Google's AI defense line has completely collapsed, with all eight core Transformer leaders leaving the company. Half of the AI industry is now made up of rivals that Google itself nurtured.

In mid-to-late June in Silicon Valley, a talent earthquake erupted in Google's AI system.

On June 20th, John Jumper, the winner of the 2024 Nobel Prize in Chemistry and the leader of the AlphaFold project, officially announced that he would end his nearly nine-year career at Google DeepMind. After a break, he would join Anthropic.

Just a few days later, two core members of the AlphaFold project under his leadership, Jonas Adler and Alexander Pritzel, confirmed that they would follow suit and switch to the same employer - Anthropic.

Two days before Jumper's announcement, Noam Shazeer, a core author of the Transformer paper and the co-leader of Gemini, announced that he would join OpenAI - which means that all eight original Google authors who wrote "Attention Is All You Need" in 2017 have now left Google's system.

Within 72 hours, two top-tier scientific research leaders and a complete core team of an ace project left one after another.

Behind this series of personnel changes is the concentrated outbreak of Google's continuous loss of AI talent over the past eight years.

From the founders of Transformer to the leaders of AlphaGo/AlphaFold, from the inventors of BERT to the core R & D team of Gemini, this "Whampoa Military Academy" of the global AI industry is experiencing an unprecedented talent diversion.

Google's Talent Turmoil in June

Two Heavy Farewells in 72 Hours

The first thunderclap of this talent storm came from a ceiling-level figure in the field of AI life sciences.

As the creator of DeepMind's most significant scientific research achievements in the past decade, John Jumper's departure is a barometer for the industry.

Just half a year after graduating with a doctorate, John Jumper led a team to tackle the protein folding problem that had puzzled the academic community for 50 years. He successively launched the AlphaFold, AlphaFold2, and AlphaFold3 series of models, analyzing the structures of over 200 million proteins at once and completely rewriting the research paradigm of structural biology.

In 2024, John Jumper and Demis Hassabis, the head of DeepMind, jointly won the Nobel Prize in Chemistry, becoming the youngest Nobel laureates in chemistry in 70 years.

For Google, Jumper is a flag figure in the AI for Science track. For Anthropic, recruiting him is a strategic move - the latter has already laid out in the AI life sciences track, acquiring biotech companies and launching a product line of large models dedicated to biomedicine. Jumper's joining directly elevates its protein calculation and drug R & D capabilities to the first echelon of the industry.

Demis Hassabis highly recognized the achievements of the nine-year cooperation in his public response on the social platform, but it was hard to hide the regret of losing core assets.

What's even more alarming is that this is not a single person leaving, but a whole team migrating in an organized way.

Jonas Adler and Alexander Pritzel, who followed Jumper's footsteps, are not only core technical members of the entire AlphaFold series but also key contributors to the Gemini large model: Jonas Adler is the core author of AlphaFold2 and the leading designer of the biomolecular interaction algorithm of AlphaFold3. At the same time, he is in charge of the R & D in the field of Google AI programming and is the core person in charge of Gemini's code capabilities; Alexander Pritzel is the third author of the milestone Nature paper on AlphaFold2, responsible for the model training architecture and optimization strategy throughout, and also deeply involved in the construction of the distributed training system of the Gemini large model.

Their following means that the core technical team of the AlphaFold project has moved to Anthropic as a whole, not only completing the transfer of AI life science capabilities but also strengthening Anthropic's weaknesses in large model engineering and code capabilities.

Two days before Jumper's announcement, another more symbolic resignation news came out: Noam Shazeer joined OpenAI to be responsible for the research on the next-generation model architecture.

As one of the eight core authors of the Transformer paper and a pioneer of the MoE mixture-of-experts architecture, Noam Shazeer is one of the founders of Google's large model technology route. In 2024, Google spent $2.7 billion to buy him back from Character.AI through technology licensing and talent return and appointed him as the co-leader of Gemini.

However, less than two years after his return, this top-tier architect chose to leave again. Noam Shazeer's departure officially put an end to the "full departure" of Google's original Transformer authors - the technical foundation of an era has no original creators left within Google.

Google's Eight-Year "History of Talent Drain" in AI

The concentrated outbreak in June is just the tip of the iceberg.

In the past eight years, from Google Brain to Google DeepMind, the loss of core talent in Google's AI system has already reached a large scale, covering four major directions: basic models, AI for Science, large model engineering, and product executives. More than 20 core authors of milestone papers and several executives have left one after another.

1. Founders of Basic Models: All Left from Transformer to BERT

Google is the technological origin of contemporary large models, but it failed to retain the people who defined this era.

All eight authors who published the Transformer paper in 2017 have now left Google: Lukasz Kaiser joined OpenAI as early as 2021; Aidan Gomez, Ashish Vaswani, Niki Parmar, Jakob Uszkoreit, Llion Jones, and Illia Polosukhin started their own businesses one after another; Noam Shazeer's second departure extinguished Google's last hope.

The same is true for another milestone, BERT.

Jacob Devlin, the first author of the BERT model, left Google in anger in 2023 to join OpenAI because he questioned the compliance of Bard's training data. Although he returned briefly later, it was hard to cover up the reality of the break of core talent in Google's pre-training era.

These people are the "pioneers" of the large model era. Their departure takes away not only technical experience but also the source power of underlying innovation.

2. Ace Troops in AI for Science: Two Milestone Teams Disintegrated

DeepMind once established an unshakable technological barrier with AI for Science, but now the cores of its two ace projects have all gone.

In the field of protein folding: With John Jumper leading Jonas Adler and Alexander Pritzel to leave collectively, the core R & D team of AlphaFold has basically disintegrated, and Google's technological advantages accumulated in this field have been greatly diluted with the loss of talent.

In the field of reinforcement learning: "The Father of AlphaGo", David Silver, officially left in January 2026 to take full charge of his founded Ineffable Intelligence, focusing on "super learners" in reinforcement learning that do not rely on human data. It raised $1.1 billion in the seed round, setting the largest seed round record in European AI history. Mustafa Suleyman, a co-founder of DeepMind, left to start his own business in 2022 and later joined Microsoft with his team as the CEO of Microsoft AI, taking away Karen Simonyan, a core contributor to AlphaZero.

From AlphaGo to AlphaFold, the core leaders of DeepMind's two most proud benchmark projects have all established their own businesses or defected to competitors.

3. Main Force in Large Model: Continuous Loss of Talent in Gemini R & D Line

In the general large model track of direct competition, the Gemini team has also become a hard-hit area for poaching, and top talents in subdivided tracks have continued to leave:

In the field of reasoning ability: Jason Wei, the core proposer of Chain-of-Thought, joined OpenAI in 2023 and became a key figure in the iteration of large model reasoning ability.

In the field of competition-level reasoning: Dustin Tran, the core creator of Gemini DeepThink, led core member Ashish Kumar to join xAI in September 2025, directly supporting the breakthrough of Grok 4's reasoning ability; in July of the same year, Meta poached Tianhe Yu, Cosmo Du, and Weiyue Wang, three core researchers of Gemini's Olympic math gold medal, at one time to strengthen Llama's mathematical reasoning ability in a targeted way.

In the field of security and efficiency: Nicholas Carlini, a top expert in AI adversarial security, and Neil Houlsby, a pioneer in parameter-efficient fine-tuning, have joined Anthropic one after another, continuously taking away Google's technological thickness in subdivided fields.

4. Executives and Product Layer: Mass Exodus from Co-founders to Project Leaders

In addition to technical talents, the loss at the product and management levels is also serious.

The departure of Mustafa Suleyman, a co-founder of DeepMind, took away a complete productization and commercialization team. Daniel De Freitas, the former head of Google's LaMDA project, left to found Character.AI and became an early leader in the generative dialogue AI track.

From underlying research to upper-level products, Google's AI talent chain is loosening in all aspects.

When a Whale Falls, All Things Thrive

Where Have Google's "Brainiest Minds" Gone?

The top talents lost by Google are not scattered across the industry but are highly concentrated and flowing to four major destinations, precisely corresponding to the current competitive landscape of the AI industry.

1. Google's Number One Rival, OpenAI: Targeted Acquisition of Core Architecture and Capabilities

OpenAI is the largest talent recipient outside of Google, and its poaching logic is highly targeted - it specifically poaches the definers of underlying architecture and core capabilities.

In terms of architecture, it has successively absorbed Lukasz Kaiser and Noam Shazeer, two core authors of Transformer, continuously consolidating its leading edge in the basic model architecture.

In terms of capabilities, it recruited Jason Wei to strengthen reasoning and Jacob Devlin to strengthen pre-training and data systems. Each poaching move targets Google's technological foundation and allows OpenAI to always hold the right to speak in the iteration of large models.

2. The Fastest-Growing Rival, Anthropic: Systematically Completing the Technology Stack

If OpenAI poaches talents point by point, Anthropic absorbs them systematically and gradually builds a complete technology stack that can compete with Google.

In the early stage, it first introduced Niki Parmar, an author of Transformer, to strengthen the model architecture, Neil Houlsby to strengthen model efficiency, and Nicholas Carlini, Milad Nasr, etc. to strengthen AI security and privacy alignment, filling in the core sectors of general large models.

Now, by recruiting John Jumper's team, it has entered the AI life sciences track in one step, forming a two-line layout of "general large models + vertical scientific computing". This poaching rhythm of "first building the framework and then adding aces" has allowed Anthropic to quickly enter the first echelon in just five years.

3. Multi-polar Players Meet Their Own Needs: Precise Positioning in Subdivided Tracks

Players such as Meta, xAI, and Microsoft do not take the route of all-round confrontation but focus on their own weaknesses and poach talents precisely.

Meta targeted the reasoning weakness of Llama and directly poached the core trio of Gemini's Olympic math, achieving a rapid leap in subdivided capabilities at a low cost.

xAI focuses on extreme reasoning and recruits competition-level reasoning experts such as Dustin Tran. Together with the founding team with a reinforcement learning background, it quickly challenges the ceiling of large model reasoning.

Microsoft, by absorbing Mustafa Suleyman's Inflection core team, has filled in the product management and technology coordination capabilities of consumer AI at one time, injecting fresh blood into all Copilot businesses.

4. Entrepreneurship Legion: Half of the Global AI Entrepreneurship Circle

More top talents choose to start their own businesses, and the companies they founded now constitute half of the global AI entrepreneurship circle.

The eight members of the Transformer group can be called the "entrepreneurial dream team".

Cohere, founded by Aidan Gomez, is already one of the three major basic model manufacturers in North America, with a valuation of over $20 billion.

Character.AI, founded by Noam Shazeer and Daniel De Freitas, was once a unicorn benchmark in the dialogue AI track.

Adept, founded by Ashish Vaswani and Niki Parmar, is an early pioneer in the AI agent track.

Inceptive, founded by Jakob Uszkoreit, focuses on AI design of RNA drugs.

Sakana AI, founded by Llion Jones in Tokyo, has become the leader of Japanese AI startups.

NEAR Protocol, co-founded by Illia Polosukhin, is already a leading player in Web3 infrastructure.

Entrepreneurs from the DeepMind system are also very important: David Silver's Ineffable Intelligence raised $1.1 billion in the seed round with just one technical direction.

Without exception, these companies are precisely positioned in the three frontier tracks of basic models, agents, and AI for Science and ultimately become direct or indirect competitors of Google.

Why Can't Google Retain the Top Talents?

Google has no shortage of funds, computing power, or brand. Why is it losing ground in the battle for top talents?

The core reason is not salary, but the most core demands of top scientific researchers - dominance, stability, and a sense of value. And these are gradually disappearing in Google's system.

1. Gap in Autonomy: The Hierarchical Barriers of Big Companies Can't Compete with the Dominant Space of Startups

For top researchers who can achieve era-defining results, the biggest driving force is not job title and cash, but the power to "do a big thing according to their own ideas".

In Google, even senior scientists often have to advance projects through complex hierarchies and approvals. Resource allocation and direction adjustment are subject to the company's overall strategy.

Llion Jones once publicly complained that Google became highly bureaucratic in the later stage, and it was almost "impossible to get anything done". In startups or growing companies, top talents can independently lead the entire technical route and have full say in everything from direction definition to resource allocation.

After John Jumper joined Anthropic, he can directly lead the entire AI life sciences product line.

After David Silver started his own business, he can fully devote himself to the direction of "super learners without human data" that he believes in.

This kind of scientific research freedom of "having the final say" is a core bargaining chip that big companies can hardly offer.

2. Strategic Swing and Organizational Infighting: The Coordination Dilemma after the Merger

In 2023, Google Brain and DeepMind merged into Google DeepMind. It was supposed to be a strong combination, but in fact, it brought about internal strife in team games, resource allocation, and route differences.

From the hasty launch of LaMDA to Bard, from Bard's poor reputation to the strategic adjustment of Gemini, Google's large model business has always been in frequent changes while in a chasing position.

Front-line researchers need to constantly adapt to new priorities, new teams, and new goals, lacking a long-term and stable scientific research environment.

For scientists who are deeply involved in basic research, strategic swing means the interruption of scientific research continuity, which is more unacceptable than insufficient salary.

3. Ineffective Incentive Mechanism: Money Can't Buy Long-Term Value

Google is not stingy with money.

Spending $2.7 billion to buy back Noam Shazeer's team is enough to prove its financial strength and determination to retain core talents. But cash and high salaries can no longer retain talents at this level.

In startups, core talents get equity, which is the long-term value bound to the company's growth. The industry reputation and personal influence brought by project success far exceed promotions and bonuses within big companies.

When the choice of top talents has changed from "how much money to earn" to "how big a thing to do and how much fame to leave", the salary system of big companies naturally loses its competitiveness.

4. Conceptual Differences: The Balance Contradiction between Security, Compliance, and Commercialization

Ideological differences are also a key reason