August 5, 2026, Alphabet CEO Sundar Pichai sent an internal memo announcing the most drastic personnel and structural reorganization in the history of Google’s AI division: Demis Hassabis, who has led DeepMind for many years, stepped down as CEO to become Chairman and Chief Scientist at Alphabet, stepping away from daily operations to focus on AGI strategy; former CTO Koray Kavukcuoglu takes over the daily battle for Gemini model R&D and developer ecosystem; even more shocking is that Jeff Dean, employee No. 30 who has worked at Google for 27 years, along with three senior researchers, left to found Discovery Loop. The news caused Alphabet’s stock to drop about 4% in a single day. The capital markets voted with their feet, expressing concern over the loss of core technical talent. Behind this major shakeup, is this a full-scale acceleration of Google’s AGI strategy, or an organizational loss of control in daily model iteration under intense competition?
Jeff Dean, Former Google Chief Scientist, Co-Founder of Discovery Loop
The 4% Drop and a 27-Year Farewell
To understand the magnitude of this reshuffle, one must first see what the departing figures take with them. Jeff Dean worked at Google for 27 years. As employee No. 30, he participated in the core construction of everything from early search infrastructure (MapReduce) to deep learning frameworks and, finally, the Gemini multimodal models. In the developer community, he is viewed as the soul who “underpinned half of Google’s technology history.” The three co-founders who left with him to start Discovery Loop carry equally heavy weight: Sanjay Ghemawat is a founding father of Google’s distributed systems, Oriol Vinyals is a DeepMind Research VP, and Quoc Le is a co-founder of Google Brain.
The collective departure of these four senior researchers directly triggered a violent reaction in the capital markets. Alphabet’s stock price fell about 4% in a single day after the personnel changes were announced. For a tech giant of such immense scale, a 4% evaporation of market value in one day is no small number, reflecting the market’s deep concern over the loss of core technical backbone and the stability of short-term model iteration.
On developer communities like Hacker News, shock and regret became the dominant sentiment. Developers widely believe this is no ordinary succession; rather, it is the physical migration of “half the empire” of Google’s AI infrastructure and frontier research. The departure of Jeff Dean and the others takes away not just individual intellect, but also the engineering rapport and foundational architectural experience accumulated over many years inside Google. When the bedrock supporting Google’s twin engines of search and AI loosens, the market naturally questions whether Google can maintain its original engineering iteration tempo in the coming arms race for large models.
Hassabis’s Separation: Stepping Away from Daily Operations, Diving into the Deep End of AGI
As Jeff Dean and others left, another main thread of Google AI underwent a fundamental change. Demis Hassabis stepped down as CEO of Google DeepMind, transitioning to Chairman of DeepMind and Chief Scientist at Alphabet. According to the official statement, he will shed daily operations to focus on AGI strategy while continuing to serve as CEO of Isomorphic Labs (AI drug discovery).
Hassabis’s withdrawal is not marginalization, but a move based on Google’s judgment of the current stage of AI development. As the application-layer competition for large models in text generation and multimodal understanding increasingly turns into a red ocean, Google believes AGI has entered deep waters that require top scientists to tackle full-time. Daily model iteration, productization, and commercialization were consuming massive amounts of management energy. Freeing Hassabis from the administrative affairs of the CEO role and allowing him to concentrate full-time on fundamental AGI breakthroughs and Isomorphic Labs’ commercial offensive is Google doubling down on its long-term vision.
This adjustment appears particularly unique in the industry context. At a time when leading AI labs like OpenAI and Anthropic are mired in push-and-pull over safety alignment and speed limits — even triggering industry safety anxiety to the point where “1,134 AI employees jointly called for hitting the brakes” — Google chose to let its chief scientist bypass the usual safety speed-limit disputes and focus his energy on scientific discovery itself. Hassabis’s transition is, in essence, an organizational separation by Google between the “long-term AGI vision” and the “short-term commercialization of models.” Google is attempting to use an independent team to charge at fundamental scientific breakthroughs that require a cycle of over a decade to bear fruit, while handing the close-quarters market competition to a different set of people.
Koray’s Close Combat: Taking Over Daily Development and the Race Against Time for Commercialization
The person inheriting Hassabis’s daily management duties is former DeepMind CTO Koray Kavukcuoglu. Having worked at DeepMind for 13 years, he is a top expert in deep learning. After being promoted to SVP of Google DeepMind, he now reports directly to Sundar Pichai, taking full responsibility for Gemini model R&D, cutting-edge AI research, the Gemini App, and developer teams.
Koray is taking over an arena marked by extremely brutal competition. The current large model market is no longer Google’s exclusive domain. OpenAI’s GPT series holds a first-mover advantage in the developer ecosystem, Anthropic’s Claude is pressing hard in long-context and coding, and domestic newcomers like Moonshot AI are showing explosive power in specific dimensions. As a recent site article, “From 18th Place to the Top: How Kimi K3 Surpassed Claude and GPT in Long-Context Coding,” reveals, competition in vertical areas like coding and long contexts has reached a white-hot stage, making engineering iteration speed a decisive factor for survival.
The biggest challenge Koray faces after taking office is proving that Google is not falling behind under the pincer attack from rivals like GPT and Claude. He must accelerate the engineering implementation and developer ecosystem building of Gemini. This means DeepMind’s center of gravity will inevitably shift from “academic and frontier exploration” to “product and engineering close combat.” Some developers have voiced concerns: after Hassabis steps back, will DeepMind’s original academic and frontier exploration genes entirely give way to product and engineering genes, causing Gemini to stagnate in foundational architecture innovation? Whether Koray can maintain underlying model innovation while withstanding the race against time for commercialization will directly determine Google’s standing in the next generation of large model competition.
Jeff Dean’s External Plug-In: Veteran Exodus from the Tech Giant and Ecosystem Extension
Jeff Dean and his colleagues’ departure to start a new company is the most dramatic scene in this reorganization. The Discovery Loop they founded is positioned as a public benefit corporation (PBC), aiming to use AI to automate the scientific research process — including proposing, running, evaluating, and iterating thousands of concurrent experiments — and to explore “recursive self-improvement,” i.e., using AI to create more powerful AI.
Worth noting is the capital and compute relationship between Discovery Loop and Google. Alphabet is a founding investor in Discovery Loop and has committed to providing compute support for its first year. The initial funding round was led by Radical Ventures and Khosla Ventures. This arrangement suggests that the departure of Jeff Dean and his team is not simply “brain drain,” but rather a kind of “external plug-in” for Google’s ecosystem.
Inside a large company, constrained by compliance reviews, departmental silos, and short-term financial reporting pressure, ultra-frontier and potentially uncontrollable explorations like “recursive self-improvement” and “automated scientific methods” are extremely difficult to advance. Through investment and the provision of compute, Google is spinning this part of the business out, allowing it to develop independently free from the shackles of big-company bureaucracy, while simultaneously converting it into an ecosystem ally via the founding investment and compute supply. This is a high-order solution of “externalizing internal incubation.” Essentially, Google is using compute to bind the frontier demands of “AI for Science.” Once Discovery Loop achieves breakthroughs in automated scientific discovery, Google Cloud’s underlying compute ecosystem will directly benefit. The developer community expresses excitement about “automated scientific methods,” believing they can break the bottleneck of sequential human iteration, but also voices deep safety concerns about “recursive self-improvement” operating outside the human loop.
Organizational Tension and Final Verdict: Acceleration or Loss of Control?
Taken as a whole, Google’s high-level AI overhaul is neither purely an acceleration of AGI nor a loss of control in daily model iteration; it is a strategic choice aimed at competitive pressure. Organizational tension between long-term vision and short-term commercialization emerged within the tech giant, and Google chose to address it by using a “separation” to let Hassabis focus on the long term, an “external plug-in” to let Jeff Dean explore ultra-frontier areas, and leaving Koray to shoulder the short-term close combat.
Sundar Pichai’s internal memo framed this shake-up as an “amicable parting” and “trying new things.” This three-pronged architecture attempts to maintain the vitality of frontier exploration while strengthening the market competitiveness of commercial products. For developers and industry observers, the success of this architecture hinges on two core metrics: first, whether Koray can hold the line in close combat, ensuring the Gemini model family does not fall behind competitors in iteration speed and engineering capability; second, whether Discovery Loop can truly feed back into Google’s underlying compute ecosystem, transforming frontier exploration into commercial growth for Google Cloud.
Google’s restructuring this time is a proactive self-adaptation to the current AI competitive landscape. When large model competition shifts from comparing parameters to competing on engineering and ecosystem, physically separating the academic and engineering factions within the organization can actually help reduce internal friction. But the risks are equally clear: the loss of core veterans is hard to make up in the short term, and Koray faces extremely low fault tolerance under intense market pressure. Whether Google AI can achieve a soft landing in this strategic restructuring of carving out and spinning off, the model iteration data over the next year will provide the final answer.