Gary Marcus: neurosymbolic AI has won — a model is an engine, not the whole car
GaryMarcus · x · 2026-11-16
In a long-form post, Gary Marcus argues you can't understand modern AI without distinguishing models from systems: any system combining neural networks with symbols is neurosymbolic by definition. He cites Lean-checked math and Claude Code as systems relying on rules, code interpreters, and loops — classical symbolic constructs — and says Hinton's dismissals of symbols as "aether" were wrong. His conclusion: neurosymbolic AI has won overwhelmingly; scaling alone was never enough.
Related event: Gary Marcus Sparks Debate: LLMs With Tools Are Neurosymbolic AI(9 posts)→
More from AGI Musings
- Writer using AI to rework drafts gets flagged as AI 100% of the time — aronchick · 2026-11-16
- AI takeover creating conscious observers could imply long timelines, not doom — jessi_cata · 2026-11-16
- "Ads are prompt injections for humans" — a developer's analogy that reframes both fields — maksym_andr · 2026-11-16
- AI safety researcher: NDAs can't hide reckless endangerment by AI companies — DavidSKrueger · 2026-11-16
- Does rapid observer-population growth plus SSA predict near-term doom? — jessi_cata · 2026-11-16
- YouTubers are firing human editors and replacing them with Claude, says AI commentator — ChrisGPT · 2026-11-16