RESEARCH DIRECTION
Humans and animals acquire common sense from observation, not annotation. We study self-supervised objectives — predictive and joint-embedding formulations — that let systems learn multi-level representations of the world directly from unlabeled streams, predicting in representation space rather than pixel space, at multiple time horizons simultaneously.
Pixel-perfect prediction wastes capacity on irrelevant detail — the flutter of every leaf. Joint-embedding predictive architectures sidestep this by predicting abstract representations of future states rather than the states themselves, with energy-based formulations handling the irreducible uncertainty. The model learns what is predictable and represents the rest as latent variables, not noise to be painted.
Reasoning at one timescale is a pathology. We build representation stacks in which each level abstracts and predicts over a longer horizon than the one below — millisecond dynamics at the bottom, task and goal structure at the top — so that planning can decompose long-horizon problems into short-horizon subgoals natively.
Supervised objectives remain useful for evaluation and steering, but a system whose representations depend on labeled data inherits the coverage limits of its annotators. Our foundations are trained on prediction; supervision is applied sparingly, at the top, where it is cheapest and most meaningful.
WORKING PRINCIPLES
CONTINUE EXPLORING