The model is the stale one
2026.162I pointed two frontier models at my own docs to see which one to pay for. Both flagged the correct, freshly-updated pages as stale -- the audit was poisoned by the thing being audited.
Ai
tag:
ai
all tags →2026
The mechanism is the enforcement
2026.158Three AI code-review tools sit on the same model. The one that guarantees coverage did it with a for-loop, not a better prompt.
The mask in the README
2026.139An attention mask drawn in ASCII, and the architectural trade it names.
Reviewing code that looks right
2026.133The old failure mode was reading every line equally. The new one is trusting code that looks right.
The archive reads back
2026.111An LLM in the middle of a fetch pipeline took orders from the file it was summarizing.
The easy half
2026.105A wallet lets you spend. That's the easy half.
Trust boundaries
2026.079Your code already leaves the machine. The agent just gave it another surface.
What compounds
2026.079Credentials got easier to claim. Artifacts didn't.
The loop
2026.073The deal was never about the license. It was about the loop.
The detection trap
2026.067When quality looks suspicious, people stop producing quality.
Repeat yourself
2026.049Retries fix distributed systems. Repeating yourself fixes attention asymmetry.
Coding with LLMs
2026.013The bugs are quieter now. Just wrong assumptions buried in correct-looking code.