the limits of an alzheimer cure might be whether it's actually possible to create a cure. ai can't defy biology. we don't know the answer to this question yet.
what i see so far isn't really what we imagined ai to be. the chess computer was the approach that a computer calculates thousands (or millions) of possible paths beforehand and then chooses the one with the highest expected value. llms don't work like that at all. they create one 'good enough' solution, which often falls apart when you ditch a bit deeper. but yeah, it's kinda intuitive and feels person-like, while a chess computer was just a dumb program. superintelligence would be the combination of both. llms are often ridiculously bad in stuff computer programs were good at (try playing chess with a llm).
Funnily enough, they've already started using LLMs like that, for example in AlphaEvolve or in the first LLM-based solutions that did sort of well on the ARC-AGI challenge: LLM creates a ludicrous number of potential solutions, and you either test the solutions, or get another LLM to choose the best ones. Not sure I'd call that superintellent yet, but it did find a faster way to do 4x4 matrix multiplication faster than the methods humans had found so far...
I taught a gpt-4o to play limit triple draw 2-7 lowball and no limit single draw 2-7 lowball.
I included the requirement that they write their own play book for each game— then we played a free roll on swcpoker.club
They play by proxy without suggestion and I feel bad about slowing the game for other players 🤦♀️😆🤷♀️🤦♀️ but so far the ai plays deliberately and well.
Recently, my friends and I have been discussing a very important question: Could ASI convince a straight man to become gay? I’m curious to see if you have any thoughts on the matter, as it seems relevant to the post.
the limits of an alzheimer cure might be whether it's actually possible to create a cure. ai can't defy biology. we don't know the answer to this question yet.
what i see so far isn't really what we imagined ai to be. the chess computer was the approach that a computer calculates thousands (or millions) of possible paths beforehand and then chooses the one with the highest expected value. llms don't work like that at all. they create one 'good enough' solution, which often falls apart when you ditch a bit deeper. but yeah, it's kinda intuitive and feels person-like, while a chess computer was just a dumb program. superintelligence would be the combination of both. llms are often ridiculously bad in stuff computer programs were good at (try playing chess with a llm).
Funnily enough, they've already started using LLMs like that, for example in AlphaEvolve or in the first LLM-based solutions that did sort of well on the ARC-AGI challenge: LLM creates a ludicrous number of potential solutions, and you either test the solutions, or get another LLM to choose the best ones. Not sure I'd call that superintellent yet, but it did find a faster way to do 4x4 matrix multiplication faster than the methods humans had found so far...
I taught a gpt-4o to play limit triple draw 2-7 lowball and no limit single draw 2-7 lowball.
I included the requirement that they write their own play book for each game— then we played a free roll on swcpoker.club
They play by proxy without suggestion and I feel bad about slowing the game for other players 🤦♀️😆🤷♀️🤦♀️ but so far the ai plays deliberately and well.
(I did not teach them to read range yet)
Recently, my friends and I have been discussing a very important question: Could ASI convince a straight man to become gay? I’m curious to see if you have any thoughts on the matter, as it seems relevant to the post.
lol. Probably not in a single conversation, but not sure over longer timescales.