I generally agree with the idea that AI raises the floor but not the ceiling on any domain of expertise.

*But* I think what this misses is that all the spare time and mental bandwidth I claw back as a domain expert gives me a lot of cycles to think deeper.

I get to imagine what can be, unburdened by what has been.

Known for the everywhere.computer (/ec), /unglish (unglish.co.uk), the /secure-design WG, WalletUncon, Causal Islands + drumming in all-genre power trios.

Sat in on a talk last night with the Google Gemini and Weights & Biases folks. Was especially interesting as I’m looking to identify ways to expand CI/CD to accommodate AI builders and their use cases.

Taking models and applying them to “fun” use cases still seems to be the norm with the tech so far, but builders are loath to apply them to actual critical use cases because evals are difficult and laborious to implement at scale and non-deterministic output is too risky to allow us to take the big risks. There’s the ceiling.

Yeah I wouldn’t trust these things unsupervised yet. Rubber ducking and batch processing of tedious shit is the strength.

But they are also pretty amazing learning aids. Synthesis across domains is gonna be big, I think.