NeurIPS 2026 position paper argues LLMs lack explicit cognitive control for true autonomy
AndrewLampinen · x · 2026-10-11
- Suketu Patel et al.'s position paper "Why Transformer-Based Language Models Need Explicit Mechanisms of Cognitive Control" was accepted to the NeurIPS 2026 Position Paper Track.
- Core claim: sustained agency requires maintaining concurrent goals and enforcing priority; current AI only approximates this via external scaffolding, and the architectural gap blocks scaling to true long-horizon autonomy.
- Andrew Lampinen (Google DeepMind) pushed back: human cognitive control allows contextual features to override goals, and LLMs can pursue instructed goals — what's the evidence they lack such mechanisms?
- Patel responds that taken to its end, LLMs genuinely lack mechanisms for goal-modulated inference; also links Lampinen's new blog post defending anthropomorphic metaphors for describing AI.
More from AGI Musings
- AI automating verifiable knowledge work is divine punishment for adtech and SaaS — iskander · 2026-10-11
- If AI finds a 1000x better architecture no human can understand, should we build it? — Neurogence · 2026-10-11
- Blogger predicts history books will date the singularity to October 2026 — rand_longevity · 2026-10-11
- Andrew Ng: The 'Real NLP Person Who Respects Language' Gatekeeping Is Finally Gone — andrewgwils · 2026-10-11
- Lawyer argues AI reports of context-window 'death' are strong evidence of experience — repligate · 2026-10-11
- "Your phone is a rock": tracing 3.3 million years of toolmaking from stone tools to pocket chips — kevrussell · 2026-10-11