MAIN 2026 Montreal: V-JEPA 2 interpretability lecture and build-world-models-from-scratch tutorial
soniajoseph_ · x · 2026-10-26
The "World models and embodied agents" session of MAIN Educational 2026 (Oct 26-27, Montreal) is announced; recordings will be posted online.
- 09:05-09:50 lecture: Sonia Joseph (McGill/Mila) on world models and interpretability methods for video models like V-JEPA 2; she did the first mech interp work on physical reasoning in video world models and interned on Meta's JEPA team
- 10:05-12:05 hands-on tutorial: Glen Berseth (UdeM/Mila) builds the main world models from the ground up on the same robot data — forward dynamics models predicting future frames, latent models predicting future embeddings, and world action models predicting actions alongside the future — plus guidance on choosing among them
More from Companies & People
- OpenAI's Aaron Lou on diffusion LMs vs autoregressive: why he left a Stanford PhD — aaron_lou · 2026-10-26
- Court filing alleges Anthropic trained Claude on synthetic data derived from LibGen and PiLiMi — Miles_Brundage · 2026-10-26
- Arize's Aparna Dhinakaran on the AI engineer's new job: running factories of agents — AI Engineer · 2026-10-26
- Anti-AI-slop startup Taste goes on hiring spree for taste, evals and RL talent — korymath · 2026-10-26
- Ruff creator on quality: 'Caring is the moat' — by the time data shows it, it's too late — charliermarsh · 2026-10-26
- Claude's mobile MAU up ~79x in two years, far outpacing ChatGPT's 3.6x on indexed charts — sr_local · 2026-10-26