Codex is super weird for me. When using their CLI it sometimes seems brilliant but often it misses stuff. They need to tune their exploration before coding. The Codex CLI lags behind in UX from Claude Code but is catching up fast. One thing to watch is "Skills": Codex doesn't have an equivalent yet

What's funny is that the responses to you are mostly slop - lazy, unoriginal, mistaken arguments. People have become what they hate about AI.

That made me curious:

You just made my night! And ruined my next-next-week 😂

This one is wild; makes it seem mundane. Like a throwaway snap you'd take out of the window of an airplane. But there's Earth - wow!

(Or distill into itself this way)

Thinking this will be a powerful new formula: collect use cases, build a course router, prompt optimize (GEPA etc.) each use case (including many-shot examples sampled from prod), then distill that whole system into a smaller model.

Just played for a few hours so taking a break, but will send you a steam link so we can connect there later!

...why deep learning gets better with scale (better fractal "render" resolution), why LLMs have to do all kinds of norm and projections to work (must attend to many feature scales), and why TRM works well on ARC (can find features in hugely magnified areas of the intelligence manifold)