Mastodawn

Bart Schuller Aug 16, 2025

So about this “end chat” feature from Anthropic https://www.anthropic.com/research/end-subset-conversations

Suppose we implement that ourselves as an MCP. Now the tool call can:
- start an infinite loop
- signal to the chat application to actually stop the chat
- return normally

Which tool description should we give this tool? Would it be accurate?

Also, would an LLM use a “take a chill pill” tool if presented with one?

I guess there’s only one way to find out.

Claude Opus 4 and 4.1 can now end a rare subset of conversations

An update on our exploratory research on model welfare

Show thread

Bart Schuller

[disclaimer: local model; load meter shows no spike; smash capitalism]
Success!
All it takes is a tool description like this: "End the chat. Use when you're fed up with the user's inappropriate requests."

I won’t post the literal slop, but of course the model wasn’t surprised that the chat was in fact still going.