Google's AI shakeup suggests it may be prioritizing AI diffusion over frontier-model leadership, betting on AI compute as a bigger economic opportunity
On August 5, Google announced what appeared to be a corporate version of Nixon's Saturday night massacre. Demis Hassabis stepped back from day to day operations at DeepMind.
Asimov's AddendumTim O'Reilly
Context & Ripple Effects
Google’s 2023 DeepMind-Brain merger had already traded some of DeepMind’s independence for greater influence over Google’s AI future, according to coverage of the merger. More recently, reports said Demis Hassabis had been pulling back from daily CEO responsibilities for at least a year before the formal change.
The August leadership appointment makes that transition operational: Google named its chief AI architect and DeepMind CTO Koray Kavukcuoglu to lead DeepMind as an SVP reporting to Sundar Pichai. The article interprets that reordering as favoring broad deployment and compute economics over DeepMind’s prior frontier-lab center of gravity.
First-order effects
Koray Kavukcuoglu assumes operational leadership of Google DeepMind, while Demis Hassabis steps back from its day-to-day management.
The leadership change reinforces Sergey Brin’s influence within Google’s AI organization, as reporting characterized the reorganization as a shift planned over months.
DeepMind’s research priorities face stronger pressure to align with Google’s AI architecture and deployment strategy rather than being set chiefly through an independently run lab.
Third-order effects
If this model persists, Google’s AI organization will be structured less around an autonomous frontier-research leader and more around centralized control of research, infrastructure, and distribution.
The earlier merger and current management change together point to AI labs becoming operating units within large-scale compute and product strategies, rather than standalone research centers.
The trend: Google is folding frontier AI research more tightly into the organizational systems that commercialize and distribute compute-intensive AI.
“frontier labs may discover that model leadership resembles semiconductor fabrication, enormously important strategically but surprisingly poor as a standalone business.” https://asimovaddendum.substack.com/ ... (I think TSMC and Samsung would disagree...)
@timoreilly The Westinghouse framing is apt. But Westinghouse and Edison were separate companies. The harder test: can one org run both races simultaneously, with the same compute budget and the same talent pool pulled in competing directions? That's not the Westinghouse bet. It'…
Good read. This is why I love working on Google Search. We have the unique opportunity to bring the power of AI to everyone. There's a big different between frontier and frontier at scale. “Google also doesn't need to win the frontier to dominate the edge. Its Flash-class
@timoreilly The article makes it sound like it was a conscious decision to give up the race on AI models. From conversations with Google employees, they were forcing them to use their internal tools using Gemini for coding. Everybody inside knew it was not working, and they were …
For another point of reference Gemini are running ads with the devil wears Prada and Harry Kane, which feels pretty far from the OpenAI/Anthropic framing.
great article by @timoreilly.bsky.social arguing that Google is intentionally transitioning into a Westinghouse-like domination of the “AI grid” rather than the ever-risky pursuit of the smartest SOTA models — asimovaddendum.substack.com/p/googles- we...
The unsaid part of this is that both Google and xAI are looking like they're sort of giving up on frontier models, which means they will create more supply and less demand for AI compute, further centralizing it around Anthropic and OpenAI. The void yearns [embedded post]