AMI + JEPA: a unified theory of world models, from robotics to finance
There was substantial publishing activity this week as three interlocking papers on hierarchical planning, unstable-mode preservation, and multi-embodiment scaling, backed by GitHub activity and a wave of high-engagement social posts, all appeared. Together, they show the team believes it can extend, stack, and scale the JEPA framework to solve problems neither pure diffusion nor pure autoregression can. The launch of RoboJEPA was billed as the first scaling law for multiembodiment world models. That gives that thesis its most ambitious public test yet. Meanwhile, the quiet confirmation of a VP of Finance and the arrival of a founding self-supervised learning scientist signal that AMI is simultaneously hardening its organizational infrastructure.
JEPA goes hierarchical and robotic: AMI's architecture is being stress-tested at scale
researchThe most consequential cluster of work this week centers on pushing JEPA beyond its original single-timescale, single-embodiment form. H-JEPA's lead author is Wancong (Kevin) Zhang, a PhD student at NYU Courant, where Yann LeCun advises him. His research focuses on world models, planning, and representation learning, aiming to build agents that can learn from observation and interact with the world in an open-ended way. He became a Research Intern at AMI in May. The other four co-authors are Basile Terver, AMI Co-Founder Michael G. Rabbat, LeCun, and Randall Balestriero. That's notable for two reasons.
First, all five authors identify themselves as researchers at Advanced Machine Intelligence. Until now, most research published by AMI team members had legacy ties to previous employers, such as Meta. Even if this is research rather than a product, it's pretty clearly tied to AMI. And second, this is the latest of several papers that identify Balestriero as an AMI researcher. The Brown professor hasn't publicly announced he's been hired, but his long-standing work with LeCun points in that direction.
As for the JEPA paper, it introduces a hierarchy of action-conditioned JEPAs where each level predicts farther ahead in its own learned latent space and passes subgoals down to the next level. The claimed result — lifting Visual AntMaze success from 18% to 73% with three levels — is the kind of jump that, if it holds up, reframes long-horizon planning as an architecture problem rather than a data problem. The paper and code dropped on 5 October, and Terver's GitHub activity on the `kevinghst/H-JEPA` repository — specifically the merging and deletion of `droid_v2` branches — suggests the team was integrating real-robot (DROID dataset) experiments up to the last moment before release.
On the same day, Artem Zholus launched RoboJEPA with a social post that earned 564 likes and 104 retweets — by far the week's highest-engagement item. The project, which scaled JEPA predictors to 8B parameters across multiple robot embodiments and claimed to establish scaling laws that correlate compute with planning performance, now has its own GitHub organization (`robojepa/robojepa.github.io`, created 7 October).
A third paper, 'Preserving Unstable Modes Through Inverse Dynamics in JEPA World Models', provides theoretical grounding for why naive next-step JEPA prediction fails for robotic control and how inverse dynamics can fix it. Together, these three works read as a deliberate package: theory (unstable modes), architecture (H-JEPA), and scale (RoboJEPA) — each paper answering a question the others raise.
JEPA's surface area is expanding — finance and autonomous driving are now on the map
researchTwo papers this week extend the JEPA family into domains that have nothing to do with laboratory robotics. Randall Balestriero's NeurIPS-accepted 'Towards Financial World Modeling' applies JEPA-style reasoning to financial time series — noisy, non-stationary data with no ground-truth state — and the post announcing it (185 likes) frames LeJEPA as a practical investment tool. This is a deliberate provocation: applying a world-model framework designed for physical prediction to market dynamics claims the framework's generality, not just a quirky application. Separately, 'AD-E2E-JEPA' benchmarks existing JEPA world models (LeWM, DINO-WM, JEPA-WM) for end-to-end autonomous driving, identifies gaps, and proposes a SIGReg-regularised improvement. That paper lands alongside Balestriero's repeated public endorsements of SIGReg as the key stabilizer for end-to-end JEPA training.
The pattern here is strategic domain expansion through the same architectural lens. AMI is not building separate models for robotics, finance, and driving — it is arguing that one framework, properly extended, covers all of them. Whether that argument is borne out empirically is a question the coming months will answer, but the positioning is clear and deliberate.
A VP of Finance and a founding SSL scientist
hiringTwo personnel confirmations this week sit at opposite ends of AMI's organizational stack, and together they suggest a company moving from founding-team mode to structured institution. Mahmoud Assran joins as a founding scientist and member of technical staff in Montreal, focused on self-supervised learning and world models — exactly the research core that underpins every JEPA paper released this week. His arrival in Montreal also extends AMI's footprint toward Meta's FAIR Montreal hub, from which LeCun has long drawn collaborators.
More telling in institutional terms is the formalization of Chloe Giraut-Pierron as VP of Finance. She has actually been at AMI since November 2025, previously listed as director general, but her LinkedIn update to VP of Finance this week reflects a deliberate re-titling that points to growing financial complexity — the kind that requires a named executive rather than a founding generalist. Her background at Pigment, Qonto, and Stratumn maps onto exactly the fintech-adjacent financial operations infrastructure a scaling AI startup needs. Pascale Fung's public statement that AMI's mission 'goes beyond commercial success' reads, in this context, as a counterweight: a signal to the research community that the institutional build-out does not mean the science is being subordinated.
- Jiatao Gu's 'Lattice of Transition Laws' (published 8 October) proposes a unified theoretical framework treating diffusion, autoregression, and hybrid models as paths on a single corruption lattice — a design principle for decoding schedules that could inform AMI's own generative modeling work, particularly given Gu's NeurIPS acceptances for STARFlow2 and NTM announced the same week. Lattice of Transition Laws
- Ayaan Malik's post arguing that robotics startup 'quadrants' are investor taxonomy and that the real endgame is a single general model trained across embodiments and post-trained for specifics maps precisely onto the RoboJEPA thesis — and suggests the team is positioning AMI's approach as the technically correct answer to a question the VC community is framing incorrectly. Tweet
The RoboJEPA scaling-law claim is the most exposed bet AMI has made publicly — independent replication attempts will likely follow within weeks, and the community's verdict will shape how seriously the multiembodiment world-model framing is taken. Watch also for whether the AD-E2E-JEPA autonomous driving work attracts industry collaboration, and whether the VP of Finance re-titling precedes any announced financing round.