Musk: Grok 4.8 pretraining done, Optimus to show unscripted capabilities within 1-2 months
FuSheng_0306 · x · 2026-11-19
Highlights from Musk's TomorrowX Summit talk (items 4–9 of a 24-point thread):
- Anti-benchmark stance: chasing leaderboards is the wrong optimization; what matters is whether models actually solve users' problems
- Grok training method: a Tesla FSD-style loop — every error triggers a hunt for wrong/biased/missing data, then retraining
- Grok Bot as feedback system: new models ship small to Grok Bot first and iterate on real users' task completion instead of lab tests alone
- xAI infra rewrite: the team dropped parts of existing frameworks and rebuilt training, RL, and inference systems in C/C++ for performance
- Optimus usefulness bar: Musk borrows John Carmack's test — unbox a RC car, open the battery cover, insert batteries, close it, and drive; that's the everyday-task bar Optimus must clear
- Demo preview: Optimus may show new capabilities in 1–2 months, emphasizing unscripted, near-daily-use abilities rather than staged demos
(Earlier items covered Starship's full rapid reusability, multi-planetary civilization, and that Grok 4.8 has finished pretraining and is in RL.)
More from Companies & People
- Anthropic fans and Freeze AI protesters outside Nasdaq — one came to ask Dario for a job — shiringhaffary · 2026-11-19
- Every staffer built an AI coach to hit level 8 on its AI adoption framework — every · 2026-11-19
- Granola cofounder on the "scorpion dance" with frontier AI labs — soleio · 2026-11-19
- Anthropic eyes investment-grade rating; GaryMarcus mocks the idea — GaryMarcus · 2026-11-19
- Musk says xAI is adding 0.5GW of compute per month, Neuralink implants pass 30 people — FuSheng_0306 · 2026-11-19
- Google names Gemini for its partner-run sovereign cloud Dedicated, S3NS opens AI hosting in France — shashib · 2026-11-19