4 posts
llm A public benchmark showed 92% of coding-agent tokens were served from cache — a 12.5x cost difference. A teardown of oMLX's prefix cache, the six gaps in it, and the three ceilings that decide how many agents one machine can serve.
mlx What MLX, mlx-lm and oMLX each own, which layer oMLX actually replaces, and three findings from reading the code: a dead SLO subsystem, an auth model with no principal, and a Thunderbolt 4 misclassification.
mlx A PR that shipped 56,190 lines by accident, a security regression that rejected every real SSH key, a 2 GB venv one flag away from deletion, and a cluster that flapped 297 times because a bundled Python could not import its own package. Every number here came off the machines.
mlx MLX issue #3910: when a peer dies mid-collective, JACCL spins forever with zero errno, zero exception, zero log output. I read the source. It is six unbounded loops, not one, plus a second unreported bug with a negative-offset pointer. Here is what is actually broken and why Rust would not have save