Demis Hassabis on Gemini 3, world models, and the AI bubble (sources.news)

On the eve of Gemini 3’s release, Demis Hassabis portrayed Google DeepMind as calm and confident: early reports peg Gemini 3 Pro as the best all‑round model to date, and Google points to massive product traction — the Gemini app sees ~650 million monthly users, Search AI Overviews reaches over 2 billion people monthly, and roughly 13 million developers are integrating Gemini into their products. Hassabis’ remarks, posted alongside internal hype from leaders including Sundar Pichai, underscore Google’s aggressive push to close past gaps in capability and reliability with this generation. Technically and strategically, Hassabis signaled two priorities that matter to the AI community: advancing “world models” research and rigorously evaluating compute bottlenecks. World models suggest deeper scene- or environment-level reasoning beyond token prediction, while attention to compute constraints reflects engineering tradeoffs for scaling and deployment. His take on the AI “bubble” was pragmatic — whether there’s retrenchment or continued growth, Google aims to emerge stronger — implying sustained investment in core model quality, product integration, and developer tooling that could shape competition, safety practices, and real‑world adoption across the ecosystem.