[iamflimflam1]

It’s pretty interesting how the work they’ve been doing on plugins has fed into this.

I suspect that they’ve managed to get a lot of good training data by calling the APIs provided by plugins and detecting when it’s gone wrong from bad request responses.