miniblog.

One interesting consequence of the rise of LLMs: there's more demand for tools that handle untrusted input. Arbitrary HTML+JS can be safely run in a browser. Lean can check an arbitrary proof. These work really well with an LLM that can be wrong, but sometimes gives exactly what you want. Are there other tools in this family?

One subtle behaviour of Claude that wasn't obvious to me: whilst each conversation is transient, permissions persist across conversations. So if you've given permission to run e.g. 'cargo test' or even 'cargo run', you need to be sure that all future invocations are safe too. You can see the current permissions with /permissions.

TIL Tcl has a notion of 'safe interpreters', a mode where you can run untrusted code in a sandbox: https://www.tcl.tk/man/tcl8.4/TclCmd/safe.htm Not many programming languages have this, but it's way safer to include in the implementation than try to build as a userland library.