> The improvement is slow for now, but undeniable. Developing superintelligence is now in sight.
Oh, is it now? So you know for a fact that intelligence comes from token prediction, do you, Mark?
Look, multi-bit screwdrivers have been improving steadily as well. I've got one that stores all it's bits in the handle, and one with over three dozen bits in a handy carrying case! But they're never going to suddenly, magically become an ur-tool, capable of handling any task. They're just going to get better and better as screwdrivers.
(Well, they make a handy hammer in a pinch, but that's using them off-spec. The analogy probably fits here, too, though.)
My POINT, to be crystal clear, is that Mark is saying that A is getting better, so eventually it will turn into B. It's ludicrous on its face, and he deserves the ridicule he's getting in the comments here.
But I also want to go one step further and maybe turn the mirror around a bit. There's also an odd tendency here to do a very similar thing: to observe critical limitations that LLM tools have, that they have always have, and that are very likely baked into the technology and science powering these tools, and then to do the same thing as Mark, to just wave our hands, and say "But I'm sure they'll figure it out/fix it/perfect it soon."
I dunno, I don't see it. I think we're all holding incredible screwdrivers here, which are very impressive. Some people are using them to drive nails, which, okay, sure. But acting like a screwdriver will suddenly turn into precision calipers (and a saw, and a level, and...) if we just keep adding on more bits, I think that's just silly.
Isn't it though? He's provided zero evidence to suggest otherwise. So of course we are all going to assume he's talking about the current, popular, SOTA architectures still as the foundational piece.
>So you know for a fact that intelligence comes from token prediction, do you, Mark?
He never said that for a start nor much of the other stuff. I mean I can accuse him of say pigs fly but it's kind of silly if he didn't say that.
I mean it's like before powered flight someone says powered flight is coming that the critic say how can you claim steam engines will take off? It's kind of silly and about the same as the superintelligence won't happen because LLMs aren't very good argument.
It's perfectly reasonable for critics of powered flight to point out that steam engines will never be truly useful for the job if that's the only kind of attempts at powered flight that anyone's shown, and they've been claiming that steam engines WILL actually get good enough, just you wait and see!
Of course we can't predict rapid introductions of heretofore-unknown technology, but that doesn't mean that every scam artist and huckster should get a free pass because there MIGHT be a shiny new technology just about to be discovered that will retroactively make their wild claims sound prescient.
Oh, is it now? So you know for a fact that intelligence comes from token prediction, do you, Mark?
Look, multi-bit screwdrivers have been improving steadily as well. I've got one that stores all it's bits in the handle, and one with over three dozen bits in a handy carrying case! But they're never going to suddenly, magically become an ur-tool, capable of handling any task. They're just going to get better and better as screwdrivers.
(Well, they make a handy hammer in a pinch, but that's using them off-spec. The analogy probably fits here, too, though.)
My POINT, to be crystal clear, is that Mark is saying that A is getting better, so eventually it will turn into B. It's ludicrous on its face, and he deserves the ridicule he's getting in the comments here.
But I also want to go one step further and maybe turn the mirror around a bit. There's also an odd tendency here to do a very similar thing: to observe critical limitations that LLM tools have, that they have always have, and that are very likely baked into the technology and science powering these tools, and then to do the same thing as Mark, to just wave our hands, and say "But I'm sure they'll figure it out/fix it/perfect it soon."
I dunno, I don't see it. I think we're all holding incredible screwdrivers here, which are very impressive. Some people are using them to drive nails, which, okay, sure. But acting like a screwdriver will suddenly turn into precision calipers (and a saw, and a level, and...) if we just keep adding on more bits, I think that's just silly.