Sonia Joseph co-founds World Mechanics, a lab building interpretable foundation models for physical world
soniajoseph_ · x · 2026-10-02
Interpretability researcher Sonia Joseph and cvenhoff00 announced World Mechanics, a new frontier lab building interpretable foundation models for the physical world.
The core argument: LLM interpretability is necessarily post hoc — reverse-engineering trained models without ground truth — while physical AI is early enough to make interpretability a first-order objective of training. Physical systems also offer unusually useful data-generating processes: researchers can intervene in simulation or on real machines, observe responses, and exploit known physical structure.
More from Companies & People
- Mila hosts agentic AI community event on Nov 10, focused on reliable and safe agent deployment — Mila_Quebec · 2026-10-02
- Customer replaced Devin at ~25% of cost after quarterly-payment spat, ex-Cognition rep reveals — adrianscottcom · 2026-10-02
- Schulman: OpenAI leaks so much it should just embrace total research transparency — johnschulman2 · 2026-10-02
- How Matei Zaharia went from Berkeley PhD and Spark to a $190B Databricks — jfiance · 2026-10-02
- Anthropic said to file for IPO as soon as mid-November at $1.8-2T valuation — firstadopter · 2026-10-02
- Scaling a million sandboxes live on stage — graceisford · 2026-10-02