Alternatively, what you’re seeing is continuous hill climbing and optimisations based on large scale experiments and simulated rollouts.

For me personally, I don’t use Claude Code because I can’t read the source.

At least with Codex, I know there’s a bunch of telemetry and all that, but I can have my agent inspect the source and have relatively greater confidence it’s not implanting spyware onto my machine, because a classifier thinks I’m working for Deepseek or something.

They're hill climbing by changing their system prompt... That's not how thats done.

They're changing system prompts to control how users use their llms.