Inference is cheap. Inference is costly. Nobody tells you that inference is loud and toasty. You learn that when you run a 2 x RTX Pro 6000 build in your office.
@skydotcs This non deterministic contraption did a non deterministic thing. I lost some things because I didn't do the deterministic thing I should have. Now I am angry.
Love Remodex by @emanueledpt for working with Codex (release the app @ajambrosino , @thsottiaux ) on the go.
Made it more personal: libghostty terminal ( @mitchellh ๐) , @pierrecomputer Diffs review, file/change accept-reject, updated some project sidebar controls, native macOS app/plugin @mentions, faster bridge recovery.
Lesson from automating Sentry fixes: donโt ask an agent to โfix the issue.โ
Ask it to prove the issue first: hypothesis, evidence, narrow patch, tests, browser verification, and PR metadata linking back to Sentry.
Automation gets safer when it adds the right friction.
A lot of โAI agentsโ are just weaker, prettier versions of `grep`, `cat`, `ls`, and `find`.
If a model can't inspect reality, it's mostly doing improv.
@thsottiaux Give me deep research in Codex. Also give me the ability to ask a question about something in the output without actually continuing the conversation.
@garrytan Built a system in the early days with Squid proxy and an explicit allowlist of domains itโs allowed to reach to. Get approval request for new domains. Working like a charm.