Don't be the tooling, build the tooling: [un]prompted.au talk

I gave a 15 minute talk at [un]prompted in Sydney last week, and the slides are below.

It's the worked-example sequel to a short post from April, using lookup.disclose.io as the example. I got pissed off about rookie bug hunters dropping 0-day on X because they couldn't find the right contact, and figured AI might have some solutions for the problem.

The ISA.md for lookup says "never hallucinate" and "determinism only (!!!)", which pretty much rules out asking a model "wtf is this thing?" on every request. What I did instead was use the best available models to answer the question over and over, reverse-engineer the trajectories that worked, then task a model with implementing those in deterministic code, with a couple of loops running across the production logs to keep improving it.

54,114 lookups so far.

The bit I'd underline is that product management > engineering for problems like this one... most of the work was defining the ideal state and pointing the agents back at it.

If the viewer isn't playing nice on your phone, grab the PDF here. Lookup is free to use if you want to have a go.