Event Overview
Noam Shazeer, Vice President of Engineering at Google, Co Technical Leader of Gemini Models, and Core Author of Transformer Architecture; Recipient OpenAI, CEO Sam Altman.
On June 18th, Shazel announced on social media that he would resign from Google and join OpenAI as the head of architecture research.
On June 18th, I posted a personal statement, and on June 20th, Google officially confirmed my resignation. In 2024, Google spent $2.7 billion to acquire Character.AI, which was founded by me, and recalled me. I left again after less than 2 years.
The Silicon Valley technology talent mobility event is aimed at the global AI industry.
Shazel hopes to focus on the research and development of the underlying architecture of general AGI, and OpenAI has been recruiting him to join for a long time.
Both social media platforms simultaneously announced, with Altman publicly welcoming and industry media following up with reports.
Core Highlights
Core level talent loss in the industry
Shazel is the core author of "Attention Is All You Need", designing a multi head attention mechanism. In 2016, he proposed the MoE hybrid expert architecture, and all major model underlying technologies around the world rely on his research results, making him the founder of AI underlying architecture.
Google's sky high repurchase price leads to talent leaving again
In 2024, Google spent $2.7 billion to acquire the Character.AI team and recalled Shazel, who was fully responsible for Gemini's core research and development. He was a key figure in Google's fight against the GPT series and had worked for less than two years before switching to a competitor.
Altman's ten-year invitation finally lands
During OpenAI's startup phase, Altman identified Shazel as his top target talent and successfully recruited him after a decade, demonstrating his irreplaceable value in the field of underlying architecture.
Google experiences a chain of talent loss during the same period
Two days after Shazel's resignation, DeepMind Nobel laureate John Jumper announced his job switch to Anthropic, causing the successive loss of Google's two core technology leaders and intensifying pressure on the AI research and development team.
Industry background
The current global iteration of large-scale models has entered a deep water zone of tackling underlying architecture, with computing power, model architecture, and sparse training technology becoming the core moats for manufacturers, and top architecture researchers are extremely scarce. Silicon Valley has erupted in a sustained battle for AI talent, with top companies investing heavily in core talent: the annual salary range for top AI researchers is between 3 million to 10 million US dollars, Meta、OpenAI、 Google and Anthropic poaching each other has become normalized.
Google relies on the dual team of DeepMind and Google Brain to layout Gemini, but the internal hierarchy is cumbersome and the product landing pace is conservative. Several core researchers have left due to limited research and development concepts; Institutions such as OpenAI and Anthropic, which focus on general AGI, have more centralized research and development paths, shorter decision-making chains, and continue to attract low-level architecture talents. Talent flow shows a clear one-way trend from Google to OpenAI series.
Expert/official citation
Noam Shazel (former co head of Gemini at Google and new head of architecture research at OpenAI): "This is a difficult decision, and I am immensely proud of the results that the Google team has built with us. I am looking forward to collaborating with the OpenAI team to explore cutting-edge architectures. X platform posted on June 18th
Sam Altman, CEO of OpenAI: "Shazel was the researcher I most wanted to collaborate with at the beginning of OpenAI's founding, and this invitation has been waiting for a whole decade. All the waiting has been worth it. Public comments on social media platforms
Gil Luria (D.A. Davidson Technology Industry Analyst): "The talent in the underlying architecture directly determines the upper limit of the next generation of big models. Compared to Google, OpenAI has a lighter organizational structure that can give top researchers complete technical leadership, and its talent attraction continues to increase. Interpretation of the Industry Research Report on June 21st
impact analysis
Regarding the competitive landscape of the industry
OpenAI will replenish its top-level research and development capabilities in the underlying architecture, significantly improving the iteration speed of MoE and Transformer. The performance of the next generation GPT series models is expected to widen the gap; Google Gemini has lost its core architecture leader, and the iteration of underlying technologies has slowed down, making it difficult to narrow the gap with GPT products in the short term. Google's competitive pressure in the general large model track has significantly increased. At the same time, the incident will further intensify the global AI talent arms race, and the salaries and acquisition costs of talent in large companies will continue to rise.
Regarding the direction of technology research and development
Shazel is deeply involved in sparse model and ultra large scale Transformer training. After joining OpenAI, he will focus on promoting the development of low-cost and high-efficiency giant models. The MoE architecture will become the core route of OpenAI's next-generation models; The industry will accelerate its diversion, with some talents staying at Google to deeply cultivate multimodal and scientific computing, while others will flock to OpenAI and Anthropic to specialize in the underlying architecture of general AGI.
Regarding the market and user end
In the short term, it will drive global model manufacturers to increase their investment in underlying technologies, leading to a new round of upgrades in model inference costs, context length, and multimodal capabilities; In the long run, it will accelerate the landing of general AI, accelerate the iteration of enterprise level large models and AI native applications, and ordinary users can obtain stronger and cheaper AI services. But the update pace of Google Gemini products may slow down, and the comprehensive advantages of GPT series products will further expand in the short term.
Shazel's job hopping is not a single talent mobility event, but a landmark signal of the redistribution of global AI underlying technology discourse power. In the intense stage of AGI competition, top architecture talents who master basic technologies such as Transformer and MoE have become the core strategic resources that determine the long-term competitiveness of enterprises. In the future, the talent attack and defense battle of top Silicon Valley enterprises will continue to escalate.
