@YourNetworkIsHaunted - eviltoast
  • 0 Posts
  • 648 Comments
Joined 1 year ago
cake
Cake day: March 22nd, 2024

help-circle


  • I actually like the argument here, and it’s nice to see it framed in a new way that might avoid tripping the sneer detectors on people inside or on the edges of the bubble. It’s like I’ve said several times here, machine learning and AI are legitimately very good at pattern recognition and reproduction, to the point where a lot of the problems (including the confabulations of LLMs) are based on identifying and reproducing the wrong pattern from the training data set rather than whatever aspect of the real world it was expected to derive from that data. But even granting that, there’s a whole world of cognitive processes that can be imitated but not replicated by a pattern-reproducer. Given the industrial model of education we’ve introduced, a straight-A student is largely a really good pattern-reproducer, better than any extant LLM, while the sort of work that pushes the boundaries of science forward relies on entirely different processes.








  • I mean, it’s obviously true that games have their own internal structures and languages that aren’t always obvious without knowledge or context, and the FireRed comparison is a neat case where you can see that language improving as designers have both more tools (here meaning colors and pixels) and also more experience in using them. But also even in the LW thread they mention that when humans run into that kind of problem they don’t just act randomly for 6 hours. Either they came up with some systematic approach for solving the problem, they walked away from the game to ask for help, or something else. Also you have the metacognition to be able to understand easily “that rug at the bottom marks the exit” once it’s explained, which I’m pretty sure the LLM doesn’t have the ability to process. It’s not even like a particularly dumb 6-year-old. Even if it’s prone to similar levels of over matching and pattern recognition errors, the 6-year-old has an actual conscious brain to help solve those problems. The whole thing shows once again that pattern recognition and reproduction can get you impressively far in terms of imitating thought, but there’s a world of difference between that imitation and the real deal.


  • Also I think he doesn’t understand MAD like, at all. The point isn’t that you can strike your enemy’s nuclear infrastructure and prevent them from fighting back. In fact that’s the opposite of the point. MAD as a doctrine is literally designed around the fact that you can’t do this, which is why the Soviets freaked out when it looked like we were seriously pursuing SDI.

    Instead the point was that nuclear weapons were so destructive and hard to defend against that any move against the sovereignty of a nuclear power would result in a counter-value strike, and whatever strategic aims were served by the initial aggression would have to be weighed against something in between the death of millions of civilians in the nuclear annihilation of major cities and straight-up ending human civilization or indeed all life on earth.

    Also if you wanted to reinstate MAD I think that the US, Russia, and probably China have more than enough nukes to make it happen.