5 ms·

As is being confirmed by this report, these large language models are absolutely building models of the real world, though not exactly at the abstraction layer

As is being confirmed by this report, these large language models are absolutely building models of the real world, though not exactly at the abstraction layer level we initially guessed. The only question is how thorough of a model it is, and what is the “waste factor”, I.e., how much more inefficient are these at learning the world compared to say the human brain. It’ll be tantalizing to see how GPT-4 performs but my best guess would be that there will be an even more amazing performance but not something that can truly match human beings (even a dumb one) without significant change to the architecture.

mattigames 4y ago
It doesn't matter, brains took millions of years to become what they are now, it wouldn't be a surprise that we may need to take a completely different path to make machines perform equally or better than a human using our current technically (incl future iterations/versions of it), including stuff that looked from a biological perspective would be "wasteful".
- pyinstallwoes 4y ago
  If we follow your logic you can say that whatever humans build is also a result of millions of years.
  - mattigames 4y ago
    Not really, that would follow my logic if I had said sometbing like "the universe took millions of years to create the human species in including the human brain", and that's not what I said, the baseline is an organism and it's ability to evolve, nothing before that. And anyway in this context "evolution took millions of years" is just another way to say that a huge amount of lesser brains died while the ancestors of the current brain survived, AKA self-selection; so clearly is unlikely to be any similar to the process we will take to improve artificial intelligence as we will measure it with a different standard and NOT over its ability to survive in the wild.
- irrational 4y ago
  For a machine to perform equally or better than a human, wouldn’t it need to be general intelligence? That is, it knows that it exists and can think about itself.
  - 082349872349872 4y ago
    I think it's the other way around: something with general intelligence, that is good at modelling others, is going to naturally wind up producing a(n imperfect) model of itself, en passant.
    - Ieghaehia9 4y ago
      Maybe it would be possible to create some kind of equilibrium/fixed-point based AI that, for everything it knows, it knows that it knows, knows that it knows that it knows, and so on. Then again, perhaps not, because PPAD is much harder than P; so just getting the AI to maintain the perfect self-knowledge invariant as it learns more things could be intractable. There might also be halting problem reductions if the AI is sufficiently powerful, although I'm not sure.
      - emporas 4y ago
        If it knows that it knows, then surely can program by itself a sufficiently large program, debug it, extend it and so on. That is an NP-hard problem, as the name suggests, it is not easy, which is to say that it is infinitely hard, i.e. impossible. Something that many people are surprised to realize, including me, is how much of our life can be be described by a context-free language. When we try to create a machine which knows-that-it-knows this is context-sensitive. Day and night comparison, Javascript vs CSS for example. Machines are currently able to compare, contrast and mimic images and text,a thousand times better than a human. Maybe even million times better than a human. Very soon this will be billion and trillion. This copying machine, copies only the context-free properties of the subject it analyzes. The moment it will try to copy context-sensitive properties, then the it hits the halting problem. No LLM currently has any problem terminating, so it analyzes only context-free properties of the subject. A human song, a painting or a novel, maybe copies only the context-free properties from one other song, painting or novel. We humans for centuries and millennia described it as creativity. It turns out it was not creativity. Machines are creative only because we conflated context-free art creation, and context-sensitive art creation. When that distinction takes place, then machines will not be described as creative anymore!
- fancyfredbot 4y ago
  Eyes took millions of years to evolve but we've built cameras which capture images just as well in a few hundred. Legs took millions of years to evolve but we've built walking robots. I'm not sure why brains are different?
  - mjburgess 4y ago
    "Capture images" is something cameras do, eyes do vastly more (baring on the neuroplascitity of the brain, and so on). We have never built a system capable of https://en.wikipedia.org/wiki/Autopoiesis https://en.wikipedia.org/wiki/Autopoiesis , and so, cameras are vaslty inferior in this respect, since they are wholey incapable of modulating the growth of an organism.
YeGoblynQueenne 4y ago
The certainty about these results reminds me of the house of cards we now understand was build in the social sciences over many years of shoddy, unreplicable work referencing other shoddy, unreplicable work. For example, the authors are careful, in the article above, to characterise the "probing" technique that they used as a "heuristic" and they even clearly say that there is no formal proof of its correctness: It’s believed that the higher accuracy these classifiers can get, the better the activations have learned about these real-world properties, i.e., the existence of these concepts in the model. "It's believed". The results reported in this article are conjecture, backed by belief, interpreting the unobservable state of a black-box model. It should be no surprise that the authors found what they clearly set out to find. And yet, already, there are multiple comments here on HN asserting that, hey, problem solved, we know that GPT's build world models. It is known.

mattigames 4y ago

It doesn't matter, brains took millions of years to become what they are now, it wouldn't be a surprise that we may need to take a completely different path to make machines perform equally or better than a human using our current technically (incl future iterations/versions of it), including stuff that looked from a biological perspective would be "wasteful".

pyinstallwoes 4y ago
If we follow your logic you can say that whatever humans build is also a result of millions of years.
- mattigames 4y ago
  Not really, that would follow my logic if I had said sometbing like "the universe took millions of years to create the human species in including the human brain", and that's not what I said, the baseline is an organism and it's ability to evolve, nothing before that. And anyway in this context "evolution took millions of years" is just another way to say that a huge amount of lesser brains died while the ancestors of the current brain survived, AKA self-selection; so clearly is unlikely to be any similar to the process we will take to improve artificial intelligence as we will measure it with a different standard and NOT over its ability to survive in the wild.
irrational 4y ago
For a machine to perform equally or better than a human, wouldn’t it need to be general intelligence? That is, it knows that it exists and can think about itself.
- 082349872349872 4y ago
  I think it's the other way around: something with general intelligence, that is good at modelling others, is going to naturally wind up producing a(n imperfect) model of itself, en passant.
  - Ieghaehia9 4y ago
    Maybe it would be possible to create some kind of equilibrium/fixed-point based AI that, for everything it knows, it knows that it knows, knows that it knows that it knows, and so on. Then again, perhaps not, because PPAD is much harder than P; so just getting the AI to maintain the perfect self-knowledge invariant as it learns more things could be intractable. There might also be halting problem reductions if the AI is sufficiently powerful, although I'm not sure.
    - emporas 4y ago
      If it knows that it knows, then surely can program by itself a sufficiently large program, debug it, extend it and so on. That is an NP-hard problem, as the name suggests, it is not easy, which is to say that it is infinitely hard, i.e. impossible. Something that many people are surprised to realize, including me, is how much of our life can be be described by a context-free language. When we try to create a machine which knows-that-it-knows this is context-sensitive. Day and night comparison, Javascript vs CSS for example. Machines are currently able to compare, contrast and mimic images and text,a thousand times better than a human. Maybe even million times better than a human. Very soon this will be billion and trillion. This copying machine, copies only the context-free properties of the subject it analyzes. The moment it will try to copy context-sensitive properties, then the it hits the halting problem. No LLM currently has any problem terminating, so it analyzes only context-free properties of the subject. A human song, a painting or a novel, maybe copies only the context-free properties from one other song, painting or novel. We humans for centuries and millennia described it as creativity. It turns out it was not creativity. Machines are creative only because we conflated context-free art creation, and context-sensitive art creation. When that distinction takes place, then machines will not be described as creative anymore!
fancyfredbot 4y ago
Eyes took millions of years to evolve but we've built cameras which capture images just as well in a few hundred. Legs took millions of years to evolve but we've built walking robots. I'm not sure why brains are different?
- mjburgess 4y ago
  "Capture images" is something cameras do, eyes do vastly more (baring on the neuroplascitity of the brain, and so on). We have never built a system capable of https://en.wikipedia.org/wiki/Autopoiesis https://en.wikipedia.org/wiki/Autopoiesis , and so, cameras are vaslty inferior in this respect, since they are wholey incapable of modulating the growth of an organism.

YeGoblynQueenne 4y ago

The certainty about these results reminds me of the house of cards we now understand was build in the social sciences over many years of shoddy, unreplicable work referencing other shoddy, unreplicable work. For example, the authors are careful, in the article above, to characterise the "probing" technique that they used as a "heuristic" and they even clearly say that there is no formal proof of its correctness: It’s believed that the higher accuracy these classifiers can get, the better the activations have learned about these real-world properties, i.e., the existence of these concepts in the model. "It's believed". The results reported in this article are conjecture, backed by belief, interpreting the unobservable state of a black-box model. It should be no surprise that the authors found what they clearly set out to find. And yet, already, there are multiple comments here on HN asserting that, hey, problem solved, we know that GPT's build world models. It is known.