---
title: "Yao Shunyu's 4-hour interview: training models at Anthropic and Gemini"
url: https://jacquesao.com/en/reading/yao-shunyu-4h-interview-deep-read/
lang: en
author: Jacques AO
type: reading-note
source_author: "Zhang Xiaojun × Yao Shunyu"
updated: 2026-05-16
description: "A close reading of episode 140 of Zhang Xiaojun's podcast Business Interview Record, a 4-hour interview. Yao Shunyu worked on training Claude 3.7/4.5 at Anthropic and then Gemini 3 at Google DeepMind, making him one of the few researchers with first-hand training experience at both top labs. Cross-checked against public reports and his own blog; about a 40-minute read."
---

# Yao Shunyu's 4-hour interview: training models at Anthropic and Gemini

> A close reading of episode 140 of Zhang Xiaojun's podcast Business Interview Record, a 4-hour interview. Yao Shunyu worked on training Claude 3.7/4.5 at Anthropic and then Gemini 3 at Google DeepMind, making him one of the few researchers with first-hand training experience at both top labs. Cross-checked against public reports and his own blog; about a 40-minute read.

## Key points

- Pre-training has not hit a wall — direct confirmation from two top-tier labs; most 'walls' are actually bugs in your own code you haven't found.
- Agentic coding is the only truly viable AI-native scenario right now; other application layers still lack a real data flywheel.
- AI is essentially simple — simple algorithms + powerful infra; the real bottleneck has shifted from 'can the model do X' to 'is the problem well-defined'.
- M-shaped polarization of programmers — system design + business judgment gain value, pure execution work disappears fast, and the middle layer is the most at risk.
- Individual heroism is over — AI takes hundreds of people working together plus hundreds of millions in compute; what matters is not genius but high-quality, reliable components.
- Robotics and quantum computing are underdog territory for young people — big tech has not locked them down yet, so there is more opportunity.

## Editor's Note

This article is a **close-reading note** on episode 140 of Zhang Xiaojun's *Business Interview Record*—a 4-hour interview with Google DeepMind research scientist **Yao Shunyu (姚顺宇)**. It is not a verbatim transcript of the original program.

The compilation is based on cross-verification of the following public materials: program pages and chapter timestamps on Bilibili, Xiaoyuzhou, and YouTube; public reports and statements from media outlets such as TMTPost, Huxiu, Zhihu columns, Sina Tech, Phoenix Tech, QbitAI, PingWest, Beijing Daily, and DoNews, as well as Yao Shunyu's own public blog. Quoted sentences in this article are **preferentially taken from the original text of Yao Shunyu's personal blog or from publicly cross-verifiable quotations across multiple media**; for viewpoints from the original program, I mainly paraphrase and summarize, avoiding falsely attributing "original words."

The copyright of the original program belongs to Zhang Xiaojun and *Business Interview Record*. Readers with time are strongly encouraged to listen to the full version (link at the end of this article). This article only offers commentary and summary, to help readers who do not have a 4-hour window quickly establish a frame of reference.

---

## Opening: A Portrait First

Yao Shunyu is not an ordinary "AI researcher." To understand this interview, one must first understand why this person is rare.

**His academic credentials may be among the strongest of his generation of researchers.** In 2015, he entered the Department of Physics at Tsinghua University and received the **Special Scholarship**—Tsinghua's highest undergraduate award, given to only a single-digit number of students university-wide each year. As a sophomore, he began taking graduate-level theoretical courses, and the research direction he pursued during his undergraduate years was **topological field theory of periodically driven systems**. In 2018 (still an undergraduate), he published a paper in *Physical Review Letters*, presenting for the first time a new method for **topological band theory of non-Hermitian systems** and defining two new physics concepts. An associate professor at a university in China's Project 211 was quoted repeatedly by multiple media outlets—to the effect that no professor at that university could surpass his undergraduate-level physics.

He then went to **Stanford for his PhD**, where his advisors included **Douglas Stanford** (a star theoretical physicist known for work on the black hole information paradox and the SYK model, and winner of the 2017 New Horizons in Physics Prize) and **Yang Zhenbin**. During his PhD, his direction gradually shifted from condensed matter to high-energy physics. In 2023, he proposed the "**Scramblon Mode**" mechanism—a concrete technical contribution to the study of quantum chaos and information transfer inside black holes.

**In 2024, he spent a few months as a postdoctoral researcher at UC Berkeley**, making the transition from physics to AI. **On 1 October 2024, he officially joined Anthropic's Claude team**—he was about 27 that year. He worked at Anthropic for **exactly 354 days**, **leaving on 19 September 2025** and **joining Google DeepMind 10 days later, on 29 September**, where he worked on Gemini 3 training. The interview was recorded in March 2026, about half a year after he joined DeepMind.

**What makes him rare is a combination of three independent scarcities:**

1. **He has first-hand training experience at both Anthropic and DeepMind**, two top-tier labs. This is extremely rare among Silicon Valley researchers. The two companies have almost opposite styles; he took part in training Claude 3.7 and 4.5 at Anthropic, then moved on to training Gemini 3 at DeepMind. An insider view that spans both labs is almost impossible to find anywhere else in public interviews.
2. **A hard pivot from physics PhD to AI**—and a pivot after genuinely producing top-level physics results, not an escape from "being unable to continue." Therefore his judgments on both fields carry credibility.
3. **Directness in his public stance.** Chinese researchers doing model training in Silicon Valley rarely voice industry opinions in public; he is one of the few exceptions. When he left Anthropic, he wrote directly on his personal blog: "**So Ant, it was good with you, but it is better without you 🙂**"—such a public stance is very rare among Chinese researchers.

His posture throughout the entire interview is consistent: direct, unpretentious, unafraid of offending people, and responsible for his words. "Please allow me to go a little crazy" is the subtitle he set for himself—he did not actually go crazy in the interview; he simply said out loud judgments that many in the industry dare not state publicly.

The interview was recorded in March 2026. At the opening of the program, Zhang Xiaojun added a disclaimer: because the world is changing too fast, a series of major events occurred between the recording and the release (Meta's acquisition of Manus was revoked, Cursor was acquired by SpaceX, etc.), so some time-sensitive judgments may already be outdated; also, due to NDAs, **he could not discuss some technical details**—the interview itself is already a version filtered by dual NDAs from Anthropic and Google.

---

## I. Two Yao Shunyus: Clearing Up the Confusion

The interview addresses this in its first minute: Silicon Valley's AI community has **two researchers from the same Tsinghua undergraduate cohort whose names are both romanized as Yao Shunyu—Yao Shunyu (姚顺宇) and Yao Shunyu (姚顺雨)**. They are repeatedly confused with one another.

**Yao Shunyu (姚顺雨)** (OpenAI → Tencent): graduated from Tsinghua's Yao Class, PhD from Princeton, has worked on agents for 6 years, and is a core author of key papers such as ReAct, Reflexion, and SWE-bench. In 2025 he jumped from OpenAI to Tencent. **Zhang Xiaojun interviewed him in episode 115**—3 hours, discussing agents, humans and systems, and a unipolar versus multipolar world.

**Yao Shunyu (姚顺宇)** (Anthropic → Google DeepMind): graduated from Tsinghua's Department of Physics, PhD in Physics from Stanford, transitioned from physics to AI. **He is the protagonist of this episode, episode 140**.

The two **entered Tsinghua as undergraduates in the same year** and **have long been confused with one another**. In 2025, a "symmetrical" pair of moves compounded this: Yao Shunyu (姚顺雨) moved from OpenAI to Tencent, while Yao Shunyu (姚顺宇) moved from Anthropic to DeepMind—both leaving leading US laboratories in the same year.

In the interview, Yao Shunyu (姚顺宇) responds to the confusion with a wry smile. He stresses that he is not Yao Shunyu (姚顺雨), and that his field is not agents: he works on model training, specifically pre-training and RL post-training.

---

## II. Competition and Escape: Does the AI Application Layer Really Have a Moat?

The interview opens with a sharp judgment—**almost all moats currently lie on the model side**.

This statement is not an abstract "model companies are strong." Expanded, it means:

**First, the data flywheel has not truly appeared.** "Data flywheel" has been the most popular narrative in the AI application layer over the past two years—the more users use it, the better the model/product becomes, which attracts even more users. Yao Shunyu's judgment is direct: no scenario has formed a true data flywheel. Most so-called "flywheels" are actually brand effects or network effects, not data-driven model self-reinforcement.

**Second, the only exception is agentic coding.** This is currently the only scenario where "AI is truly native and has become very successful." People coding on products like Cursor, Claude Code, Codex, and Windsurf genuinely leave high-quality training signals for the model in every action—these signals have clear success/failure determinations (whether the code runs, whether tests pass, whether the user accepts or rejects a completion), and they are generated at a massive scale. This is a property no other AI application layer has. **Writing, customer service, chat, search**—the success signals in these scenarios are much more ambiguous, and the flywheel simply cannot spin.

**Third, there are only two paths for the application layer to survive:**

- **Path one: escape fast enough.** The product's growth rate must exceed the reaction speed of model companies, **occupying users' minds before model companies notice this scenario**—and not just occupying minds, **but also evolving its own model**. Mindshare can only protect you for one or two foundation model cycles; when the next generation of foundation models comes out, if you haven't grown your own model differentiation, you will be eaten. Cursor and Perplexity are already trying this path, but the vast majority of application-layer companies do not have this execution speed.
- **Path two: the market is small enough.** Do vertical scenarios that model companies cannot be bothered to manage. This path can survive, but the ceiling is right there—you will not become the next big company.

**Fourth, his implicit attitude toward the current application layer is cold.** He did not explicitly say "99% of the AI application layer will die," but the logic of his argument actually pushes to this conclusion—most application-layer companies can neither outrun the iteration speed of model companies nor are willing to do small markets, so they are merely "surviving for now."

This passage cuts deep in AI venture-capital circles—over the past two years many funds bet on "models will rapidly commoditize, and the application layer will be winner-takes-all." Yao Shunyu's judgment is exactly the opposite: **models will not rapidly commoditize, and there will be very few application-layer winners**.

---

## III. Pre-training Is Not Over: First-Hand Counter-Evidence from Two Labs

One of the biggest narratives in the industry over the past 12 months is that "**the pre-training Scaling Law has hit a wall**." Investors used this to judge that the valuation ceiling of model companies had been reached and that money should be moved to the application layer.

Yao Shunyu directly gives the opposite judgment: his first-hand experience at both Anthropic and DeepMind is that **it has not hit a wall**.

The gist: pre-training has actually become stronger over the past few months, and in the coming months he sees no sign of hitting a wall. The credibility of this "dual insider perspective" is an order of magnitude higher than that of external observers.

He lays out a very important research methodology—when a researcher "feels the pattern has come to an end," there are actually **three possibilities**:

1. **The pattern's scope of applicability has truly come to an end**—the rarest case.
2. **Some precondition is not satisfied**—for example, data quality has declined, infra has become unstable, some hyperparameter is not aligned, or the tokenizer was changed causing downstream performance misalignment. Medium probability.
3. **There is a bug in your own code or experiment that you have not found**—**the vast majority of cases**.

This is the passage in the entire interview that most violates the intuition of "grand narratives"—many phenomena that appear to be "the end of Scaling Laws" may, back at the engineering site, simply be a bug somewhere, or a problem with the data cleaning pipeline, or leakage in the evaluation set itself.

From this extends a core view of his on the qualities of a researcher: **the progress brought by fixing a bug is often far greater than those seemingly magical tricks**. This sounds plain, but in a field full of "magic narratives" it is very counter-trend—most AI big shots talk about "algorithmic intuition," "understanding of models," and "aesthetic taste," while he instead emphasizes "systematically fixing bugs."

His **specific technical expectation for 2026** points to a very concrete direction: **"Train with finite context, use as infinite context."** Train with a finite context length, but at inference time be able to use a near-infinite context. This is the next capability leap he identifies—currently long context means feeding long sequences during training, which is costly and unstable in effect; if it becomes possible to train with a short window and extrapolate at inference to extremely long or even infinite context, the economics of models would be rewritten. Gemini's generational advantage in long context (2M tokens) also suggests that DeepMind is internally pushing in this direction.

This passage also contains an implicit judgment: **the Scaling Law has not hit a wall, but the form of Scaling is changing**—no longer simply "bigger parameters, more data, more compute," but joint scaling across several dimensions: pretraining + RL post-training + data synthesis + long context. The Scaling Law is not dead; it has become fractal.

---

## IV. The Coding Explosion: 90% of Code Is Already Written by AI

He speaks here with near certainty—**AI's explosion in programming is a fact that has already happened, not a future tense.**

**At the factual level**: in his own and many surrounding researchers' daily work, **more than 90% of code is already AI-generated**. This is not a prediction; it is the real workflow of this group of people in early 2026. Tools like Claude Code, Codex, and Cursor are not assistants but the main force—the researcher's role shifts from "writing code" to "reviewing code, designing tasks, defining goals."

But he makes a very clear-eyed split—**the programmer's job has not been replaced as a whole; what has been replaced is one category within it**:

**The part that will not be replaced**:
- How to reasonably design an implementation plan
- Making the plan fit the company's future direction
- Cross-system integration judgment
- Understanding of the business
- Making the call on security and performance trade-offs

In sum, this is the **"product manager + system architecture" layer**. He even mentions that being a good product manager is something he currently cannot figure out how to train AI to do. This sentence is very important—it reveals a real structural limitation of current LLM training: **a model can learn any task that can be clearly defined, but cannot learn "defining the task in an ambiguous context" itself**. The core work of a product manager is the latter, so it escaped the first wave of AI replacement.

**The part that will be replaced**:
- Execution-type tasks where a manager says "implement this plan and give it to me next Friday"
- Feature development described clearly in standardized language
- Code refactoring for which there are already many similar precedents

His judgment on this part is cold: it may no longer exist in the future.

From this, his judgment on the programmer population is an **M-shaped polarization**: the upper half (system design + business judgment) becomes more valuable (because they can direct AI to amplify output), while the lower half (pure execution) disappears quickly. The middle layer is the most dangerous—it has neither the judgment nor the leverage that AI amplifies.

**Advice for students in school**: the core competitiveness is "how to collaborate effectively with AI." This is not a platitude; it is very concrete—it means do not treat "knowing how to write code" as the core ability, but treat "knowing how to use AI to write code, knowing how to review code written by AI, knowing how to design the parts AI cannot solve" as the core ability.

The implicit message of this passage for Chinese programmers is: **the domestic hiring market's criteria for assessing fresh graduates' coding ability are already outdated**. Rote knowledge, drill questions, LeetCode—AI does all of these better than people. What should really be assessed is "under an ambiguous requirement, can you reasonably break down the problem, can you review errors in AI output, can you take over where AI is not good at."

---

## V. Seedance, Hard Distillation and Soft Distillation: An Underrated Engineering Topic

This passage is somewhat more technical, but it is crucial for understanding the real barriers in the model industry.

**Model distillation** is the technique of compressing a large model's capabilities into a small model. Yao Shunyu distinguishes two approaches:

**Hard Distillation**: use a large model to generate data (for example, have a large model generate 1 million answers), then have a small model perform supervised learning on this data. This is the most common and crudest approach. Almost all "small models benchmarked against top closed-source models" in the open-source community use hard distillation. Its problem is: you only learn the large model's "sampling results," **losing the probability distribution information of the large model at each token**—the large model says "A is very likely, but B also has a 20% chance," while the small model produced by hard distillation only learns "the answer is A," and B's possibility is completely lost.

**Soft Distillation**: directly align with the large model's output **probability distribution** (logits). The small model learns not only the answer but also the large model's confidence in each candidate answer. **The information density is an order of magnitude greater**—in theory, a small model of the same size can approach more of the large model's capabilities.

**Why are there only a few teams in the industry that do soft distillation well?** Because it places extremely high demands on infra and details:
- It requires the ability to fully save the large model's logits on the training data (both storage and bandwidth are huge overheads)
- It requires ensuring that the tokenizers of the large and small models are completely identical
- It requires handling the trade-off between the distillation objective and the original pretraining objective
- If anything goes wrong anywhere in the engineering pipeline, the distillation effect degrades to worse than hard distillation

This passage returns once again to his theme throughout—**simple algorithms + powerful infra are what produce results**; algorithmic cleverness alone is useless.

**Seedance** is ByteDance's Doubao-family video generation line. In this passage, Yao Shunyu implicitly places ByteDance's engineering capability in a relatively leading position—being able to deliver in a direction like video, which is "extremely picky about infra," shows that the engineering team is there. In China's AI industry, ByteDance is one of the few companies he mentions relatively positively in the interview.

**Why is this segment undervalued in public discussion?** Because the technical barrier is high and it doesn't grab attention. But this is precisely where real model companies pull apart from one another—not who has the smarter idea, but whose infrastructure is more solid and whose handling of details is cleaner.

---

## VI. Robots + Betting on Underdog Territory

This segment is about robots and embodied AI, and Yao Shunyu's attitude is far cooler than the current Silicon Valley mainstream.

**The current mainstream narrative on robots/embodied AI**: foundation models have solved the "brain," and what remains is the "body" problem; a ChatGPT moment is expected in 2026-2028—humanoid robots will explode. Nvidia is pushing "physical AI" as the next trillion-dollar narrative.

**Yao Shunyu's view**: the opportunity exists, but the pace is widely overestimated. Reasons:

- **The feedback loop in the physical world is inherently slow**—training a chat model, you can run countless inferences per second; training a robot to grasp a cup, each failure may require repositioning the cup and recalibrating the camera, one or two orders of magnitude slower.
- **Data cannot be copied infinitely**—text can be scraped from the entire internet, but robot data can only be collected live. Simulators are useful but there is a sim-to-real gap.
- **Hardware iteration cycles are long**—software versions can ship daily, but a robot hardware design takes at least several months per iteration.
- **The cost of accidents is high**—a robot falling over, hitting someone, or damaging objects is a real cost every time.

But he is **not pessimistic**. He used a very key phrase: **"Underdog territory."**

When a track has no clear winner yet, the giants are still watching, and the technical path has no standard answer—**such places actually offer more opportunities for young people and small teams**. The reason is simple: in fields where big-company resource monopolies have not yet formed, what you build is credited as a "small-team achievement"; in already mature fields, whatever you do will be replicated by big companies with more resources.

**He recommends robotics to young people alongside quantum computing**: quantum computing and robotics/embodied AI are currently the directions offering relatively more opportunities to young people. Common traits:
- Big companies have not yet formed absolute dominance
- The technical path has genuine uncertainty; it is not merely a contest of engineering volume
- It still requires considerable time, which is actually friendly to new graduates and young PhDs

**His implicit attitude toward young people in the pure-software AI field** is cool—this area is already locked down by the compute monopoly of OpenAI, Anthropic, and Google, and it is hard for an individual young person to make a directional contribution anymore.

---

## VII. From Non-Hermitian to AI: How He Made This Turn

This is the richest personal-narrative segment of the interview, and also the most unconventional part of his life path—**having done physics to the level of top journals and top scholarships, why leave?**

### 7.1 How hardcore was his physics research direction

Many media outlets describe him as a "Tsinghua physics top student," but exactly how hardcore requires unpacking:

- **Undergraduate Special Scholarship**: only a single-digit number of recipients per year at Tsinghua.
- **2018 PRL paper**: as an undergraduate, he published a first-author paper in *Physical Review Letters*, giving for the first time a topological band theory for non-Hermitian systems.
- **Stanford PhD**: advised by Douglas Stanford (a star theoretical physicist working on black holes, the SYK model and quantum chaos, and a 2017 New Horizons in Physics Prize winner) and Yang Zhenbin.
- **2023 proposed the Scramblon Mode mechanism**: a concrete technical contribution to the study of quantum chaos and black hole information transfer.

In other words, he did not "fail to continue studying physics"; he **left voluntarily after reaching the top 1% of his generation**.

### 7.2 Why he left physics

The reasons he gave in interviews and blog posts are very concrete, not an abstract "AI is the future."

**First, high-energy/theoretical physics has reached a stage where "experiments cannot keep up at all."** String theory, supersymmetry, and various conjectures about quantum gravity have been proposed for decades, but the energy scale of the LHC generation of accelerators cannot go higher, so they cannot be verified. **Theory has run too far ahead, and experiments cannot keep up**—researchers are in a state of **no objective feedback**. In his own blog there is a precise formulation: theoretical physics has lacked experiments for years, making it very difficult to objectively judge the importance of a theoretical work. This is the **core** reason he gave for leaving.

**Second, AI is the opposite—you can do any experiment.** Physics cannot be understood without labeled experimental data; AI is not bound by this—it doesn't matter if you can't understand it, you can still move forward. In AI he **can do any experiment he wants**, as long as compute and infrastructure keep up. This "democratization of experiments" is enormously attractive to a physicist accustomed to "theory must have experimental verification."

**Third, the boundary between empirical laws and scientific laws is inherently blurry.** The laws of thermodynamics were initially empirical laws too (the macroscopic laws summarized by Kelvin and Clausius), and only later did Boltzmann establish a microscopic statistical-mechanical explanation. Scaling Law currently has no first-principles explanation, but that does not mean it is not true.

**Fourth, he chose AI rather than quantum computing.** He believes AI's development prospects are more mature than quantum computing's. Quantum computing is still waiting for the critical point of "practical quantum advantage," possibly another 10 years; AI is already creating value. This is an engineer-style judgment, not a physicist-style one.

### 7.3 Hard skills physics brings to AI: almost none

Yao Shunyu says this very bluntly: physics actually offers little help for hard AI capability. The transfer of instrumental skills is **very, very small**.

He did not say studying physics was wasted—what remains is **character and ways of thinking**:
- **The habit of getting to the bottom of things**—when encountering a phenomenon, he does not accept surface explanations and must keep asking down to the root.
- **Intuition for complex systems**—people who do condensed matter are used to "many-body interactions emerging into macroscopic behavior," and this intuition transfers naturally to neural networks.
- **Patience with experimental iteration** — a physics experiment takes months per run, while an AI experiment takes a few hours per run — by contrast, iterating on AI experiments feels almost relaxed to him.
- **A conviction that theoretical hypotheses must be experimentally verifiable** — this point is crucial, making him wary of work that "looks beautiful on paper but can't be run."

**Why does Anthropic hire physics PhDs at scale?** Dario Amodei himself comes from a physics PhD background. He has publicly explained that hiring people with physics PhDs enables them to quickly learn machine learning and make contributions. Anthropic currently has about **30-40 researchers with a physics background**. Yao Shunyu joining in 2024 was part of this wave. He himself has also commented that Anthropic remains one of the best places for physics PhDs transitioning into AI research — even after he later left, he still holds this judgment.

---

## VIII. Training Claude 3.7 and 4.5 at Anthropic, and That Public Departure

This is the most technically substantive and longest segment of the interview (about 1:52:32–2:35:03, around 45 minutes).

### 8.1 Timeline

- **1 October 2024**: Joined Anthropic's Claude team.
- **October 2024–September 2025**: Helped establish the reinforcement learning foundations team, working on the framework for Claude 3.7 Sonnet and the reinforcement learning theory behind the Claude 4 series. Claude 3.7 was released in February 2025; Claude 4.5 was released around the time of his departure.
- **19 September 2025**: Left Anthropic.
- **29 September 2025**: Joined Google DeepMind, working on Gemini 3 training.

In total, **354 days** at Anthropic.

### 8.2 Core methodology learned at Anthropic

**In one sentence**: **Simple algorithms + powerful infra are enough to produce results**.

This is the most profound lesson he learned at Anthropic. He repeatedly emphasizes not to chase algorithmic tricks — often Team A and Team B start from almost identical papers and ideas, and the gap comes down to: **whose infra is more stable, whose experiments run more cleanly, whose bugs get fixed more thoroughly**.

The research culture behind this is consistent throughout Anthropic's founding team (Dario, Jared Kaplan, etc.) — the literal meaning of the "Scaling Hypothesis" is: just make simple things bigger, don't try to bypass scale with complex algorithms. The advantage of this culture is focus; the disadvantage is that researchers' individual creative space is relatively small.

**Training focus of Claude 3.7 / 4.5**: improving coding ability through RL. This is Anthropic's bet on the Coding direction — and this bet has been validated by the market: Claude Code became one of the de facto standards in the developer market in 2025-2026, and Cursor long used Claude as its default model. Yao Shunyu's work on the RL foundations team directly served this product success.

### 8.3 Anthropic's organizational character: top-down

Yao Shunyu's description of Anthropic's organizational style: **top-down**. Major directions are decided directly by co-founder Jared Kaplan and others, and researchers execute within the defined directions. The advantages of this organization:
- **High efficiency** — directions don't swing back and forth
- **Concentrated resources** — everyone focuses on a few common goals
- **Strict quality gates** — the core team reviews key decisions

Disadvantages:
- **Less "innovation autonomy" left for researchers** — it's hard to open a new direction yourself
- **Not suitable for people who like bottom-up exploration**
- **The tension with individual creativity may cause frustration**

In the interview he did not directly criticize this Anthropic style, but the implicit attitude is that **this culture suits a particular stage and a particular type of researcher, and not everyone is suited to stay long-term**.

### 8.4 Work intensity

He describes Anthropic's work intensity as "super busy" — so busy that there was no time to write papers recording his thoughts and journey. This point is very important — it means Anthropic researchers' output is mainly internal (the models themselves), with relatively few public papers. This also explains why Anthropic gives the outside world a stronger "technical mystique" than OpenAI — their people have no time to write papers explaining what they are doing.

### 8.5 Departure: the precise ratio of 40% / 60%

The interview is relatively restrained, but combined with his **departure blog** the picture is complete.

**In the blog he gives a precise ratio:**
- **40%** is the **publicly disclosable reason**: divergence in values
- **60%** is the **non-disclosable reason**: internal details, protected by NDA

**What exactly is the content of that 40%?**

On **5 September 2025**, Anthropic listed **China among "adversarial nations"** in an internal document and announced that it would **stop providing AI services to "companies controlled by Chinese entities."** This provoked a strong reaction among Chinese researchers in Silicon Valley.

Yao Shunyu's public statement in his blog was roughly: he strongly opposed the anti-China remarks published by Anthropic; although he believed that most people at Anthropic would oppose such remarks, he felt he could no longer stay at the company.

This is the reason he publicly gave for leaving. Note that his wording is very precise—he **did not** accuse the entire company of being anti-China, and he **explicitly** said that "most people would oppose such remarks"—what he criticized was the **official statement at the company level**, not his colleagues.

**Farewell message** (written in the blog): **"So Ant, it was good with you, but it is better without you 🙂"** ("Ant" is short for Anthropic, a pun on "ant")

This kind of public stance is very rare among Chinese researchers in Silicon Valley. In the interview he did not elaborate much on this episode, but he did not avoid it either—he did not want the company he worked for to take such a position on China-related issues.

### 8.6 Researcher Qualities: Reliability, Self-Reflection, and a Big-Picture View

This section is also the part of the interview that most has the flavor of "a master teaching an apprentice." The core qualities he listed are:

**1. Reliability is ten times more important than intelligence.** The gist—AI work actually does not require much brainpower; the most important trait is being reliable, meticulous, and responsible for what you do. This sounds less like something a researcher would say and more like something an engineering manager would say. But this is exactly what he wants to emphasize—**AI research today is no longer driven by "inspiration"; it is driven by "execution."** Ten thousand clever ideas cannot produce one usable model, but one hundred reliable people can.

**2. The ability to reflect on oneself.** "If what I did has good results, is it really good?" This is the core of judging whether a researcher can be trusted over the long term. **Good metrics do not mean the work was done right**—it could be data leakage, evaluation-set bias, or some spurious correlation. A researcher unwilling to self-check is more dangerous the "better" the results look.

**3. A big-picture view.** If a researcher cannot consider the whole picture, they are not a good researcher. This point has particular relevance for researchers trained in Chinese academia—many PhDs are trained to "do small problems to perfection," but systemic engineering like large AI models does not allow you to look only at your own part.

**4. People are components of a system.** He has repeatedly expressed that "a person should become a more trustworthy and reliable component of a system." This sounds like belittling individuals, but what he means is: **people who can truly do AI well over the long term are not here to be heroes; they are here to be high-quality nodes in a system.**

---

## IX. "AI Is Essentially Simple": The Epistemology of the Current Stage

This is a judgment that forms its own section in the interview (about 2:35:03).

He does **not mean that AI is not hard**; he means that at the current stage, the driving forces behind AI progress are a very small number of simple principles—pretraining + post-training + compute + data. **It is not because of cleverness; it is because of execution.** He does not really believe any narrative that makes it sound mysterious or full of "magic."

The force of this view needs to be seen in the discourse of the AI industry—over the past three years, AI companies' marketing, technical blogs, and CEO speeches have generally followed the narrative of "we made some very creative algorithmic breakthroughs." Yao Shunyu has been at two of the top labs, and his counter-observation is: **real breakthroughs rarely come from algorithms; most come from engineering.**

From this he derives a core judgment about the current research paradigm: **AI has entered a stage where people are not worried about whether AI can do something, but about whether the thing is well defined.**

This is a very important epistemological shift. In the past few years, the core question was "can it"—can it write code, can it be multimodal, can it handle long context. Now most of these "can it" questions can be answered yes. **The bottleneck today has become "how to define it clearly"**—how to write the reward function? How to design the evaluation set? What counts as success? What counts as failure?

He elaborates: on paper everyone looks similar; what is harder is how to define the problem and define the desired behavior; when the problem is not clearly defined, model differences come from things you cannot imagine.

An example to help understand: if two companies both want to make a model "safer"—but Company A defines "safety" as "refusing any risky request," while Company B defines "safety" as "giving constructive alternatives to risky requests"—then even if the two companies train with exactly the same algorithms and data, the resulting model behavior will be completely different. **The real difference lies in the precision of the problem definition**, not in the algorithm.

**What does this imply for Chinese large-model companies?** He did not say it explicitly, but the implication is—catching up to the capability ceiling of GPT-4 / Claude 4.5 is not the hardest part; the hardest part is "defining clearly" each specific capability, which requires a great deal of product judgment and a meticulous evaluation system. This is often an underestimated weakness of Chinese companies—many companies have compute, data, and papers, but **they have not defined "how the model should behave" to that level of precision**.

There is also an extension in this section: **systematicness in doing things**. For him, what matters more is the system of doing things—when things differ from predictions, can you systematically rule out various possibilities? This returns to the research methodology that runs through the whole interview—**the value of a researcher is not in flashes of genius, but in systematic troubleshooting**.

---

## X. Training Gemini 3 at Google DeepMind: The Opposite Organizational Temperament

He joined DeepMind on 29 September 2025; by the time the interview was recorded in March 2026, he had been there for about half a year. In this section, he compares the two organizations' cultures.

### 10.1 Cultural comparison: DeepMind vs Anthropic

**Anthropic**: top-down, focused, research directions are locked in, resources concentrated, high execution efficiency, low individual autonomy.

**DeepMind**: **bottom-up, decentralized, researchers have greater autonomy over direction, relatively lower coordination efficiency**.

The differences he described are very concrete—at Anthropic you are given a pre-framed bet ("we are going to RL a top-tier model on coding, and everyone works around that"), while at DeepMind you have to **prove the bet yourself** ("you want to pursue this direction? Convince me, show me the data").

Neither culture is absolutely superior—companies at different stages and different types of researchers suit different cultures. **His own preference is clearly the latter**—which is also why he left Anthropic after 354 days.

### 10.2 DeepMind's historical advantages remain

DeepMind is the home of AlphaGo and AlphaFold—its research culture is naturally "long-distance running, deep cultivation, bottom-up exploration." This culture seemed disadvantageous in the first two years of the large-model race (OpenAI ran faster), but the comeback of the Gemini series in 2025-2026 proved that DeepMind's foundation is still there—especially in **long context**, **multimodality**, and **scientific applications**, where Gemini began to establish a clear differentiated advantage in 2025-2026.

Yao Shunyu did not elaborate much on the specific training details of Gemini 3 (due to NDA restrictions), but he emphasized one judgment: **Gemini's generational advantage in long context is real**. This echoes the direction he mentioned earlier: "train with finite context, use as infinite context"—the Gemini team has accumulated experience in this direction.

### 10.3 Another layer of logic in choosing DeepMind

Besides value factors, there was also a research motivation for choosing DeepMind—**Google, as the parent company of a search + content ecosystem, can provide training resources for model training that Anthropic cannot obtain**. Video, long text, cross-lingual, multimodal—Google's internal ecosystem has structural advantages in the coverage of these data. For a researcher who "wants to tackle hard training problems," this is attractive.

### 10.4 The subtlety of company sentiment

In the interview, his description of DeepMind is more relaxed and gives a greater sense of control than his description of Anthropic—but he also reminds himself (and readers): he has only been at DeepMind for half a year, **and has not yet gone through the test of "maintaining stability at large scale"**. His judgment of DeepMind is currently positive, but he does not draw absolute conclusions. This **restrained honesty** is also one of the reasons this interview is convincing.

---

## XI. Technology Predictions and Team Building: Specific Bets for 2026

This section is the core of the final part of the interview (about 3:01:28).

### 11.1 Technology predictions

**1. Pretraining continues to scale**, and at least in the coming months no wall has been seen. It is not that "pretraining will never hit a wall," but that at present it has not yet reached the point of hitting one.

**2. Paradigm shift in long context**: "Train with finite context, use as infinite context." This is the most specific next technological leap direction he pointed out.

**3. Agentic coding continues to be the only truly native and successful scenario for AI**—other scenarios are still being explored. This is completely consistent with his earlier product judgments.

**4. The real implementation of agents is limited by "problem definition" rather than "model capability"**. In other words, the agent breakthrough in 2026 will not come from smarter models, but from **defining more precisely what the agent is supposed to do**.

**5. Robotics and quantum computing are underdog opportunity areas**—friendly to young people, unfriendly to resource monopolies of big companies.

**6. AI centralization intensifies**—most value is concentrated in the hands of a few companies, a few products, and a few people.

### 11.2 Organization building

On how to build a good AI team, his core view: **A good AI team is not a gathering place for geniuses, but a network of high-quality, reliable components.**

The trade-off when hiring: **Better to have reliable average people than unreliable geniuses.**

This point hits managers especially hard—many AI team failures are not because they failed to hire smart people, but because they hired too many "stars" who cannot be trusted. **One unreliable genius can pollute the entire system**—you cannot trust their experimental results, others dare not touch the code they write, priorities are not aligned with the team—in the long run it is a negative asset.

He extended a specific suggestion: **The essence of trust is reflection**. A truly trustworthy researcher will reflect on "whether it is really true that what I am doing works well"—rather than only working on metrics. **A researcher who does not self-check is more dangerous the faster they run**.

### 11.3 "Maintaining stability at large scale"

He also mentioned an atypical management philosophy: the key question for a researcher is not "can I produce good results," but **"at truly large scale, can I remain stable"**.

It is not hard to get a good result from a single experiment; **what is hard is maintaining quality across dozens of parallel experiments over a scale of months and quarters**. **There are many researchers who shine in the short term, but very few who are stable over the long term**—the latter are the truly scarce resource.

---

## XII. The Triumph of Collectivism: The Age of Heroism Is Over

The most memorable passage of the entire interview, and also the subtitle singled out in the Bilibili description.

### 12.1 Main Argument

**The era of individual heroism in AI is over, so there are no heroes either.**

This sentence is not sentimentalism; it is a structural judgment.

### 12.2 Structural Argument

**First, scale itself determines that an individual cannot be the key variable.** Building a top-tier model today requires the coordination of **hundreds of people, hundreds of millions of dollars in compute, and several months of time**. Pre-training routinely involves 10²⁵ FLOPs, requiring clusters of tens of thousands of GPUs, hundreds of engineers maintaining the infrastructure, dozens of researchers handling data/algorithms/evaluation, and several product-side people responsible for aligning objectives. If any one link fails, the whole thing collapses.

**At this scale, no single individual can determine the outcome.** Even figures of the caliber of Ilya Sutskever or Demis Hassabis could not produce a GPT-4-level model on their own, outside the organization they belonged to at the time.

**Second, therefore the individual's position in the grand narrative has changed**: everyone now is a surfer, but essentially we are the wave, not the surfer. The subtlety of this sentence lies in "essentially we are the wave"—meaning that individual success is not due to individual ability but is lifted up by the wave of the times. **The researcher has gone from "hero" to "a point on the wave".**

**Third, what really matters is not the upper limit of individual ability but the stability of the collective**: at a truly large scale, can stability be maintained? This requires understanding the system and a sense of responsibility toward the company.

Accordingly, his description of his own work is also restrained—he emphasizes that he merely had the opportunity to join an important project at that time. Note the wording—it is "had the opportunity to join", not "accomplished something". He repeatedly guards against mythologizing narratives about individuals.

### 12.3 Corollary: AI's Social Impact

From the triumph of collectivism, he derives a sharp judgment about social impact: **AI is a highly centralized technology; it will make a small minority stronger while causing the majority to lose their unique value.**

This is his coldest statement on AI's social impact—**neither optimistic nor pessimistic, but structural**.

**A small minority become stronger**: they stand at the fulcrum of AI as a lever, and their work output is amplified by orders of magnitude—researchers at leading model companies, product managers who can use AI efficiently, the few who can define AI tasks.

**The majority lose their unique value**: their past work happens to be exactly what AI can now replace—execution-oriented programmers, junior translators, junior writers, junior customer service, junior analysts. "Unique value" disappearing does not mean people have no value; it means **the differentiated contribution they can offer to the market has been flattened by AI**.

This judgment applies not only to the industry but also to individual choices—**the gap between those standing at the fulcrum of the lever and those flattened by the lever will widen rapidly**.

### 12.4 Advice for Young People (the Densest Passage)

**1. Leaving school means your career is your own**: at school you can have a student mindset, but after leaving school it is your own career, and you have to find a path for yourself. Warning: many people still use a student mindset after leaving school—waiting for teachers/bosses to assign tasks, waiting for mentors to point the way, waiting for the organization's recognition—and this mindset will be rapidly eliminated in the AI era.

**2. Don't wait on old geezers**: a person's life is not that long, so why waste your time waiting on old geezers? "Old geezer" (laodeng) is an internet term referring to senior authority figures (especially those who rely on authority relationships rather than real contributions). This sentence is very direct, and it is a strong implicit critique of Chinese workplace culture—many young people in domestic workplaces are trained to "first serve authority, then fight for space", and he says outright that **this is a waste of life**.

**3. Don't be afraid of offending people**: don't worry too much about provoking someone because of your views; as long as your views are self-consistent and you have your own understanding, people will respect you. He practices this himself—when he publicly left Anthropic he directly wrote "better without you", offending a company worth a hundred billion dollars, yet **his judgment was instead respected more by the industry because of it**.

**4. Objective evaluation standards exist**: ultimately, how well you do in this field is subject to objective evaluation standards. It is not about how well-connected you are or how popular you are, but about what you have produced. This is a fundamental challenge to the path of "getting ahead at a big company through connections"—**in the long run, connections without objective contributions depreciate**.

**5. No one can coast into retirement**: **in the field of generative AI, no one can coast into retirement**, unless you have completely lost interest in technology and have no aspirations for yourself. This includes himself—he left Anthropic to continue fighting at the frontier, not for better benefits or work-life balance at Google.

---

## Closing: Condensing This Interview into a Single Card

If you can remember only 10 things:

1. **Pre-training has not hit a wall**—verified with first-hand data from two top-tier labs. The illusion of hitting a wall is mostly a bug in your own code.
2. **Agentic coding is the only truly viable AI-native scenario right now**; other scenarios have not yet produced a real data flywheel.
3. **AI is essentially simple—simple algorithms + powerful infrastructure**. The gap lies in execution, in bugs, in problem definition.
4. **The current bottleneck is not "whether AI can do it" but "whether the problem can be clearly defined"**. Model capability is constrained by the precision of task definition.
5. **M-shaped polarization among programmers**: system design + business judgment become more valuable, while execution-type work rapidly disappears. The middle layer is the most dangerous.
6. **Individual heroism is over**. Doing AI means coordination among hundreds of people + several hundred million in compute; what matters is not genius, but high-quality, reliable components.
7. **Reliability matters ten times more than cleverness**. AI work does not really require brains; the most important trait is reliability.
8. **AI is a centralized technology**: a small number of people become stronger, while most people lose their unique value.
9. **Robotics and quantum computing are the underdog territory for young people**—the big tech companies have not yet established absolute dominance.
10. **No one can coast into retirement in generative AI**, unless they have already lost interest in the technology.

### Yao Shunyu's stance in this interview

**Do not romanticize technology, do not romanticize the individual, do not romanticize the future.**

His observations on the AI industry resemble neither Sam Altman's grand narrative of "AGI changing humanity" nor Yann LeCun's reserved attitude that "current LLMs don't work"—he **calmly stands at the engineering site and says**: what works now works, there is still a path ahead, but do not package it as something divine.

His attitude toward personal choices corresponds as well: **do not wait to be recognized, do not cater to the old guard, do not fear offending people, work according to objective standards, and there is no retirement in generative AI**. The direction these suggestions give to young Chinese researchers is especially clear—do not take the path of personal connections; take the path of objective contribution.

His own words at the end are the best footnote to this interview: **"Please allow me to go a little crazy."** He did not actually go crazy in the interview—he simply said out loud judgments that many people in the industry dare not state publicly.

---

## Original Program and Public Sources

**The original program (listening to the full version is strongly recommended):**

- [Xiaoyuzhou: Zhang Xiaojun Jùn | Business Interview Record](https://www.xiaoyuzhoufm.com/podcast/626b46ea9cbbf0451cf5a962)
- [Bilibili full video of episode 140 (with chapter timestamps)](https://www.bilibili.com/video/BV1YR5E6EE9o/)
- [YouTube full version of episode 140](https://www.youtube.com/watch?v=Gk_KUg3qED0)

**Public secondary reports used for fact cross-verification:**

- [TMTPost - Yao Shunyu's 4-hour in-depth interview, which we summarized in 30 sentences](https://www.tmtpost.com/7987040.html)
- [Huxiu - reprint of the same 30 sentences](https://www.huxiu.com/article/4858125.html)
- [Zhihu column - summary of Google DeepMind's Yao Shunyu 4-hour interview](https://zhuanlan.zhihu.com/p/2037497180299186879)
- [Sina Tech - Legendary Tsinghua physics student Yao Shunyu leaves, disagrees with Anthropic, joins DeepMind](https://finance.sina.com.cn/tech/roll/2025-10-08/doc-infteazp0504775.shtml)
- [Phoenix Tech - Strongly opposes Anthropic's China-related remarks](https://tech.ifeng.com/c/8nHo3SfgTN5)
- [QbitAI - Fundamental disagreement with Anthropic's values](https://www.qbitai.com/2025/10/339852.html)
- [Huxiu - Yao Shunyu leads Google's "revenge" on Anthropic: it's better without you](https://www.huxiu.com/article/4835161.html)
- [Sina Finance - Dissatisfied with American AI company's "anti-China remarks," Chinese scholar changes jobs](https://finance.sina.com.cn/jjxw/2025-10-09/doc-infthnhk9671307.shtml)
- [Tencent News - Claude team happily welcomes Tsinghua physics top student Yao Shunyu](https://news.qq.com/rain/a/20241028A064WF00)
- [PingWest - Yao Shunyu joining DeepMind draws attention](https://www.pingwest.com/w/308129)
- [Beijing Daily - Strongly opposes American AI company's anti-China remarks](https://news.bjd.com.cn/2025/10/09/11343076.shtml)
- [DoNews - Legendary Tsinghua physics student Yao Shunyu joins DeepMind](https://www.donews.com/news/detail/1/6167272.html)

---

## Copyright and Reprinting

- **Original program** © Zhang Xiaojun, *Business Interview Record*. This article does not copy the original program's content; it only provides secondary synthesis and commentary based on public reports.
- **This article** is a personal reading note; reprinting is welcome, but please retain the author's byline and the original link.
- If you have enough time, it is strongly recommended to bypass these notes and listen directly to the 4-hour original program—the density of viewpoints, Yao Shunyu's tone, and Zhang Xiaojun's pacing of questions cannot be replicated by secondary synthesis.
