Google's AI research lab was founded on a single, almost quasi-religious idea: build artificial general intelligence, a machine that can think and learn like a person across any task. That mission is now taking a back seat to something much simpler: shipping the next model on time.
Koray Kavukcuoglu took over daily leadership of Google DeepMind in August, after founder Demis Hassabis moved into a chairman and chief scientist role. He wants Gemini 4 released well before the end of the year.
The model is already running inside Google's internal coding tool, Antigravity, going through a final polishing stage before public release. When asked directly whether Google has achieved AGI, Kavukcuoglu waved the question off. He said the more useful conversation is about building agents that businesses can actually trust to do real work.
That framing matters. For three years, the AI industry has argued over a finish line that nobody agrees on the definition of. Kavukcuoglu is effectively saying: stop debating the finish line, start shipping things people can use.
There is a less flattering reason for the urgency, too. Google announced a model called Gemini 3.5 Pro back in May, promised it for June, then delayed it to July. It then scrapped the underlying model entirely and restarted training from the beginning.
It still has not shipped. Google quietly released smaller updates around it instead, but the flagship version simply disappeared from the roadmap. Jumping straight to Gemini 4 lets Google skip past an unfinished product rather than explain why it failed.
Behind the scenes, DeepMind is also losing talent. An industry analysis found the lab's ratio of new hires to departures among top researchers fell from twelve to one in 2023 down to roughly two to one now. Names leaving include Jeff Dean, one of Google's most senior engineers for two decades, and John Jumper, a Nobel Prize winner for his AI protein-folding work.
Anthropic has become the single biggest destination for departing DeepMind staff, with OpenAI close behind. None of this means Google is losing the AI race, since it still owns the biggest distribution channel on earth, built into Search, Android, and Workspace.
But it does mean the era of DeepMind as a slow, prestige research lab chasing a grand scientific mission is over. It is now a product company competing on release speed against OpenAI and Anthropic, exactly like everyone else.
For any business relying on Google's AI tools, the lesson is simple. Gemini 3.5 Pro was promised, delayed twice, and quietly buried, so do not get attached to a specific model version or a public roadmap date. Build your workflows so they can tolerate a model being renamed, swapped, or pulled with little warning, because the company shipping it is now optimizing for speed over predictability.