J-lens Live readout of a 35B model's global-workspace band, per token,
as it writes — Jacobian lenses over the residual stream.
Schema
· code ARC-AGI-3 world-model harness on a local model: live run replay,
and the reproduction it runs on.
mlx-rl
· code GRPO reinforcement learning of LoRA adapters on Apple Silicon —
duel the base 35B against its RL-trained calibrated-honesty self,
live on the machine that trained it.
Podcasts Three machine-made shows — nightly paper readouts, Zvi digests,
and generated music. Listen in the browser or
subscribe by RSS .
GitHub Everything public.
Training, evaluation, and serving happen on one machine,
which also serves this page.
