The AMI Labs Observatory
One of AI's Most Watched Companies

The most consequential publication this week is AD-E2E-JEPA, dated September 28, which evaluates existing action-conditioned JEPA world models — including LeWM, DINO-WM, and JEPA-WM — for end-to-end autonomous driving. The methodological choice is telling: rather than train a policy, the authors use a goal-conditioned zero-shot planning setting, isolating world-model quality from policy learning. This makes world-model performance falsifiable on its own terms, and the paper's finding that existing JEPA-based approaches fall short is the kind of honest internal audit that precedes a new generation of the architecture.

This sits in direct conversation with Randall Balestriero's post on October 1, in which he publicly identifies a specific failure mode in LeWM — that applying SIGReg per time-step leaves temporal collapse unaddressed — and points to AMI's own prior work as the fix. The AD-E2E-JEPA paper and Balestriero's intervention are two faces of the same move: positioning AMI as the authority on where JEPA world models currently fail and what comes next.

• AD-E2E-JEPA paper
• Balestriero on temporal collapse in LeWM
• Balestriero keynote at MICCAI-FOMO26

Amir Bar's post on October 1 introduces FAR, described as a memory framework for world models that draws on multiple sensory cues — vision, audio — and takes explicit inspiration from retrieval-augmented generation in LLMs. AMI is not just asking how a world model predicts the future, but how it recalls and retrieves relevant past context to do so. This is precisely the kind of architectural extension that would be required to move JEPA-style models from short-horizon prediction toward the kind of persistent, multi-step reasoning needed for autonomous agents.

Paired with Julia Kempe's announcement of a World Modeling for Physics Workshop in Aspen — with applications closing October 15 — the signal is that AMI is simultaneously pushing the research frontier and convening the community around it. The physics workshop is notable in its own right: two papers attributed to LeCun's co-authorship this week (SCOPE and PDE-OBS) address recovering complete physical fields from sparse observations, suggesting that world modeling for physical simulation is an active thread, not merely a community-building gesture. Caveat: As always, these papers represent a mix of work begun before some team members were formerly hired by AMI, as well as outside research teams that include AMI employees via their university or research affiliations.

• Amir Bar on FAR memory framework
• Julia Kempe on World Modeling for Physics Workshop
• SCOPE paper
• PDE-OBS benchmark paper