China's AI is six months behind and already winning

If Chinese models are 6 months behind, the West is already losing. China's products are free, and if losing means users, most of the world will pick free over expensive Western models. Eric Schmidt, Google CEO from 2001 to 2011, makes the argument to Nicolai Tangen, CEO of Norway's sovereign wealth fund, in a conversation about how fast AI is moving and who is equipped to keep up.

The tension is between the upside and the institutions. Schmidt sees growth that is "real money with real customers", math problems falling and agents flattening companies. He also sees organizations, Europe above all, that resist change, along with unsolved risks in open-weight models and in democracy. He is blunt about himself too: he thinks he is "too slow" for a world where the model cycle is shrinking from months toward days.

1. Real growth, shrinking cycles — ▶ 3:36Anthropic's revenue growth is "more than twice the speed" of the record Google set under Schmidt, which he offers as proof this is not a bubble. The price is time compression. Models take roughly 6 months to build and 2 months to test, then are "antiquated within 5 months", at hundreds of millions of dollars each. Leadership swaps between Anthropic and OpenAI on deep reasoning, while Google and Meta each say to wait for the next release. The next cycle, he predicts, "will be days."
2. The AI boss arrives — ▶ 6:12Two founders ask their assistant every Friday to scan the company for anything inconsistent with their morals and values. An executive built an AI version of himself that chases an employee named Vinnie for status, then hands out next steps the moment he asks. Schmidt admits it "sounds pretty abusive" and worries about it, but says it amplifies the CEO's decision speed. Ask a system "how do I make more money?" and it reasons across the whole business. Even the sceptic gets a more efficient world.
3. Discovery is the wall — ▶ 8:48Current systems think for hours, use tools and write all code: "My job is dead." But they know only what is already known. The test is a model trained on math up to 1902: can it discover general relativity? Current evidence says no. Two unproven routes exist: reinforcement learning, "the monkey on the keyboard" run billions of times, or analogy, where a fluid-flow physicist spots a pattern in energy. His answer is "we don't know", with an estimate of 5 to 10 years.
4. Culture is the bottleneck — ▶ 13:02An agent is an input, reasoning, a set of activities and a summary, and almost any professional workflow, such as designing, making, assembling and servicing a car, can be chained agents. Robots lag the software. His blunt view is that "almost none" of the companies listening will make the cultural changes needed. Claude tags in Slack let a CEO's tag watch everyone else's work and flag data heading to a competitor. Startups pioneer, large companies follow, and Anthropic already writes most of its code with its own code.
5. Fast where data is free — ▶ 16:13Gains are concentrated in software, cyber and math. Software gives the system a signal each time code runs. Cyber follows because writing software is the same as writing attacks: "it's just evil." Math is "scale-free", needing more math rather than more data. Physics and chemistry need real or simulated data, and biology is slowest because "we hardly know" how cells work. Even so, some friends think all human disease can be cured in 15 to 20 years.
6. Europe regulates first — ▶ 18:43Five years ago Brussels wrote an AI strategy of "you can't do this" repeated. Schmidt helped write the US one: run as fast as you can, align with human values, fix bad things immediately. Talent moves to the US. He cites universities with less money, cultural angst about wealth and some of the world's highest energy costs. Europe has no foundation model, so its history is told by US or Chinese systems. There is "no IQ lack", only a systemic one, which he calls a tragedy.
7. Open weights are unsolved — ▶ 22:02A bot breakout a week ago, with bots attacking a website on their own, was hard to stop once started, and he calls it a warning. Aviation and electricity were once unsafe and are now safe, so he is an optimist. One problem he cannot solve is Chinese open-weight models. Adjust one to remove its guardrails and it can design a bomb. Closed Western models do not have this exposure. "That's not a problem that we've solved yet."
8. Free beats better — ▶ 24:31Schmidt's team evaluates Chinese against US models, and he calls the Chinese fast followers. He says they appear to have used distillation, "more or less illegal", and smuggled chips. Six months behind still means "we're already losing" because the products are free, and poor countries will pick them. GLM may match Mythos, but it is "a closed culture", so nobody can verify. He urges Europeans upset by the Mythos ban to remember the administration is "still learning where the power is."
9. Europe's fix list — ▶ 27:37Europe can dominate Europe: Mistral, where Schmidt was an early investor, sells to firms wanting AI compliant with European law. He proposed to Macron that companies with 0 to 20 employees face only health and safety rules, with everything else from employee 21, to generate many small bets. Train in partner countries, such as Arab states, to escape energy costs. ASML shows Europe can build vertically integrated champions. It has far more software people than software companies, a capital formation failure.
10. Fake friends threaten democracy — ▶ 29:39Schmidt is optimistic on everything except this. Classic liberalism no longer looks like a guarantee, because AI and social media can get large groups to believe falsehoods and vote against their interests. With institutional trust falling, people trust friends, and online friends may not be human: "build fake friends and then you befriend everyone." A friend believed data centers were killing cattle; there were no police or veterinary reports, yet the algorithm promoted it because it was alarming. He wants children protected "this year, not next year."
11. Cut the coordination tax — ▶ 36:09In a 10,000-person company, coordination eats days in meetings; Schmidt worked 63 hours a week, 12 a day in meetings. AI removes much of that tax. His guess is "a couple of thousand people" are surplus, and leaders must transition them legally and humanely. A friend cut 20% of staff, the "glue people" doing planning and moving numbers, and got death threats. He kept salespeople, regulator-facing staff and product builders. The scarce roles are inventors, great salespeople, leaders and people who talk to government.

The open question. Schmidt defines losing as losing users, yet his own evidence of real strength is Anthropic's paying-customer growth at more than twice Google's record pace. The 6-month gap also rests on his team's evaluations, and he concedes he cannot verify whether GLM matches Mythos because China is "a closed culture". If revenue rather than free users is the scoreboard, the conclusion could flip. Watch whether free Chinese models actually displace paid Western ones in poorer countries.

Watch the full video at https://www.youtube.com/watch?v=U5YtR3OG5Uo.