TP

If you release a paper, please release a podcast too.

If we recorded model evaluations like we record git commits, I'd look like a 10x engineer.

Nous world_sim

The gap between researchers and labelers should be as small as possible

Tulu3 is such a gift

Do we have any reason to fear vocab collapse? Maybe we use too many words?

I expect that highly controlled corporate environments are a great place to experiment with agents (simulated first). Would love some ideas from the community on experiment design. Related GitHub: github.com/socragpt/Age...