I think this is more akin to an RL'd LLM calling tools (image creation tool), passing results back into the context window, manipulating outputs with further tool calls (rotations, color changes) The brain is not a transformer, but the context window is like conscious attention.

The linked slideshow does a really nice job breaking down how the big labs finetune RL'd models

NYC, SF, DC, and Boston feel like places where you can walk out of your house, get around, be part of civic life, all without a car. I've lived in 3 out of these 4, it's painful to go back once you've gotten used to it.

Bill Belichick apparently met his 24yo girlfriend while discussing Warren Goldfarb's Deductive Logic?!

Sonnet 3.5 was AGI

Artifacts -> Canvas Projects -> Projects Claude Code -> Codex All the best 'product company' products first come from the dev tools and enterprise api company

I genuinely think Sonnet 3.6 was AGI for me.

UCSF construction, at Parnassus and 3rd SF Builds!

Something about the LLM api market that I didn't expect: switching costs are high. It's not a about 'how much intelligence' you can get for a cost, but the RL'd models succeed and fail in unique ways, you need to prompt around failure modes, and so prompting needs to be updated

I liked Sonnet 4,5's answer to what makes good philosophy: