Ok I'll wait for the final draft to post the publication but time to explain what we're doing because I'm really excited about it. This is our ICCM 2025 paper. We're making polynomial algebraic types learnable! #cogsci #neurosky #AI
Comments
Log in with your Bluesky account to leave a comment
Big question: why? Humans are remarkably efficient at spontaneously acquiring novel skills in unfamiliar settings. "Skills" we traditionally describe as procedures for satisfying some goal; in other words, skills are programs. We can define skill acquisition as a sort of program synthesis.
Is this related to @fchollet.bsky.social's work? Yes. It's also connected to Hutter's "universal intelligence", an interesting recent finding from @standehaene.bsky.social's group (linked) indicating that the programs we devise tend to be convergently space-optimal
This is additionally supported by some old work my former supervisor was a part of showing that the pace of skill acquisition vis-à-vis *time* optimality is optimal up to a polynomial term, which beats existing neural network scaling laws by a lot, particularly scaling laws for test time compute
*Unlike* Chollet, we are claiming there's a line between general and sub-general intelligence, and it's defined by being well-modeled by a time optimal finite approximation of universal intelligence. The implication here is that most useful procedures will not behave like brute force...
... therefore, we can impose a strong constraint on our search space, restricting it to only polynomial-time programs. This motivates us to turn towards type theory.
Comments