The number of supply chain compromises through Github Actions is quite fascinating and it's not going to disappear since people don't learn Btw, bitwarden-cli just got hit by that

As expected, DeepSeek V4 is mainly about multi-head latent attention + mHC + Muon. Kimi K3 might be all about hybrid KDA + attention residuals and other niceties

Free hardening flags for Chromium: - strict-site-isolation (enabled) - origin-keyed-processes-by-default (enabled) - partition-alloc-with-advanced-checks (all processes) Specifically for Brave: - brave-v8-jitless-mode (enabled) - brave-extensions-manifest-v2 (disabled)

I guess I was right about that one :)

I think the main issue with continual learning is not opposing weights-based with harness-based memory, but rather what should get consolidated and when/where. Evolution may have taught us that a reliable memory mechanism is intrinsically hierarchical

Hitachi Seaside Park

Prompt injection is to AI agents what memory unsafety was to systems software