Google: LLM Frontier Over AGI Dreams
Losing its top researchers looks like a crisis for Google. But it was probably a necessity, and it's bullish for Gemini.
From The Information:
Google announced major changes to its artificial intelligence operations, with Demis Hassabis stepping aside as Google DeepMind’s chief executive to become its chairman and longtime AI scientist and executive Jeff Dean leaving the company to co-found a startup.
CEO Sundar Pichai said Hassabis, who shared a Nobel Prize in 2024 for work on using AI for protein folding, will also become chief scientist of Google’s parent company, Alphabet. Koray Kavukcuoglu, who had been serving as Google DeepMind’s chief technology officer and Google’s Chief AI Architect, will lead Google DeepMind as senior vice president.
The changes are likely to strike a blow to morale at Google DeepMind, which has already seen a string of departures in recent weeks, and could complicate Google’s efforts to catch up in the race to develop the best AI models, particularly if more researchers leave as a result. Google has fallen behind in the AI model race, particularly in coding. Its next flagship model has been delayed as rivals Anthropic and OpenAI have released new models that are strong on coding and cybersecurity.
Google’s stock price dropped 4% on Wednesday’s announcement.
Earlier this summer, star AI researcher Noam Shazeer left Google to go to OpenAI. John Jumper, who received the Nobel Prize alongside Hassabis for their work on the protein-folding project, called Alphafold, recently left the company to join Anthropic.
This is a good summary of what happened, but I think the interpretation that this “could complicate Google’s efforts to catch up in the race to develop the best AI models” is wrong.
Logan Kilpatrick, a member of Google’s technical staff, posted the opposite on X last night:
Of course Kilpatrick, who has been a sort-of unofficial spokesperson for Gemini on X, would remain bullish. But in this case, I actually tend to agree with him. This shakeup seems bullish for Gemini in particular.
Researchers vs. Hackers
Many forget that Google actually presented LaMDA – its own LLM chatbot – at its May 2022 I/O conference1, six months before ChatGPT. It was already good enough to make one Google engineer believe that LaMDA was sentient. Nevertheless, Google didn’t publicly release an AI chatbot until OpenAI forced its hand2.
Why didn’t Google ship first? Gaurav Nemade, LaMDA’s founding product manager, told Alex Kantrowitz on the Big Technology Podcast that it came down to reputational risk. LLM answers were far below the traditional quality bar for Google Search. Executives were concerned about embarrassing or dangerous responses. They were also skeptical that users would be interested in a chatbot that often hallucinates. OpenAI, meanwhile, didn’t share those reservations. Another case of the Innovator’s Dilemma.
But that wasn’t the only reason Google took a while to catch up; another was a big cultural difference between OpenAI and Google’s DeepMind. A former DeepMind researcher described in a Tegus interview in early 2023:
DeepMind has historically recruited primarily elite researchers from academia [...] [who were] building their CV as a machine learning researcher. That often means a certain type of credentialing. [...] publishing pieces where you are the first or last author on that paper. [...] What that means is there’s a space for two people to get credit [...] That cultural makeup basically then incentivizes and flows down into small projects of two or three people on quite speculative research.
In stark contrast to DeepMind’s academic culture, the former researcher explained how OpenAI ended up with an organization of hackers and builders:
[...] OpenAI initially started off by trying [...] to mimic DeepMind. [It] hired a bunch of superstar researchers. And they all left within about a year.
[...] They’ve come up with a model [...] that is really focused on building engineering systems and products that look like AGI. And don’t really care about the credentialing system. [...] [They] hired a lot of people who are exceptionally smart, exceptionally good engineers, exceptionally good hackers. That hadn’t really published AI papers before.
They obviously had some researchers too. […] The contributions of that sort of skill set were as important to building this AGI system. […] Consequently, [they] did away with this concept of research scientists, which they originally had and now had this concept of member of technical staff. I think that is profoundly different to DeepMind.
[...] A large group of people is a far preferable thing to work at and culturally allows them to have a very significant amount of focus [...] they kind of said no, this is not an academic research organization, if you want that, go somewhere else. That’s very different.
How different? He moves on to explain:
There’s huge amounts of arguing about who gets the sort of first author or last author at Google or DeepMind within their research organizations. People would not work on projects, where those people, if they feel that they’re not going to get that credential. And that is encouraged by promotions and things where you will get promotions for being first author on a paper.
So the organizational incentives are set up to encourage them.
[...] I imagine it’s been quite a difficult cultural transition at OpenAI, but one that paid dividends. Now the value proposition to potential employees is pretty clear, which is that you get to be part of GPT-4 or GPT-5. And people are now, I think, willing to make that trade-off in order to work at OpenAI. And I think that has been a huge benefit for them even if it was probably pretty tough.
[...] Until just before I left DeepMind, the idea of 50 people [...] working on a project was unheard of, and it was only really stimulated, I think, by the large language model changes.
DeepMind functioned as an exploratory research lab, where small teams of two or three researchers pursued profound breakthroughs. OpenAI, meanwhile, had created an organization of exceptional hackers, building and scaling “systems that look like AGI.” When Large Language Models graduated from exploration to exploitation, OpenAI was ready to capture the moment.
ChatGPT forced DeepMind to change.
“No More Goofing Around”
Google, famously, declared an internal code red in response to ChatGPT. A great article on WIRED last year reported how the company scrambled to catch up with OpenAI, which included an organizational change to its AI teams:
The twin AI research labs that joined together to build Gemini, Google’s new language model, seemed to differ in their sensibilities. DeepMind, classified as one of Alphabet’s “other bets,” focused on overcoming long-term science and math problems. Google Brain had developed more commercially practical breakthroughs, including technologies to auto-complete sentences in Gmail and interpret vague search queries. Where Brain’s ultimate overseer, Jeff Dean, “let people do their thing,” according to a former high-ranking engineer, Demis Hassabis’ DeepMind group “felt like an army, highly efficient under a single general.” Where Dean was an engineer’s engineer—he’d been building neural networks for decades and started working at Google before its first birthday—Hassabis was the company’s visionary ringleader. He dreamed of one day using AI to cure diseases [...]
It was Hassabis who became the CEO of the new combined unit, Google DeepMind (GDM). Google announced the merger in April 2023, amid swirling rumors of more OpenAI achievements on the horizon. “Purpose was back,” says the former high-ranking engineer. “There was no goofing around.” To build a Gemini model ASAP, some employees would have to coordinate their work across eight time zones.
Concerns over being named first in an academic paper became a luxury Google researchers could no longer afford. And Jeff Dean could no longer let his engineers “goof around.” The open-ended exploration had ended, replaced with an intense cross-organizational focus:
Of the seven Google services with more than 2 billion monthly users, including Chrome, Gmail, and YouTube, all had begun offering features based on Gemini. Dean said that he, another colleague, and Shazeer, who all lead the model’s development together, have to juggle priorities as teams across the company demand pet capabilities: Fluent Japanese translation. Better coding skills. Improved video analysis to help Astra identify the sights of the world.
Two of the most accomplished researchers alive – Jeff Dean and Noam Shazeer3 – dedicated themselves to triaging a cross-organizational pool of feature requests. The engineer’s engineer who’d been building neural networks since before Google’s first birthday, and the researcher who had co-invented the Transformer, were spending their days stack ranking Japanese translation against better video analysis. But it was a code red, and that’s what Google needed at that moment.
This is worth highlighting, as things played out differently in other labs. Once they crossed from exploration to exploitation, they lost their explorers. During such a shift, OpenAI parted ways with its co-founder and chief scientist Ilya Sutskever4. Meta went through this last year, when FAIR wound down, replaced by Meta Superintelligence Labs. Enormous offers were made to people with hands-on LLM training experience, while Yann LeCun — who doesn’t think scaling LLMs leads to AGI — departed.
Google held out longer. The explorers went on to build and scale Gemini. But you can only delay the inevitable for so long.
Another Code Red?
I’ve experienced a code red or two while I worked at Google. Once the alarm goes off, anyone who can help drops what they’re doing and joins the war room. No one tells you to work nights and weekends; it’s just what you naturally do when your team is in a tough spot. As exhilarating as it can get at first – no org charts, no bureaucracy, just a group of talented people working together on a tough problem – it is also exhausting. Especially when it lasts for weeks, or even months. At some point, though, it always ends. The goal is achieved. And the victorious troops return home from the battle. Showered with spot bonuses, they can lean back for a while.
For a brief moment in 2025, it seemed like Google DeepMind finally got there: In May, CEO Sundar Pichai claimed that the Gemini 2.5 family of models were dominating the Pareto frontier. Gemini 3 came out in November, leaping past OpenAI, which then declared a code red of its own. The tables had turned.
But not for long. Throughout 2026, a series of models from OpenAI and Anthropic kept pushing the frontier further and further away. While recent Gemini models are notably cheap and fast, they fall behind on performance benchmarks. The widest gap is around code-writing and long-running agentic workflows. A key capability for building self-improving models.
Which calls for another code red. Or maybe the first code red never really ended.
At some point, though, a company has to recognize that this isn’t about winning one specific battle. The AI race is turning out to be a long-running war, and it requires the right soldiers. It could no longer be an academic research organization; it needed exceptional hackers, who would truly enjoy working in large teams on improving model capabilities.
Relying on world-class researchers to do this job, however, could lead to extreme situations – from a recent Financial Times article titled Google DeepMind dismantles Nobel-winning AlphaFold team in strategy shift:
Earlier this year, Jumper and fellow AlphaFold researcher Jonas Adler moved to Google’s internal “Code Strike” team, a group assembled to improve the company’s AI coding capabilities as it seeks to catch Anthropic and OpenAI.
Last month, Jumper announced he was leaving for Anthropic, where he will be joined by Adler and AlphaFold colleague Alexander Pritzel.
Nobel Prize-winning researchers aren’t the best fit for debugging a model’s coding performance. DeepMind was able to get its researchers to do it once, under emergency conditions, but they can’t be expected to spend their careers on benchmark leaderboards.
Singularity Is Not Near Enough
Last May, after Google failed to announce a frontier model in its last I/O conference, Ben Thompson wrote that it might have to do with Hassabis’ focus on building World Models, which have a much bigger long-term potential compared to LLMs:
I think it’s possible that the reason Google is widely considered to be behind both Anthropic and OpenAI in terms of coding, particularly long-running agentic workflows that depend just as much on the harness as the model itself, simply comes down to their research team having other priorities. That’s why the coding parts of this keynote fell on the Antigravity team, not DeepMind, and why Hassabis was barely on stage.
Still, he did have the closing, and it’s clear he [Hassabis] thinks DeepMind’s approach is working, at least in terms of AGI [...] This — particularly given that it came on the heels of explaining how Google was working to end all diseases — is certainly inspiring; it also has nothing to do with coding or video generation or Search. That’s fine, as far as it goes; I just suspect that if the company’s AI leader doesn’t care about such prosaic matters, such prosaic matters may not, in the long run, go that far. And, if Hassabis is right and Google gets to AGI first, it may not matter.
Google, however, has yet to get to AGI. And Google’s announcement from yesterday makes it seem like “prosaic matters” such as LLM coding do start to matter to CEO Sundar Pichai. I found it to be quite telling, beginning with the title, The next chapter of our AI momentum, having no mention of AGI; Pichai wrote:
[...] We have to accelerate all this work and stay focused on the AI frontier. At the same time, there’s never been a more important moment to shape the future of AGI and science. Today Demis, Koray and I are sharing a few changes to our Google DeepMind teams that will enable us to do both.
AGI and science: Demis has described us as standing in the foothills of the singularity, and has been spending a lot of his time engaging externally. He and I have been long discussing a role that allows him to put his full attention on actively shaping the future of AGI. [...] Demis will become the Chair of GDM and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs. [...]
Google DeepMind: We are building strong momentum: Flash is in high demand, our Cyber model is live, and Gemma models have surpassed 900M+ downloads. We are committed to being at the frontier, and are super focused on the areas where we need to improve. I’m really excited for our upcoming model releases and the progress we’re seeing. [...]
My reading is that Pichai isn’t willing to bet the company on reaching singularity. Hassabis still believes the singularity is near, and so that remains his assignment. But everyone else at Google DeepMind – the company Hassabis had originally founded to pursue AGI – is shifting their focus to the prosaic matter of getting Gemini to the AI frontier.
This long and painful transition – from a scattered portfolio of academic explorers, to an LLM-scaling organization of hackers and builders – is now official. While it probably decreases the likelihood of Google ending all diseases, it is indeed bullish for Gemini.
It’s obviously not the happiest moment in Google’s history, seeing a hero like Jeff Dean sending his farewell email. But that’s the price of being a for-profit business, not a scientific research institution. As blurry as these lines often get at Google.
“Organizing information is clearly a trillion-dollar opportunity, but a trillion dollars is not cool anymore. What’s cool is a quadrillion dollars,” Noam Shazeer5 said last year, when he and Jeff Dean appeared on Dwarkesh Patel’s podcast. I think Pichai made the right call, though, choosing to settle for the trillion dollars.
For new posts to land directly in your inbox:
Disclaimer: Long GOOG. Not financial advice. This post is for educational and general purposes only and should not be relied upon for investment decisions.
And that was LaMDA 2; an earlier version of LaMDA was presented in 2021.
Following a code red, Google launched Bard in March 2023, nearly four months after ChatGPT.
Shazeer left Google in 2021 to start Character.ai, which was acquired by Google. He returned to co-lead Gemini, and recently left Google again to join OpenAI.
This followed OpenAI’s 2023 board crisis, which can be viewed as part of the process where a research lab transforms into a profit seeking product company.
Shazeer has recently left Google and joined OpenAI, where he may or may not be pursuing a quadrillion-dollars opportunity.
Great post. Thank you Matan. Interesting to see where they go from here.