Somewhere right now, a counter is ticking. With every tick, a company loses another dollar to a machine that hasn’t even finished learning. This isn’t a crypto scam. It’s Xiaomi’s live-streamed AI training session.
In the early hours of the morning, Xiaomi pulled a move that made everyone look twice: it put the training of its next-generation models on live TV. Two training pipelines at once. Tokens climbing. Benchmarks climbing. Failures flashing on screen. And the number everyone couldn’t stop staring at—the cost counter, past a million dollars and still climbing.
But the real story isn’t the money. It’s the woman holding the matches.
Meet Luo Fuli. She used to help build DeepSeek’s models. Now she’s at Xiaomi, in charge of MiMo. In other words, a DeepSeek alum is spending Lei Jun’s fortune to build a model that may one day go head-to-head with her alma mater. The workplace drama writes itself.
The most dangerous person in AI isn’t the one publishing breakthrough papers. It’s the former employee who left with the playbook and now has a bigger budget.
DeepSeek has quietly become the AI world’s Whampoa Military Academy. It doesn’t just build its own models—it graduates people who go on to build models for the competition. And Luo Fuli is its most visible graduate yet.
For months, she went quiet. No flashy announcements. No fake breakthroughs. Then she reappeared with a live stream: “Stop guessing. Come watch us work.”
That’s when Lei Jun’s money finally got a picture. And it looks like a furnace.
Previously, AI companies showed you the graduation ceremony. Xiaomi is showing you the homework. You see what tasks were attempted, which machines crashed, how many times the token counter blinked. If you’re a CFO, do not open that live stream.
But here’s the twist nobody saw coming: Luo Fuli isn’t training a chatty little sweetheart. She’s training an agent—a worker.
The new MiMo models write code, open terminals, call tools, and run tests. They fail. They correct themselves. They try different paths for the same task. This isn’t “answer my question nicely.” This is “here’s your computer, here’s the requirement, get the job done.”
We’ve spent years teaching AI to talk. The next decade will be about teaching it to work.
And that changes everything. Conversational AI is table stakes. The real war is over whose AI can actually execute—close a ticket, ship a feature, run a business process without a human babysitter.
So why should you care? Because the next computing platform belongs to whoever can turn intelligence into action. Xiaomi has the hardware, the users, the capital, and now, a DeepSeek-trained general leading the charge. That combination is dangerous.
Luo Fuli’s “homework” may or may not make DeepSeek lift its head. But one thing is certain: Lei Jun’s 60-billion-yuan promise just gained a heartbeat.
Drip. Training fee deducted.
Still deducting.
FAQ
Q: Is this live stream just a marketing stunt?
A: Partly. But it's also a recruitment ad, a transparency play, and a warning shot. Xiaomi is telling the world: We have the money, the talent, and the guts to train in public. Even if the model fails, the story works.
Q: Does this mean DeepSeek is in real trouble?
A: Not overnight. DeepSeek still has strong models and research momentum. But if its top talent keeps leaving to richer rivals who can outspend it in the agentic AI phase, its edge will fade. Talent flight is a slow bleed—until it isn't.
Q: Why is agentic AI such a big deal?
A: Conversational AI is table stakes. The real gold is AI that can execute tasks: write code, run tests, manage workflows, fix errors. That requires deeper integration, more compute, and larger budgets. It's an execution war—and Xiaomi just showed up with a checkbook and a chip on its shoulder.