Evals on actual research workflows is the right direction, most agent benches are toy tasks.

I'm glad to see that GPT Sol beats Opus at least in Mathematical Sciences, because that's my need right now, and I much prefer GPT's prose style.

Damn. These things aren't AGI... but I don't care.

Luna is good enough for me to give a parser spec and have it write one.