AMI's world-model thesis and the search for external validation
Researcher Randall Balestriero took the JEPA framework to Harvard, arguing publicly that the field needs more mathematical rigor to advance world models.
Balestriero at Harvard: AMI is actively evangelizing JEPA theory, not just building it
researchBrown University's Randall Balestriero's post on 11 September announcing a Harvard talk on "world models, the Le* family and why we need more theory and mathematics to advance JEPAs" is a compact statement of AMI's current intellectual stance. The framing—that the community needs more rigorous theory—is a positioning move as much as a research claim: it implicitly critiques the empirical-first, scale-first approach that dominates the LLM mainstream and reasserts the JEPA program as a principled alternative.
This public advocacy role for Balestriero complements what the GitHub record shows: Quentin Le Lidec reviewed a pull request in the `galilai-group/stable-worldmodel` repository this week, suggesting active engineering work on world-model infrastructure is running in parallel with the theoretical outreach. The combination — a researcher on stage arguing for the mathematical foundations of JEPAs while an engineer reviews world-model code — points to a team that is simultaneously building and recruiting intellectual consensus around its approach.
- Julia Kempe's wry congratulations to OpenAI for "not letting any good customer transcripts go unmined" — alongside praise for AI-assisted mathematics — drew 295 likes, suggesting her commentary on AI ethics and capability trade-offs resonates well beyond AMI's immediate research circle. Tweet
- Min Lin opened a new branch in the `torch2jax` repository focused on a lazy module constructor, a low-level bridging tool between PyTorch and JAX. Sustained work here would suggest AMI is hedging its framework dependencies or building infrastructure for more efficient JEPA training pipelines. torch2jax repo