Blog
Steering experiments to reduce sycophancy (linear probes on LLM activations)
TruthfulQA + epistemic-integrity probes, steering on MHA/MLP/residuals — citations, pipeline, and what to expect from the experiments.
Crafter rollout collector: a web app for human & agent data
Full-stack Crafter rollouts, policy demos in the browser, and world-model imagination side-by-side with real frames — motivation, stack, and literature.
Correcting overestimation bias with SIGReg + InfoNCE
Main takeaway from JaxGCRL experiments: pairing Sketched Isotropic Gaussian Regularization with InfoNCE reins in optimistic contrastive signal.
Scaling deep goal-conditioned RL with SIGReg-ISO
Applying Sketched Isotropic Gaussian Regularization to the actor trunk prevents representation collapse at depth, improving humanoid evaluation success in goal-conditioned RL.
LeDEEP: monocular depth estimation with LeJEPA and SIGReg in production
ViT-Small encoder trained with LeJEPA multi-view self-supervised learning and SIGReg regularization on 1280×720 drone footage — served via a REST API with sub-second inference on arbitrary images.