Gary Marcus: it was the harness, not the model — hidden risk of opaque scaffolding
GaryMarcus · x · 2026-09-09
Gary Marcus amplified a take arguing that a claimed capability came from the harness, not the model — meaning the real capability lives in opaque, vendor-controlled scaffolding that holds state and executes actions, a bigger hidden risk than benchmark scores.
Marcus calls it more sneaky use of unacknowledged domain-specific neurosymbolic AI doing the real work. The commenter warns enterprises handing execution to such a black box are no longer just using a model.
More from AGI Musings
- AI-assisted writing will become the norm, making 'hand-made' text indistinguishable — dbasch · 2026-09-09
- Researcher warns against turning mathematics into an AI benchmark — konstmish · 2026-09-09
- Former OpenAI researcher Aidan Clark: for the first time I'm asking if AI is moving too fast — _aidan_clark_ · 2026-09-09
- willcb: A 'niche' approach may turn out to be the right way if search scales up — willcb · 2026-09-09
- Dev: AI has advanced so much you need months of study to parse frontier problem statements — zetalyrae · 2026-09-09
- OpenAI signals it may deliberately pace capability advances after next-gen model's math breakthrough — OpenAI · 2026-09-09