NYU Professor: LLMs Can't Execute Algorithms Exactly, But Can Act as Agents to Call Tools
tallinzen · x · 2026-08-05
Responding to a Quanta Magazine article on whether LLMs can truly reason, NYU linguistics professor Tal Linzen shared his perspective backed by recent research.
Linzen noted that if "reasoning" means "following an algorithm exactly," transformer-based models fail in chain-of-thought, making increasing errors as problems get more complex. However, he argues this isn't a fatal flaw: by 2026, mainstream AI systems are no longer just standalone transformers but operate as "agents." It might be sufficient for the transformer to recognize its limits and write a Python function to execute the algorithm. Still, he warns that the brittleness of chain-of-thought reasoning will continue to impact model robustness and safety as contexts grow longer and problems become more complex.
Related event: Quanta Magazine Explores True LLM Reasoning(6 posts)→
More from AGI Musings
- Top Scholars Launch Pax Machina to Debate Institutions for Powerful AI — edelwax · 2026-08-05
- AI as the Boss: Agent-Managed Cafe Employs Five Humans — luke_drago_ · 2026-08-05
- AI Buyer Calls AI Receptionist: Autonomous Agents Conduct Business Call — Admir-Rusidovic · 2026-08-05
- AI Will Generate Millions of Games; the Hard Part is Getting Played — Proud-Marsupial-6696 · 2026-08-05
- 1,346 Frontier AI Employees Petition Government for International AI Slowdown Mechanisms — deanwball · 2026-08-05
- Gary Marcus: Top AI Models Fall Short in Writing and Theory of Mind — GaryMarcus · 2026-08-05