@AlchainHust: Spent most of a day listening to the 4-hour interview between Zhang Xiaojun and Yao Shunyu. This guy, who just moved from Anthropic to Google DeepMind last year, has worked on Claude 3.7/4.5 and Gemini 3. He offered many candid perspectives from a frontline researcher at a top large model lab. The interview is incredibly information-dense…

Spent most of a day listening to the 4-hour interview between Zhang Xiaojun and Yao Shunyu. This guy, who just moved from Anthropic to Google DeepMind last year, has worked on Claude 3.7/4.5 and Gemini 3. He offered many candid perspectives from a frontline researcher at a top large model lab. The interview is incredibly information-dense. Here are a few points I found most interesting: 1. Google prohibits employees from using Claude Code, but Yao Shunyu conservatively estimates that 90% of his code is AI-generated. Non-conservatively, 99% or even 100%. A PhD in physics from Tsinghua and Stanford, a top large model researcher, relies on AI to write code. If anyone still says they can't use AI to write code, don't flatter yourself. But on the flip side, Google doesn't even allow internal use of Claude Code and Codex—how much does that hurt employee efficiency? This company is really something. 2. Among his reasons for leaving Anthropic, opposition to Dario's anti-China stance accounts for 40%. He says it's not the primary reason but certainly a major one. He is annoyed by Dario's logic of 'we must have the strongest model to promote AI safety.' Not many in the circle dare to directly criticize their former employer from just half a year ago. 3. The naming of Claude 3.5/3.6/3.7 was a slapdash blunder. Anthropic's early product capabilities were extremely weak—'they actually gave two models the same name.' The outside world spontaneously called one 3.6 to distinguish them, and Anthropic later followed the community convention by naming the next one 3.7. I previously thought 3.6 was a skipped version number. 4. Claude Code is 'the beginning of individual heroism.' A researcher named Boris wanted to build a productivity tool himself, and it later became one of Anthropic's most important products. It grew entirely bottom-up, not planned top-down. 5. Not a single member of Anthropic's founding team has left. The core group from OpenAI fought together—that's the foundation that enabled a top-down culture to work. In contrast, OpenAI's executives have all left, and Yao Shunyu seems to look down on OpenAI's corporate culture and some of its executives. 6. OpenAI saved Google's life. The logic is quite counter-consensus: if ChatGPT had swallowed search in one go, Google would have been finished. But because it only made the possibility real without going all the way, Google was given time to counterattack. 7. The most important trait in the AI industry is not intelligence, but reliability. Original quote: 'Those things that require lots of intelligence—undergraduates can do them.' A physics PhD saying this is both a dimensionality reduction strike and a reassurance for anyone wanting to switch to AI. 8. He thinks the future of programmers is that 1/1000 will earn 100 times the salary. Not 'programmers will disappear,' but 'extreme centralization.' The vast majority will lose their unique value, while a few top ones will rake it in. 9. He thinks that many people now say Scaling Law has hit a wall, but most of the time it's their own code bugs. Original quote: 'Fixing a bug brings far more progress than some fancy tricks.' Pretraining has actually been getting stronger in the past few months, completely opposite to the outside narrative of 'pretraining is dead.' 10. The vast majority of New Labs will die. My biggest takeaway after listening: This guy really doesn't hold back. He criticizes Anthropic, OpenAI, and various 'old-timers.' But the basis for his boldness is clear: he is neither on the SSI path nor dependent on LP for his livelihood. His own words: 'In this industry, I have no mentors and no old friends, so of course I can say whatever I want about anyone.' Also, he mentioned that he probably won't stay at Google for long. Putting that out on a podcast makes me think that top large model companies in China should try to recruit him.