Gary Marcus

Rank 19 of 47
|
Score 64
12/21/2024, 3:06:31 AM
X
In reply to:
Colin Fraser
@colin_fraser
·
195d
Anyway you gotta admit $1000 per arc problem is insane. Sorry no offence but for that price I want 100%.
Colin Fraser
@colin_fraser
·
195d
The efficient reliable verifier is roughly what I think is the missing piece.
Colin Fraser
@colin_fraser
·
195d
What’s really wild to me is not as much that you can bop around token space randomly for an hour and stumble on the right answer, but more that you can continue bopping around and verify that the answer is right. But this also strikes me as an inefficient and unreliable verifier.
Colin Fraser
@colin_fraser
·
195d
I won’t be surprised to learn that o3 still loses at tic tac toe and thinks 5.11-5.9=0.21 and also produces perfect solutions to graduate level math problems and I have no idea what that will mean
Colin Fraser
@colin_fraser
·
195d
FWIW I know I’m a curmudgeon on here but I do believe there’s a lot of juice to squeeze out of inference time compute. It’s a different direction. I don’t know how much juice and I think somethings still missing, but I don’t think it’s all smoke and mirrors. It is very mysterious

The statement is a lighthearted and conversational remark about the capabilities of a system in making illegal moves in chess. It does not engage substantively with public issues or contribute to civic dialogue.