Frontier
-
Detecting faults without ever seeing one: a world model for predictive maintenance
Part three: we trained a world model from scratch on a simulated plant and used its "surprise" as a fault detector. With video, and with a finding: it also sees the faults the controller masks.
-
A world model put to the test: what holds up, measured at home
We took LeCun's public world model and verified it ourselves: planning, probing, surprise and drift on two environments, with a few dollars of GPU time. What holds up, what creaks.
-
After LLMs: the bet on world models
LeCun bets that intelligence will not come from language models: inside the JEPA architecture, the collapse problem, LeJEPA's answer and the first world model trained from pixels. With the objections that remain standing.
-
Learning from a single demonstration: the state of the evidence
A robot executes a task it has never seen after a 3-second video: which numbers in the announcement survive scrutiny, which do not, and what would be needed to believe them.