Google's DeepMind Retreat: A New Chapter in AI History

The End of an Era at DeepMind

Last week, the atmosphere at Google's Mountain View headquarters was thick with farewells. Employees lined up for one-on-ones with Jeff Dean and Quốc Lê, both departing to launch their own ventures. DeepMind staff, in particular, felt a chill in the air—uncertainty about their department's future and their own job security hung over every conversation.

Jeff Dean's new company, Discovery Loop, overlaps directly with DeepMind's work. All three other co-founders are senior Google employees. For many, these meetings were more than goodbyes; they were chances to secure a lifeline—maybe a transfer to another team through an old boss's recommendation, or even a priority interview slot at Discovery Loop.

From Frontier to Flash: A Strategic Shift

According to exclusive information obtained by APPSO, Google's DeepMind team will no longer chase the development of frontier models. Instead, they're focusing on the more cost-effective Flash-level models. This isn't just a tweak—it's a fundamental redirection. With the reorganization, layoffs could hit one-third of the team or more.

The team's size is around 7,000 to 8,000 people. While details like timing and scope remain uncertain, the goal is clear: cut redundant roles. For example, some employees hired for algorithmic positions never actually did algorithm work. Internal transfers are possible to avoid layoffs, but the writing is on the wall.

Just before this article's publication, Google publicly released Gemini 3.7 Flash, less than a month after 3.6 Flash. Yet there are no plans for a Pro update in the near term. Resources are being funneled into Flash models, signaling a major pivot in Google's AI business.

The Cost-Benefit of Not Being First

Google isn't abandoning research or pretending it doesn't care about falling behind in flagship models. The company has been a pioneer in AI: DistBelief, TensorFlow, Word2Vec, Transformer, BERT—these are foundational. Google Brain, DeepMind's predecessor, was crucial. Google will continue to push breakthroughs in key technologies.

But the chase after OpenAI, Anthropic, and Chinese open-source model makers has left Google exhausted. Endlessly betting on frontier models is no longer viable. As the saying goes, "It's not that Pro isn't worth training; Flash is simply better value."

Why Flash Wins Over Pro

Flash models are cheaper to evolve and show quicker results. They serve Google's core products—Search, Gmail, Android, YouTube, Maps—which handle billions of users daily. These products need fast, low-cost AI, not a massive, resource-hungry flagship model.

DeepMind has struggled to secure resources for big model training. In the last review cycle, the team's OKR score was just 0.5 out of 1.0. That's not a team poised to get more funding. The reality is that Google's core products consume as much AI compute as DeepMind does. Search uses proprietary models for intent understanding and low-latency responses on TPU clusters. YouTube relies on TPUs for recommendations, ads, and content moderation. Google Photos uses them for image recognition and semantic search.

A larger, more powerful model would be nice, but for a conglomerate like Google, it's not essential. What Google needs is a smart, cheap, fast model that can power its existing services. Flash fits the bill.

Not Competing for Top Three—and That's Okay

Internally tested Gemini 3.5 Pro already lags behind competitors like Meta's Muse Spark 1.1 on benchmarks. Gemini has slipped out of the top three in North America, trailing even Grok. Rumors suggest Google has stopped trying to compete for the top spot. That's likely true.

But this isn't a problem. Google's ecosystem is still solid. The company is repositioning itself not as a frontier model lab racing against OpenAI, but as an entity focused on applying AI to its core products. Search, under Jen Fitzpatrick, and Google Cloud, under Thomas Kurian, together generate 73% of Alphabet's revenue. They no longer need to be held back by DeepMind's ambitions.

The Human Element: Leadership Shuffles and Morale

Demis Hassabis, the long-time face of DeepMind, has stepped back from day-to-day management to become Alphabet's Chief Scientist and DeepMind's Chairman. His successor, Koray Kavukcuoglu, has less real power. The reporting structure now leads to Jen Fitzpatrick, whose influence has grown.

Meanwhile, Josh Woodward, VP of Google Labs, Gemini, and AI Studio, is rising. He's been a key presenter at recent I/O events. Despite the struggles of the underlying Gemini models, the Gemini App has thrived. CEO Sundar Pichai announced Dean's departure and the DeepMind reorganization, then immediately tweeted that Gemini App hit 1 billion monthly active users—600 million added since May last year. It's Google's fourteenth product to reach that milestone, and the fastest to do so. Pichai specifically thanked Woodward.

Fitzpatrick and Kurian are the defenders of the castle; Woodward represents the new wave of breakthrough operators.

Lessons from the Great AI Pivot

Google's internal shift mirrors a broader change in Silicon Valley's AI strategy. The industry has been obsessed with AGI and scaling laws, buying up compute and paying young researchers astronomical salaries, all while racing to release new benchmarks monthly. But the fervor is cooling. Even Google, a giant, has decided to draw a line under the vanity of chasing the frontier.

This is a historical moment—not just for Google, but for the entire AI landscape. The era of infinite resources for frontier models is ending. What replaces it is a more pragmatic, product-focused approach. For historians of technology, this pivot will be studied as the moment when the industry grew up, realizing that sustainable innovation often comes from serving real users, not just topping leaderboards.

The changes at DeepMind are a reminder that even the most advanced research labs must eventually answer to the bottom line. The shift from Pro to Flash isn't just a technical decision; it's a philosophical one. It says: we'd rather be useful than famous. And in the long run, that might be the most revolutionary choice of all.