Supervise the Patches You Did Not Hide
V-JEPA understood video but its feature maps were noise. One extra loss term on the visible tokens turns a world model into a dense one. With the algorithm and the full ablation.
Open → Self-supervised vision · CVPR 2023Predict the Meaning, Not the Pixels
I-JEPA hides most of a photograph and predicts what the hidden parts would mean — no cropping, no colour jitter, no decoder. With the algorithm and every ablation.
Open → Policy optimisation · The default algorithmProximal Policy Optimization
Take the biggest step you can without ruining the policy you already have. One clipped ratio replaces a second-order trust region — with the clipping surface, GAE and the full surrogate ablation, all playable.
Open → Policy gradients · Convergence theoryOne Episode at a Time
Everyone runs REINFORCE with a single trajectory per update, but every proof needed the batch size to grow without bound. This closes the gap — with every symbol explained and drawn.
Open → Sparse sensing · Kriging + learned residualKriging-Conditioned Reconstruction
Place a few sensors by greedy variance reduction, rebuild the flow with kriging, and let a network learn only what kriging misses.
Open → Multi-agent RL · Quadrotor swarmsThe Composition Axis
One set of weights for any team size and split — flown on a physically grounded quadrotor twin, with the scripted controller as a training participant.
Open → World models · Reinforcement learningMastering Diverse Domains
DreamerV3 builds a model of its world and improves by imagining futures inside it. One configuration, 150+ tasks, and the first diamonds in Minecraft from scratch.
Open →