I took my cs degree I learned about the AI winter - the decades that came after we came up with effective planning, reasoning and problem solving algorithms, but realized we couldn't use them on anything in the real world, because it is fuzzy and imprecise and we don't have an interface to put it into the clean mathematical world of those algorithms. I think basically, we realized that the higher reasoning and planning capabilities of a human mind that we were emulating were worth nothing without the fuzzy pattern recognition capabilities of our reptile brain to recognize and categorize objects for us to even reason about. The way I see it, the deep learning revolution of recent years is basically providing us with that missing link, the fuzzy recognition engine that serves as a foundation for higher reasoning. It still remains to be seen if this approach will scale to AGI on its own or it needs to be combined with others. But saying we are running up against a wall with that is like saying we have run up against the wall of what a visual cortex can do - the real fun begins when we combine it with other components, and those other components are still being developed.
