2 ms·
The current foundational models have basic reasoning with glimpses of brilliant reasoning.
by riddlemethat 3mo ago
The current foundational models have basic reasoning with glimpses of brilliant reasoning.
- krona 3mo agoLLMs have almost no fluid intelligence or capacity for abstract inductive reasoning, relying on crystallized intelligence for basically everything.
- imdsm 3mo agoI think they have some fluid intelligence, if somewhat brittle
- singularity2001 3mo agoIf anything their capability of abstract inductive reasoning is way beyond the average human given how much better LLMs are at solving math problems, it's the paradox that they can do complicated reasoning before they can do intuitive peep-pe-boo.
- krona 3mo agoabstract inductive reasoning requires dynamic learning in unfamiliar contexts. All the benchmarks which measure this specifically (in particular, task composition) see LLMs fail catastrophically. There is quite a lot of research published on these limitations now.