Post by @red_tomato@lemmy.world

2026-10-29 23:42 UTC

I doubt models are “topping out “. It feels like no prediction about the supposed “fundamental limits” of AI models have held. Two years ago it seemed like LLMs were fundamentally incapable of doing maths. It’s just a next word predictor! Of course it can’t do maths! Now it’s solving high profile maths problems left and right faster than any mathematician alive. Still a next word predictor.

Replies (3)

@FlashMobOfOne@lemmy.world 2026-10-29 23:55

Now it’s solving high profile maths problems left and right faster than any mathematician alive. This is a controversial take. It’s not unlikely that the models simply managed to steal users’ research.

Open ##5489283
@brucethemoose@lemmy.world 2026-10-30 03:36

Reasoning proofs, using formal proof languages, and tool assisted math were always in their realm. The pure LLM is still really bad at the “calculation” part though. In fact it’s worse than predicted: many thought byte-level models would improve that, and… those never came to be.

Open ##5489739
@criss_cross@lemmy.world 2026-10-30 01:40

I mean the underlying tech still can’t do math. It just has the tooling to call a python script to do it. If you don’t let it execute commands or call tools it can’t do math. Also I feel like part of the reason they’re “topping out” is cost. We’re getting “reduced rates” here and it’s still astronomically more expensive to let one of these run free than it is for a human to use one in a targeted fashion.

Open ##5489767