Comments by "Lepi Doptera" (@lepidoptera9337) on "Scaling LLMs further is an artform - Demis Hassabis (Google DeepMind CEO)" video.

There isn't. A random coin flip predicts the weather in many regions of the world with 50% accuracy or slightly better. A trivial stochastic model that knows nothing about atmospheric physics can probably do 60-70%. No amount of computing power we have thrown at that problem, so far, gets much past 80-90% as far as I know. In general computation we even have hard, theoretically provable limits like the halting problem. What AI has proven, so far, is that some problems that were deemed hard, like the generation of appealing images, are actually "quite easy". That's a statement about the average human taste, it's not so much a statement about the difficulty of general AI.
3
@endoflevelboss 99% of people can't see the obvious, hence I am bringing not just the table but the entire dining set. ;-)
1
Human chess has already changed greatly. I also don't see your point. My 20 year old car is faster than the fastest human over any distance longer than 60m and maybe 100m. Should we cancel all of the Olympic racing disciplines other than hurdles because of the existence of a machine that is faster?
1
 @tejassharma1849 I was suspecting some humorous thing going on, but the "which is better" question is actually a very sad side of the human affair. It's tribalism masking as a pseudo-sport.
1
@tejassharma1849 A chess game (including the set of all possible games) is a search problem that can be represented as a tree. The only question you are asking is how to cull the dead branches the most efficiently with the least amount of computation. A good deal of that has already been done by opening libraries and many endgames can probably be searched completely on state of the art hardware. That really just leaves the middlegame unsolved unless you want to introduce new rules... like the requirement that an algorithm should not search more than n-moves ahead or that only m evaluations are allowed in total (and the algorithm can chose how deep/broad it wants to search within that search volume). From a theoretical point of view chess is already a dead game (and probably always was).
1
@tejassharma1849 A chess game (including the set of all possible games) is a search problem that can be represented as a tree. The only question you are asking is how to cull the dead branches the most efficiently with the least amount of computation. A good deal of that has already been done by opening libraries and many endgames can probably be searched completely on state of the art hardware. That really just leaves the middlegame unsolved unless you want to introduce new rules... like the requirement that an algorithm should not search more than n moves ahead or that only m evaluations are allowed in total (and the algorithm can chose how deep/broad it wants to search within that search volume). From a theoretical point of view chess is already a dead game (and probably always was).
1
Throwing more resources at this doesn't make a whole lot of different. A system that can't do X for fundamental reasons still can't do X, even if you scale it up by a factor of 10 or even 1000.
1