Andrew Lampinen challenges Gary Marcus on what really counts as neurosymbolic AI

AndrewLampinen · x · 2026-07-24

Andrew Lampinen is pressing Gary Marcus on what should count as neurosymbolic reasoning.

He asks whether a pure autoregressive LM, without a harness or tools, can solve novel math problems at all, or merely do worse than a tool-augmented system.
He also asks whether a system becomes “neurosymbolic” if the LM writes a program or calls a solver while proving something.
A further question is whether intelligence partly lives in the classical symbolic part, and whether the same conclusion would hold if a human wrote the helper program.
He extends the argument to multi-LM harnesses where models answer, review, and vote on each other’s outputs, asking what evidence would settle the issue.

More from AGI Musings