Яо Шуньюй в Tencent: почему главная звезда AI-исследований начала с основ

Imagine the world's best striker joins a new club and in the very first match, instead of simply scoring goals, starts rewriting the tactical playbook. That's roughly what Yao Shunyu's first scientific paper under Tencent looks like. For those who missed it: Yao is a legendary figure in narrow research circles.

He's the one who gave us ReAct (Reason + Act) and Tree of Thoughts, teaching neural networks not just to produce text, but to plan their actions and reason. His recent move from Princeton to Tencent's research division caused a bombshell effect in the industry. And now we're seeing the first fruits of this union.

Instead of immediately announcing a GPT-5 killer, Yao and his team decided to look under the hood of one of the most mysterious phenomena in the world of large language models — In-Context Learning (ICL). If you've ever given a neural network a couple of examples in a prompt and it magically understood the task, you've encountered ICL. It's the model's ability to learn without changing its weights, right in the process of dialogue.

Until now, this process resembled alchemy: we knew it worked, but didn't quite understand how the model structured this knowledge internally. Yao's work attempts to turn this alchemy into rigorous science. Why did Tencent throw its best minds at this direction now?

The answer lies in the current state of the market. The era when you could simply bury a model in data and graphics cards is coming to an end. We're entering the second half of the game, where victory will go to whoever makes models smarter and more efficient at the architecture level.

In-Context Learning is the key to creating truly autonomous agents. If the model better understands context, it makes fewer mistakes in complex chains of reasoning and requires fewer computational resources to adapt to specific business needs. Tencent is clearly betting on the quality of reasoning, not the quantity of parameters.

In their work, researchers analyze how exactly input data distributions influence the model's ability to generalize. This isn't just theoretical inquiry. Understanding ICL mechanics allows for creating more stable systems that don't break from a single poorly chosen word in an instruction.

In the face of fierce competition with Alibaba and DeepSeek, it's vital for Tencent to have a technological advantage that can't simply be copied by buying another ten thousand H100 chips. They need algorithmic breakthroughs, and Yao Shunyu is exactly the person who can deliver them. It's also interesting how this work fits into the broader strategy of Chinese tech giants.

We see a clear shift from copying Western architectures to attempting to lead fundamental research. If before Chinese companies often played the role of followers, now they're hiring top scientists directly from the benches of leading American universities and giving them a blank check for deep research. This creates a new dynamic in the industry, where the boundaries between academic science and corporate development are finally blurred.

The bottom line: Tencent is investing in the foundation, not the facade. If Yao Shunyu unravels the mechanics of In-Context Learning, the next iteration of their models could leave competitors far behind through efficiency, not brute force. Can understanding ICL become the very lever that will overturn our understanding of the possibilities of current architectures?