Training language models on one Mac, in the open.

We started Akeratos to do our own research on a Mac, not on rented data-centre GPUs. Akeratos Billet pairs a core of 162 million active parameters with a learned table of 6.0 billion. The table was trained resident in the unified memory of one MacBook Pro (M4 Max, 48 GB), with no offloading code. Billet’s twin trained on the same tokens with the table switched off, so the gap between the two measures the table’s effect.

Held-out bits per byte, lower is better: 0.8866 with the table, 0.9030 without it. One seed per arm, on one MacBook Pro with an M4 Max (40 GPU cores, 48 GB) and MLX 0.32.3.

Source:Akeratos’s result page on GitHub, 14 November 2026

Row-level security is the database rule that keeps one user’s rows from another’s. On 14 October 2026, UpGuard reported 16,326 Supabase databases exposing readable tables. Our agent harness runs row-level security tests against a real Postgres when Claude Code tries to end its turn, and sends it back to work if they fail. CI runs the same tests on every pull request.

Source:UpGuard, 14 October 2026