JEPA-style world models
Representation-first models that predict in latent space, ignore irrelevant detail, and learn useful abstractions from video and sensor streams.
Open research lab
The next useful models may be smaller, stranger, and more efficient.
ReplayAI is building models outside academia with selected students and independent researchers, using GPU access, implementation help, and open-source experiments that can be read, reproduced, and argued with.
Self-driving and embodied systems
Small planning, perception, and world-state models for agents that need to understand motion, scene structure, and consequences.
Video and vision
Efficient visual models for temporal understanding, long context, compression, and multimodal perception.
New graph models
Graph-based architectures for structure, relationships, spatial reasoning, memory, and systems where tokens are the wrong primitive.
The lab is for careful wheel-reinvention.
We are interested in people who want to rebuild basic ideas from first principles, test them honestly, and publish what actually happened.
apply / propose a track