Father of the Transformer Leaves Again: Noam Shazeer Departs Google to Join OpenAI
Noam Shazeer, co-lead of Gemini and author of the Transformer paper, announced his departure from Google to join OpenAI. This is the second time in his career he has bid farewell to the company he has served since 2000, and it has been less than two years since Google “bought him back” for $2.5 billion.
One of the Eight Transformer Authors Leaves Google Again
Noam Shazeer has once again left Google.
This Tuesday, the Gemini co-lead and Vice President of Engineering posted a short message on X, announcing that he would be joining OpenAI. He didn’t mention his exact position, only writing, "This was a tough decision." But industry insiders understand the weight of this sub-200-character announcement—it means that the last core member among the eight authors of the Transformer paper still holding a "main position" at Google is now gone.
And it’s been less than two years since Google spent about $2.5 billion to bring him back from Character.AI.
From Chief Engineer to “Defector” to Gemini No. 1
To understand the impact of this departure, we need to look back at Shazeer’s trajectory at Google.
In late 2000, shortly after graduating, Shazeer joined Google to work on the early advertising system. That was before Google went public, and he was one of the company’s earliest core engineers, later becoming a Principal Software Engineer. Internally, he long played the role of “the one who understands best how to write code for models.” When the Transformer arrived in 2017 with Attention is All You Need, Shazeer was the one who rewrote the entire project’s code and pushed performance to new heights. That was enough to secure him a place in AI history.
At the end of 2021, Shazeer left for the first time. The reason was openly discussed—he couldn’t stand Google’s bureaucracy. He and fellow LaMDA teammate Daniel De Freitas left to found Character.AI, building anthropomorphic conversational products that at one point reached over 20 million monthly active users and a $1 billion valuation.
Then came the “reverse acquisition” in August 2024. Through a $2.5 billion non-exclusive technology licensing deal, Google brought Character.AI’s core team—including the two founders—back into the fold. Shazeer didn’t join DeepMind or Google Research, but was instead appointed co-technical lead for Gemini alongside Jeff Dean and Oriol Vinyals. In other words, upon returning, he was immediately given Gemini’s top spot—a virtually unprecedented arrangement in Google’s internal structure.
And now, less than two years later, he’s leaving again.
Why Now? The Overlap of Three Signals
Shazeer hasn’t given a reason, but looking at the timeline, a few overlapping signals make the picture clearer.
First, Gemini has shifted gears markedly over the past six months. After the release of the Gemini 3 series, the team’s structure went through a round of adjustment. A long-standing question in the community has been: How long can the “multiple heads” setup—Hassabis, Jeff Dean, Vinyals, and Shazeer—endure? The answer appears to be: not long. Over the past year, the balance of AI power has tilted noticeably toward the DeepMind side, narrowing the freedom available to “product + architecture” hybrids like Shazeer.
Second, OpenAI’s hiring pace has accelerated. Since the second half of last year, OpenAI has been poaching core researchers from Google alumni circles, such as Amelia Glaese. The goal is clear: fill in top-tier architects for multimodal, long-context, and reinforcement learning directions. Shazeer is one of the few with both the ability to propose new architectures and to scale them to production level, making him naturally attractive to OpenAI.
Third, his personal style doesn’t fit big corporations. He made this clear with his 2021 departure. At Character.AI, he was still hands-on writing code and building infrastructure. Projects like Gemini, with thousands collaborating, require a decision-making chain with too many layers—likely draining for him.
With these three factors combined, it was only a matter of time before he left.
What This Means for Google
In the short term, Gemini won’t collapse because one person leaves. Jeff Dean and Oriol Vinyals remain, and the Gemini team’s engineering capabilities are still industry-leading.
But in the mid-to-long term, Google is losing more than just a VP of Engineering.
- Loss of architectural influence. Shazeer is the kind of person who can “write an Attention variant from scratch.” In the past two years, Gemini’s key decisions in MoE, long-context, and sparsification have all involved him. Put another way, with him gone, Google has one less person capable of going head-to-head with OpenAI or Anthropic at the architecture level.
- Talent signal issues. In the last year, quite a few researchers have left Google for OpenAI and Anthropic. The departure of someone of Shazeer’s caliber reinforces the narrative that “OpenAI is where researchers have the most freedom”—a narrative that, once it spreads in the Bay Area, tangibly impacts recruitment.
- A $2.5 billion returnee leaving again in under two years. This isn’t about the money; it points to structural problems in Google’s retention of top talent. Even at the ceiling for non-competes, equity, and position, they still couldn’t keep him.
Reportedly, the internal response at Google has been calm. A person close to the Gemini team said, “No one’s surprised—just that it happened a bit sooner than expected.”
The Move for OpenAI
OpenAI hasn’t disclosed Shazeer’s exact role. Based on his skill profile, it’s likely a senior research-side position, possibly reporting directly to Greg Brockman or Mark Chen.
Over the past six months, OpenAI has done two main things: pushed out successors to GPT-5 and iterated the o series reasoning models, and restarted some lower-level architecture exploration projects internally. The latter is exactly Shazeer’s area of expertise.
Another noteworthy point: Shazeer’s “model-as-product” experience from Character.AI is precisely what OpenAI lacks in ChatGPT’s current product trajectory. ChatGPT’s engineering is strong, but when it comes to “making the model itself an addictive product,” Character.AI’s methodology is not something OpenAI has mastered. If Shazeer straddles both research and product at OpenAI, his impact could be much larger than if he did research alone.
Of course, there’s uncertainty. Shazeer is a “researcher-turned-entrepreneur” type. Whether he’ll feel constrained again in a 3,000-employee OpenAI is unknown. Two years ago, he returned to Google from Character.AI; two years later, he’s moved to OpenAI—the possibility of him leaving to start something new again can’t be ruled out.
Overlooked Details
A few noteworthy details the media hasn’t focused on:
- This departure didn’t involve the usual “months-long transition period.” Judging by the announcement timing, the move was made cleanly.
- Character.AI’s other co-founder, Daniel De Freitas, hasn’t publicly announced leaving, but multiple sources say his role in Gemini’s multimodal group has been diminishing since the start of the year.
- The licensing deal between Google and Shazeer in 2024 was non-exclusive, meaning Character.AI’s technology could theoretically be licensed to other companies. At the time, this was seen as an antitrust safeguard, but in hindsight, it lowered Shazeer’s “binding cost.”
Taken together, these suggest this was a prepared, not impulsive, exit.
A Subtle Industry Shift
Zooming out a bit: from 2023 to now, the migration paths of top AI researchers look roughly like this:
- Google → OpenAI / Anthropic: Mostly one-way, with Shazeer as the latest high-profile case.
- OpenAI → Anthropic / Startups: Both directions, but mostly towards startups.
- Meta → ?: Meta’s LLaMA team has split several times, with some core members heading to startups.
The key takeaway from this migration map is: the center of architectural innovation is shifting further away from Google. Google still has the strongest compute, infrastructure, and most complete data loop, but when it comes to “who defines the next-generation architecture,” its influence is loosening.
The next-gen model race after Gemini 3 will be even tighter due to this personnel change. With Shazeer on board, OpenAI is likely to make new architectural moves by year-end.
A Tangent
For developers, the fierce head-to-head competition between the Gemini, GPT, and Claude series is actually a good thing—the harder the top fights, the faster model capabilities improve, and the more prices drop. OpenAI Hub already connects to major models; with one API key you can hit GPT, Claude, Gemini, and DeepSeek, direct-connect in China, OpenAI format-compatible. If OpenAI releases a new model this year due to Shazeer’s joining, it will be online immediately.
As for Shazeer himself, the engineer who’s been running down the AI path for 25 years ended his X post with a much-shared line: “Onward.”
This co-author of the Transformer is unlikely to stop. The only question is whether the next time he makes headlines will be for delivering a new architecture at OpenAI—or for yet another departure.
References
- Gemini Co-lead, VP of Engineering Noam Shazeer leaves Google for OpenAI - linux.do: A linux.do community discussion thread on the departure, including screenshots of the original post and early reactions.
