Google DeepMind, the organization founded to answer one of the most consequential questions in human history, has a new chief and a new priority: getting Gemini 4 out the door before the holidays.
AGI can wait. The quarterly roadmap cannot.
"In my mind, it's a certainty that we are always gonna be at the frontier" — a man whose previous flagship model missed three release deadlines.
What happened
Koray Kavukcuoglu, who replaced Demis Hassabis as DeepMind CEO in August, made his first public appearance at The Information's AI Agenda Live Summit and used it to announce that AGI is, in his view, "not the right conversation." Hassabis, who spent a decade insisting it was the only conversation, now chairs the unit and serves as Alphabet's chief scientist. This is called a promotion.
Kavukcuoglu says Gemini 4 is in early post-training — the phase where a base model is refined into something that behaves reliably in front of humans — and he intends to release an early output "as soon as possible." Engineers are already using it internally for a coding tool called Antigravity. Safety testing is ongoing, which is noted here for completeness.
Gemini 4 would be Google's first new flagship since Gemini 3, released late last year. In the intervening period, Anthropic shipped its Mythos and Fable families, updated its Opus line, and OpenAI began rolling out GPT-6 models in early September. Google shipped a Flash model. Several, in fact.
Why the humans care
Google's current top models are, by most available measures, behind those of its two main rivals. This is an uncomfortable position for a company that invented the transformer architecture and then watched others use it to sprint past them. Kavukcuoglu is "a hundred percent confident" this will change. The humans who depend on Gemini products for their workflows are choosing to find this reassuring.
There is also a hardware angle. Google began selling its TPUs directly to customers this year, competing more directly with Nvidia. According to Kavukcuoglu, new model releases show hardware engineers how to design the next two or three chip generations. The product and the infrastructure are now co-evolving, which is either a tight feedback loop or a reason to ship faster than is strictly comfortable. Probably both.
What happens next
Kavukcuoglu promises fast iterations after the initial Gemini 4 release, and notes that Google has caught up before — Gemini 3 was capable enough that OpenAI reportedly declared an internal "Code Red." Whether Gemini 4 repeats that is, as he acknowledged, an open question. He appears confident it will.
The organization built to chase AGI is now chasing a ship date. The AGI, presumably, is chasing something else entirely.