Listen: 2037. AGI. China probably.

The most interesting story in artificial intelligence right now is not about what we are adding, but what we are choosing to leave out. In a leaked investor meeting, the founder of the Chinese AI firm DeepSeek, Liang Wenfeng, laid out a strategy built on saying no. No to video generation, no to massive super-apps, and no to what the industry calls world models. Instead, his strategy is deliberate restraint. To reach artificial general intelligence, or AGI, he argues we must focus on what people actually want: models that can function as real labor. Today's systems start every task from scratch. To become useful, they need continual learning, which is the ability to pick up skills on the job and keep them. This sets up a major clash. Many Western labs are betting billions on heavy world models that try to simulate physical reality. But evolution tells a different story. Biological brains are light, efficient, and always on. They don't simulate the entire universe; they use cheap, quick reflexes. Combining lightweight models with continuous learning is the real path forward. Meanwhile, the Chinese open-source ecosystem is quietly dominating, winning on cost discipline and hardware independence. Everyone knows how to add features. But almost no one with billions of dollars has the discipline to say no to the hype. In the race to AGI, the winner might just be the one who travels lightest.