honestly i dont even care that dsv4-pro was weaker than expected.

this paper and the new deepseek harness makes me feel that deepseek is still on track to their goals.

if they crack continual learning first, all other shortcomings and compute poverty are forgiven.