Probing in production, and where this leaves things

Last piece of the probing block: how any of this gets used in production as a lightweight real-time monitor rather than a one-off research artifact — cheap linear probes running alongside inference, flagging when an internal direction lights up, versus expensive full evals. Trained a probe on Gemma and clocked it against a full forward pass; overhead is close to nothing, which is the whole point.

That closes out everything before steering. Looking back, the throughline for this stretch has been the shift from behavioral claims to actual causal ones — patching, path patching, causal abstraction, self-repair — that’s the part that didn’t exist anywhere in my prior ML background and is clearly the actual core skill here.

Honest overall note before moving on: a lot of what “worked” in this stretch worked on IOI-style toy tasks and repeated-token diagnostics specifically because they’re clean and legible enough to make a circuit-finding story work. That’s probably the right way to learn the tools, but I’d be surprised if the same techniques come apart less cleanly on messier, more realistic capabilities — and most of what I’ve read so far doesn’t dwell on the cases where the method was tried and didn’t produce a clean story. Selection effect worth remembering going into steering, not a reason to distrust the tools themselves.