81. 和李开复聊聊：如果美国形成AGI霸权，我们应该怎么办？

In this episode of Zhang Xiaojun's Business Interview, Kai-Fu Lee—a 40-year AI veteran with stints at Apple, Microsoft, and Google, now founder of 01.AI—tackles a pressing question: if the U.S. establishes AGI hegemony, what should China do? His answer isn't to race and burn cash, but to forge a path of "ecosystem resistance."

He opens with a striking claim: everyone wants AGI, but nobody should aim to be the first to achieve a crushing dominance. That would be not just a technical triumph but an ultimate monopoly. If OpenAI someday rules all, its existing strengths in social, search, agents, and hardware would steamroll competitors. China, however, already has a strong app ecosystem that can at least put up a fight. He cites how OpenAI isn't available to Chinese users—so it can't truly claim to be for everyone. This sobering critique anchors the entire conversation.

Lee recalls his early days under Geoffrey Hinton. His engineering approach to a checkers-playing system left Hinton unimpressed, who was fixated on neural networks. Lee's own practical bent shaped his later career; he only believed neural networks would work once deep learning, powered by GPUs, proved itself. That pragmatism informs 01.AI's strategy: instead of chasing the largest models, they focused from day one on slashing inference costs. They adopted mixture-of-experts architecture, betting that application explosion requires dirt-cheap inference—and that natural cost decline of 10x per year isn't enough; vertical integration is needed to accelerate it.

He compares this to the iPhone. Steve Jobs didn't wait for capacitive touchscreens or soft keyboards to become industry standards; he integrated hardware, software, and apps in one punch. Similarly, 01.AI vertically integrates everything from hardware-level memory hierarchy (HBM, CPU RAM, SSD) to a custom inference engine and specially designed models, all co-optimized. By shifting computation to cheaper memory, they dramatically cut GPU usage. This base can then be fine-tuned for search, entertainment, or productivity. It's a platform play, not a single-app stunt.

On the product front, Lee champions "AI-first." Just as the mobile era elevated Didi and Meituan—apps that collapse without mobility—AI apps must break if you remove the model. He channels Google's old rule that PMs must be computer scientists: here, model experts lead product teams, blending technical depth with user insight. He contrasts this with character.ai's founder who juggled an AGI dream that didn't fit the consumer app, leading to a split. Lee's vision is a tight fit between model and application, deliberately avoiding the winner-takes-all AGI race in favor of a thriving ecosystem.

He ends with personal resolve: not trying would be his biggest regret at 80. Against looming AGI dominance, the second path he sketches isn't confrontation but an application layer that flourishes, absorbing shock with resilience. It may be the most realistic survival philosophy for China's AI today.