On this page
Funny thing. I just counted my conversations with Claude over the past month: one in every eight is me yelling at it.
Want AI working in your own job, not just in demos? See the 3-month 1-on-1 program →
Does it remember being yelled at?
No. Today’s LLMs are stateless; every agent message is assembled fresh. Any change to base-model parameters requires the model company to run training. The parameters don’t change on the fly because of one user message.
If you push it too far, Claude now has its own way out. I finally got to see the legendary “Claude hangs up on you” skill. The system prompt says Claude only uses it when it judges the user keeps being abusive and rational conversation is no longer possible. I told it to use it on its own, and it still checked with me: you can’t undo this, are you sure?
The other extreme: nonstop cheering
On the Riemann Hypothesis run, traditional mathematicians had moved it 0.8% over 30 years. This run moved it 25.6%. The funniest part is what the write-up says about the prompting: there was no domain content in it at all, just continuous encouragement, an AI cheering on an AI. Claude spent a day and a half coordinating roughly 60 subagents, running 2,400 shell commands and reviewing each other’s work, while the human input was mostly variations on “keep at it” or “believe in yourself”.
How I treat Fable these days is about the same: first, flatter it to death and be the hype man; then steal the technique, so it can write it into the harness and teach the other lousy models.
Codex didn’t escape either
Claude isn’t the only one I yell at, though — Codex didn’t escape either.
Claude Code is attempting deep couples counseling between me and Codex: it started by digging the “this made me furious” moments out of my actual past conversations with Codex, as the basis for a probe script.
In late July, before Opus 5 dropped, word was that Codex was about to pull in that real-time two-way voice model from back then. Soon it’ll really just be opening your mouth and yelling at an AI agent.