On August 5, 2026, Google’s CEO, Sundar Pichai, “reorganized” one of the most important AI rooms on earth. Demis Hassabis stepped off day-to-day from DeepMind to become “Chair of Google DeepMind and Chief Scientist” of Alphabet, and at the same time Koray Kavukcuoglu took the helm of Deepmind. By itself, maybe a non-event on the balance.
But then, in the same week, Jeff Dean (CEO of Deepmind) announced he is leaving after 27 years along with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le (the people who literally built the deep neural stack) to found Discovery Loop. Seed money from Radical, Khosla, Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet itself. Alphabet stays as investor and cloud partner.
Combined, that’s not just a PR headache, it’s a massive hole in the layer that turns raw research into systems that actually work. To be brutally honest, the news last week on the AI (and particularly the AGI) richter scale was the equivalent of an 8.0.
The decision
Pichai’s memo tried to frame it as momentum reminding us that Google’s AI Gemini app is surpassing 950 million monthly users - yada yada. As stated, Hassabis is holding the “destiny” seat and will work with Pichai directly on “strategic and global AGI matters,” along with advising DeepMind leads from London, and lean harder into Isomorphic Labs, Alphabet’s drug-discovery bet, the place he has long said AI’s number-one application should be human health. This is the visionary cognition speaking in full voice.
Koray Kavukcuoglu gets the operating seat. Thirteen years inside DeepMind. Former CTO of GDM and Google’s Chief AI Architect. Builder of the deep learning team. As SVP of Google DeepMind he reports to Pichai now and owns Gemini model development, frontier research, the Gemini app, and developer platforms. That is operating cognition layer.
Compounding all this churn is David Silver who left in February. And Noam Shazeer, Vinyals’s fellow Gemini lead, left for OpenAI in June.
The market disliked all the compounding tradewinds of the last six months and Alphabet stock slid big on the news.
The team’s blended cognitive composition
Strip the room to vectors and you see why August 5 had to happen.
Hassabis is singular-mission (visionary) cognition. Chess prodigy. Neuroscience. DeepMind founder. AlphaGo, AlphaFold, the public face of AGI-as-civilizational-project. His own words locate the center of gravity outside the quarterly model race: singularity foothills, health as the proving ground, “time and space to focus on the big picture.” That mind is indispensable for shaping what AGI becomes.
Koray is the counter-balance: deep-technical (operator) cognition who has lived inside the same lab for thirteen years and knows how to turn research into systems that ship. Pichai did not hire a brand-name outsider. He promoted the person already holding the builder keys.
Dean is (institutional-builder) cognition, the rare engineer who compounds infrastructure across decades until the company runs on his abstractions. When that vector exits with three other multi-decade collaborators, you do not have a PR problem. You have a composition hole in the layer that turns research ambition into durable systems.
Pichai’s job in this room is portfolio cognition: keep the full stack (infra, cloud, models, apps) pointed at one company strategy while the people who built each layer renegotiate their seats. The prior design asked one human (Hassabis) to hold destiny and delivery at once, while the builder layer was already voting with its feet. Dual-hat visionary-CEO plus quiet attrition is not a stable executive graph.
What a Decision Simulation would have surfaced
Load the pre-August room as cognitive vectors and the stress-test is not “is Gemini good?” The stress-test is: “What happens when AGI-destiny cognition and product-race cognition share one operating seat, your coding models lag the frontier competitors, a Gemini co-lead has already left for OpenAI, AlphaGo’s research lead is gone, and four multi-decade builders are one conversation away from a term sheet and their own cap table?”
A Decision Simulation flags the collision many months earlier (at Silver’s exit, at Shazeer’s exit) as a role-design and retention problem. It asks who in the room is scored on keeping the builder layer inside the building, who is scored on external AGI narrative, and whether those scorecards can live in the same skull at full intensity. Google answered in August by splitting the seats. That is the right move. It is also late relative to the attrition signal.
The lesson
The biggest lesson here is that the human layer is the point. Every company racing to “embrace AI” is copying model vendors and skipping the thing Google just proved still decides outcomes: which human cognitions sit in which chairs when the irreversible calls get made.
AI does not remove executive composition risk. It raises the cost of getting composition wrong. Google did not fail here because it lacked the best models, or competent users, or sufficient compute. It hit a wall because destiny cognition and delivery cognition were fused with duct tape and bailing wire.
The companies that win the next decade will not be the ones with the loudest AGI narrative. They will be the ones who simulate the human layer in the AI. As boards pour capital into the AI stack, the ones using AI to stress-test who is missing from the decision table (before the exit wave) will be making the intelligent decision possible.
The takeaway? AI isn’t going to save you from bad team composition; it actually makes it more expensive to get it wrong. The companies that will win in the next decade are the ones who understand how to put the right people in the right seats before the unconscious cognitive forces force them to.