David Poole

What @Ylecun got right (and wrong) while under siege the last several days. He was 👉Wrong to be vague about what he thought could not be done; he should have said eg, “reliably, physical accurate prediction” instead of leaving it open that he merely meant make nice-looking videos 👉Wrong to combine a vague claim about what he could do with an advertisement for his own not-yet-proven model 👉Wrong to talk about a “generative” part of the brain, with no clear reference either to what meant by “generative” or any particular anatomy, particular when most neural processes are distributed, and “part of the brain” talk is largely obsolete. But he is right that 👉Generative AI is not rich enough to solve AGI. 👉You can’t get physics right just by predicting sequences of frames, because the world is too complex 👉We are unlikely ever to get physics right without internal models. And here’s the the kicker 👉The pattern of successes and failures for Sora bear all that out. Sora had an immense amount of data, and access to immense compute, and yields superficial beautiful images. Yet it still doesn’t get the physics (or biology) correct. The pure, model-free approach was just given an enormous shot, perhaps a $1b in resources, insane amounts of data, and still has objects winking in and out of existence, violations of gravity, and more. That is *exactly* why we should consider alternatives, LeCun’s or otherwise.

When I started working on AI four decades ago, it simply didn’t occur to me that one of the biggest use cases would be derivative mimicry, transferring value from artists and other creators to megacorporations, using massive amounts of energy. This is not the AI I dreamed of.