Yao Shunyu: Let Me Go a Little Crazy! Training Models at Anthropic & Gemini, Heroism Is Over

In this candid conversation, Google DeepMind researcher Shunyu Yao traces his unusual path from condensed matter physics at Tsinghua and theoretical high-energy physics at Stanford to AI. Unlike his near-namesake, Tencent’s chief AI scientist, Yao is a latecomer to the field, which gives him a distinctive perspective: he focuses not on hype but on how problems are defined.

According to Yao, the top three AI labs—OpenAI, Anthropic, and Google—have stopped worrying about catching up with one another. On public benchmarks like SWE-bench, the scores have converged around 80%, with differences of just a point or two now amounting to noise. Yet users still feel real distinctions: Claude excels at tool use and agentic tasks, Gemini at general reasoning and conversational fluidity. In the past, companies could pour resources into a chosen direction and produce clear gaps; today, the harder challenge is defining precisely what behavior you want and curating the right data. Yao insists that pretraining is far from plateauing—models are learning more efficiently than ever—but the critical bottleneck has shifted from raw capability to problem specification.

He is unfazed by the recent hype around OpenClaw and similar tools. Internally at large labs, such agent demos already existed, he notes, but big companies can’t release a product that might accidentally take over a user’s computer. Individuals can throw up a rough open-source repo and spark a realization: models can orchestrate long-horizon tasks across multiple tools. The real value of these projects, he says, was consciousness-raising, not a technical leap.

For startups building “shell” products on top of foundation models, survival requires either blistering growth that outruns the platform, like Cursor did by gaining massive developer mindshare before Anthropic launched Claude Code, or retreating into a niche so small that giants don’t bother, as Midjourney has done. Even Cursor now faces a delicate relationship with Anthropic: from close partner to something close to competitor. Yao views Meta’s acquisition of Manus less as a product buy than as a talent anchor in Singapore, tapping Asia’s deep pool of AI engineers.

Looking to 2026, he is most excited about models that are “trained with finite context but used as if they have infinite context.” This would let a persistent assistant engage with a user continuously, discarding irrelevant information and remembering what matters, finally fulfilling the dream of a truly personalized helper. Despite all the talk of slowdowns, Yao sees no deceleration in model progress. Pretraining continues to improve, and what’s needed now is clearer thinking about what problems to solve—because the AI wave is still rising, but it’s up to the surfers to choose the right direction.