Michael Tomasello on MLST: LLMs skipped the infrastructure of shared agency
Machine Learning Street Talk · youtube · 2026-11-19
Machine Learning Street Talk released a long interview with developmental psychologist Michael Tomasello, who spent twenty years at the Max Planck Institute running identical experiments on children and great apes, and is now at Duke. His book The Evolution of Agency asks what it takes to act for one's own reasons.
Key threads:
- Unpredictability breeds agency: where the environment is predictable, nature hard-codes behavior; where it isn't, it builds a decision maker — from thermostats and robot vacuums to reptilian control loops, a squirrel imagining itself falling and taking another branch, and apes that gather information before committing.
- Shared agency: humans form joint goals, share attention, and learn that one object has multiple perspectives; kids cut off from call audio invent gestures and turn them into conventions.
- On LLMs: they "started at the endpoint and skipped that infrastructure," which is why they don't create conventions.
Also covered: the stag hunt, why lying only works among cooperators, whether capital is an agent, why the Turing test is far from sufficient, and Dennett's intentional stance.
More from AGI Musings
- Caltech math professor stops grading homework because everyone uses AI — Distinct-Question-16 · 2026-11-19
- Economist Paul Novosad: game designers should build products that change us for the better — paulnovosad · 2026-11-19
- LATO raises $4.7M as its agents interview 29 experts in 5 days on Anthropic's $900B compute bet — annbordetsky · 2026-11-19
- Gary Marcus mocks market for spiking on news of another $100B into a company that may never profit — GaryMarcus · 2026-11-19
- The 90s vs 2026 husband meme capturing the AI companion shift — hey_abusiddik · 2026-11-19
- S.J. Murray essay: stripping friction from AI companionship strips love's power to transform us — AnnaCiaunica · 2026-11-19