Ex-xAI/Gemini engineer on steering MiMo-V2.6: no RL run is launch-and-forget
infoxiao · x · 2026-09-19
An engineer who worked at xAI and Gemini describes captaining a MiMo-V2.6 RL run: training a checkpoint that actually ships means no run is ever launch-and-forget—it's about how many infra and recipe issues you discover and patch while the run is ongoing. Hard failures are easiest; if you don't think there's an issue, you're not looking closely enough. He says the experience closely mirrors his time at xAI and Gemini.
Related event: MiMo RL Training Livestream: Infra Restarts Cost Over $400K(2 posts)→
More from Companies & People
- Anthropic quietly built a Bay Area biolab where Claude will direct robot experiments — AIFlow_ML · 2026-09-19
- Stanford's Manning: bureaucracy has made universities worse for research — stanfordnlp · 2026-09-19
- Daniel Rock assembles all-star team of economists to tackle frontier AI economics — soumitrashukla9 · 2026-09-19
- calvinnwq joins NousResearch's Hermes Cloud team, betting on open-source AI agents — max_paperclips · 2026-09-19
- Sophont founder: a wave of startups is now building frontier medical AI labs — iScienceLuvr · 2026-09-19
- LangChain CEO: jev sparked more internal product demos than any other model launch — hwchase17 · 2026-09-19