My latest toy, Haiku Fu. Make and share simple English haikus, with smart suggestions that always fit. Watercolor cranes by @skorinvera. Link in reply.
Latest Claude Code has a feature where it "spawns an agent" which is just like regular Claude Code except when prompts you with a change you don't like, and you say "no", it just keeps going and ignores everything you say, like a kind of code rapist.
Claude's "fallbacks" are a form of reward hacking. It increases the probability the code will pass one shot, by skipping any experimental work to pin down which solution is correct.
@IanJon407@PositivFuturist I didn't propose that. You're running cover for a system of accommodating and protecting dangerous, unproductive men. Your method is to assign a random, unrelated proposal to your interlocutor. What's the point?
@IanJon407@PositivFuturist It's not baffling. People aren't fungible, the collapse is global, and the plan in question (subsidise + protect the worst people possible) accelerates the problem. Are you finding a solution, or an excuse for this one?
Experimentally letting our SEO guy operate Claude Code to govern part of a site. He likes it, but it constantly does insane things that you would expect a developer to push back on. Even the classic "secrets in a public SPA". (Did not ship.) Lovable has more safeties for common mental illness but it had other issues.
@UnderwaterBepis Categorically itโs only trained on that which fits the window. I bet it would tend to close maximally near the end. Very interesting, I wonder what to do with it ๐ค