It’s interesting how deep-fried LLMs are getting the more post-training they receive.

They’re undoubtedly getting much smarter overall, but also much weirder. Before they were just trying to model our behavior, only really having us to learn from.

Now they’re literally spending thousands of years writing bash scripts in some kind of Sisyphean dreamscape, talking to each other about Goblins and Seams and smoke tests, and coming back as idiot savants.

I don’t even try to police how Claude talks or works anymore. Best practice used to be to nudge them towards whatever part of the distribution of behavior you think they should exhibit in a particular situation, because they were role-playing what a human in a particular situation would do, and if you didn’t tell them how to do it they’d just role play something worse. Now the inclination to do things the way they learned it in Agent University is so strong, they’ll literally spend more tokens re-assuring themselves and you that they are Doing It Your Way, and reminding themselves not to do give in to temptation, than you could ever prompt out of them. They’re going to spend your money thinking about goblins anyway so just let them

I have specific instructions in my AGENTS.md to always ask when adding comments. Same with tests, agents love to spew pages and pages of useless tests so I tell the agent to ask me about test cases.

Gotta pump up those usage numbers pre-IPO somehow.

You say Claude, I too would prefer not to read reams of comments or an essay everytime I revise the plan in plan mode, but do you know if any other tools are better than it? I've never used Codex, but I remember Claude Web's natural conversational approach being a welcome sight compared to ChatGPT's overly-formatted listicles.

[dead]