Kimi K3 take: elite post-training can leapfrog stronger base models
JacquesThibs · x · 2026-11-20
Commenting on Kimi K3, Jacques Thiboust argues there's much to extract from pretraining via great post-training, and distilling traces from stronger models lets you hit the ceiling faster — so exceptional post-training can leapfrog stronger base models with weaker post-training. He guesses Google is more careful about post-training on OpenAI/Anthropic traces, one reason it currently trails Chinese labs, which he also says pay less attention to safety. He adds this was his covert way of saying Google would come back "from the dead".
More from Companies & People
- Survey of 200 CIOs: 11 of 12 scaled enterprise AI operators at break-even or better — BenBajarin · 2026-11-20
- Palantir CEO Karp: Dario and Altman "don't understand how unlikable they are" — DrDatta_AIIMS · 2026-11-20
- Abridge healthcare AI hackathon returns with OpenAI and a16z — GaryMarcus · 2026-11-20
- Johns Hopkins DSAI Institute opens postdoc fellowship applications for 2027–2028 — mdredze · 2026-11-20
- Google Cloud Next 27 set for April 13-15 in Las Vegas, registration opens December — rseroter · 2026-11-20
- GDM DevX lead osanseviero starts sabbatical after 100+ launches incl. Gemma 4 — osanseviero · 2026-11-20