In a rare four‑hour podcast with journalist Zhang Xiaojun, Yao Shunyu—a physics PhD from Tsinghua and Stanford who jumped from Anthropic to Google DeepMind last year—delivered a blunt, insider take on the state of large language models. He worked on Claude 3.7, 4.5, and Gemini 3, yet he speaks with a freedom rare among top researchers. His observations cut through hype, targeting sacred cows at Anthropic, OpenAI, and Google itself.
Yao estimates that 90% to 99% of his code is now generated by AI, despite Google officially banning Claude Code and Codex for internal use. If a frontline researcher relies this heavily on AI, the narrative that “I can’t use AI to code” becomes pure vanity. The contradiction is striking: Google’s productivity may be drastically lower than it could be, raising questions about corporate rigidity in an industry where speed matters. Yao’s data point also supports a broader trend—Stack Overflow’s 2023 survey found 44% of developers use AI tools daily, and that figure is rising fast among elite teams.
His departure from Anthropic was partly driven by dissatisfaction with CEO Dario Amodei’s anti‑China stance, which he says accounted for about 40% of the decision. While not the primary factor, it reflects a deeper unease with the “we must build the strongest model to push AI safety” logic, which Yao finds intellectually dishonest. This friction is not isolated: a 2024 report by the Paulson Institute noted that over 30% of top Chinese‑born AI researchers in the US have considered leaving due to geopolitical tensions, costing American labs critical talent.
The naming fiasco of Claude 3.5/3.6/3.7 reveals Anthropic’s early product chaos. Two separate models were accidentally given the same version number, forcing the community to invent “3.6” as a differentiator. Anthropic later followed the crowd with 3.7. This “grassroots naming” mirrors how Linux kernel versions were once managed—community‑driven, not top‑down. For a company positioning itself as the safety leader, such operational sloppiness is revealing.
Claude Code, now one of Anthropic’s most important products, began as a bottom‑up side project by a researcher named Boris. It was personal initiative, not strategic planning, that created the tool. This echoes the origin of Kubernetes at Google, which started as a small team’s experiment. Yao contrasts this with the dysfunction at OpenAI, where top executives have fled in droves. Anthropic’s founding team—core members from OpenAI who fought together—has stayed intact, which Yao believes is the root of its top‑down culture working at all.
One of Yao’s most counterintuitive claims is that OpenAI “saved Google.” If ChatGPT had immediately dominated search, Google would have collapsed. But because OpenAI showed the possibility without fully executing, Google got a lifeline to retaliate. This mirrors Clay Christensen’s innovator’s dilemma: incumbents often survive disruptors that stop short of total conquest. Google’s current AI news, like the Gemini 2.5 Pro release, suggests they are using that window.
Yao dismisses the widespread belief that scaling laws have hit a wall. “Most of the slowdown comes from bugs in the code, not fundamental limits,” he said, noting that fixing a single bug can yield more progress than fancy tricks. Pre‑training has continued to improve in recent months, contradicting the “pre‑training is dead” narrative. This aligns with recent work from DeepMind and Meta showing that careful data curation and architecture tweaks maintain scaling returns.
He also predicts a brutal future for programmers: only 1 in 1,000 will earn 100 times the salary of the rest. Reliability, not raw intelligence, is the key trait. For a physics PhD to say “brilliance is undergraduate work” is a humbling reminder that execution outranks insight. Yao’s view echoes a growing consensus among AI leaders—Sam Altman has hinted that AI will commoditize coding, while Demis Hassabis emphasizes “groundedness” in researchers.
Yao believes most new AI labs will die. The combination of capital intensity, talent scarcity, and the need for long‑term commitment makes survival unlikely. He himself doesn’t expect to stay long at Google, saying openly in the podcast that “there are no mentors, no old friends in this industry—I say what I want.” Chinese AI companies should take note: a top researcher with a track record at both frontiers and a willingness to relocate is a rare catch.
Yao’s interview offers a rare, unfiltered glimpse into the mind of a practitioner who has seen both Anthropic’s chaos and Google’s cautiousness. His honesty is a corrective to the polished narratives that dominate tech conferences. For anyone building in AI, the lesson is clear: focus on reliability over brilliance, fix your own bugs before blaming scaling, and remember that the best products often emerge from bottom‑up tinkering, not top‑down strategy.