Gary Marcus doubles down on 2020 thesis: LLMs alone aren't enough for robust AI
GaryMarcus · x · 2026-09-30
Gary Marcus reposted an endorsement of his 2020 position, arguing that while LLM utility and depth of understanding have improved, his core claim still holds: large language models alone are not enough.
The referenced paper, "The Next Decade in AI" (arXiv:2002.06177), argued that:
- AI research was over-indexing on general-purpose learning, bigger datasets, and more compute
- A hybrid, knowledge-driven approach centered on reasoning and cognitive models is needed for robust AI
- Marcus noted figures like LeCun and Bengio had converged on the view that LLMs are insufficient, with deeper understanding and richer cognitive models as the key to the next decade
More from AGI Musings
- Dario pushes capability, Sam pushes dirt-cheap intelligence — who wins? — haider1 · 2026-09-30
- Vicki Boykis Keynotes on AI Anxiety: Why Good Engineering Still Matters — vboykis · 2026-09-30
- MIT prof Phillip Isola on PhD vs industry: what matters is the spirit of open inquiry — CSProfKGD · 2026-09-30
- OpenAI's 18-month cyber/bio risk deadline quietly expired with nothing to show — xeophon · 2026-09-30
- RSI Means Different Things to Different People, and There's No Benchmark to Prove It — Acceptable_Stress154 · 2026-09-30
- A Farmer-Writer Proposes One Clause for the AI Age: 'AI Robots Shall Not Build or Repair AI Robots' — Wrong-Expression6981 · 2026-09-30