The Dwarkesh Reference
← Back
Claim≈ Approx

KataGo (2020, David Wu) achieved roughly a 40x reduction in compute to train a strong Go bot tabula rasa versus prior state of the art.

Who
Eric Jang
Topic
KataGo
Verification note
KataGo made major efficiency gains over AlphaGo Zero; the exact 40x is from the paper and plausible but not independently checked here.