Models, research, and the decisions behind better AI systems.
Latest update
Amazon Marketplace froze the foundation model, fused holistic and readability signals, and trained a cheap calibrator. Anthropic's 207,049 sessions and GPT-6 Cyber show why task and scope diagnostics still matter.
A Linux-kernel audit shows why terminal success is not proof of the assigned path, while a 2B document extractor and Kimi's open-weight release sharpen task-specific deployment decisions.
Google's forced model substitutions make compatibility testing urgent, while new browser and harness evidence shows why the orchestration layer belongs in every model release.
Pando narrows coordination to one Git ref, stores immutable pack dependencies, and moves cross-ref snapshots into the application contract. The design is a useful case study in matching consistency scope to the invariant that actually matters.
A malicious package in Cursor's marketplace shows why extension identity, update provenance, endpoint evidence, and served-model behavior all belong in the production AI control plane.
Just-in-Time tests turn a code diff into an executable counterexample, but the production design still needs a separate oracle, review economics, and a policy boundary the generator cannot rewrite.