DeepMind's New Chief Prioritizes Gemini 4 Ship Date Over AGI Quest

Google DeepMind's new chief, Koray Kavukcuoglu, wants Gemini 4 released "much earlier" than the end of the year. The model has already entered post-training. It runs internally in Google's coding tool Antigravity. Those facts alone would make news in any week, but the framing around them signals something larger: the research lab Demis Hassabis built to chase artificial general intelligence has, under new leadership, become a product organization with a shipping schedule.

Kavukcuoglu has drawn a clear line between his priorities and those of his predecessor. He calls the AGI question that drove Hassabis "not the right conversation." In its place, he puts a different goal: trustworthy AI agents. The distinction matters. AGI is an open-ended research target with no agreed definition and no deadline. Trustworthy agents are an engineering problem with measurable failure modes, deployable features, and customers waiting for them. A lab organized around the second goal behaves very differently from one organized around the first.

What we know about Gemini 4

The concrete details are sparse but specific. Gemini 4 is in post-training — the phase after the base model has been trained, when teams refine behavior, alignment, and tool use through techniques like reinforcement learning and fine-tuning. Entering post-training typically means the large-scale pretraining runs are done and the clock has started on final calibration and safety evaluation.

The model is not a paper artifact. It already runs inside Antigravity, the coding tool Google has positioned against rivals such as Cursor and Claude Code in the increasingly crowded AI-assisted programming market. Internal dogfooding in a production-shaped environment is the last station before external testing, and it suggests the Gemini 4 effort is far enough along that Google is stress-testing it where it counts: real developer workflows.

Kavukcuoglu's stated ambition to ship "much earlier" than year-end reads as a direct response to the cadence war. OpenAI, Anthropic, and Google have spent the past eighteen months leapfrogging each other on benchmark releases, and Google's Gemini 3.x cycle has kept the company competitive on paper. A compressed Gemini 4 timeline would be a statement that DeepMind intends to compete on shipping speed, not just research output.

The AGI question, set aside

The sharper signal in Kavukcuoglu's comments is what he declined to endorse. Hassabis built DeepMind — first as an independent lab, then as the merged Google DeepMind — around the explicit premise that AGI was the destination and safety-conscious research the route. Kavukcuoglu's dismissal of that framing as "not the right conversation" is not a nuance. It is a redefinition of the organization's purpose.

His alternative priority, trustworthy agents, aligns with where the commercial stakes now sit. Agents that can take multi-step actions on a user's behalf — writing code, booking, purchasing, orchestrating software — are the central product battleground of the current AI cycle. Enterprises are buying reliability, not metaphysics. A chief who says trustworthiness "matters more" than AGI is making a market argument, and Google's revenue mix makes that argument easy to defend internally.

The quiet disappearance of Gemini 3.5 Pro

The Gemini 4 push lands against a backdrop of turbulence. Gemini 3.5 Pro quietly disappeared. No decommissioning narrative, no victory-lap blog post — it vanished, and its absence now reads as a deliberate clearing of the deck ahead of the next major release. In a healthy product lineup, a mid-cycle model usually persists as a cheaper tier. Its removal suggests Google wanted no ambiguity about which frontier model developers should target next.

The departures tell the other half of the story. Top researchers have left for OpenAI and Anthropic, the two competitors most closely identified with frontier research prestige. Attrition at that level is rarely about salary alone; it is about what people want to work on. Researchers drawn to DeepMind by its AGI mission have watched the lab's center of gravity shift toward product delivery, and some have concluded their ambitions fit better elsewhere. Kavukcuoglu's own rhetoric — AGI as "not the right conversation" — will not slow that particular leak, though a faster shipping cadence may attract a different kind of engineer who prefers it.

From mission lab to product shop, for good

The Decoder's assessment is blunt: "the research lab with an AGI mission has turned into a product shop for good." The phrase is harsh but defensible. Google created Google DeepMind in 2023 by merging DeepMind with Google Brain, an integration widely understood as an effort to shorten the distance between research breakthroughs and shippable Google products. Under Hassabis, the lab preserved a dual identity — Nobel-recognized fundamental science alongside Gemini releases. Under Kavukcuoglu, the identity is singular.

This is not necessarily a loss for Google. Product-grade focus has arguably produced the most usable things in the Gemini line, and Kavukcuoglu inherits an organization with unmatched compute, distribution through Google's surfaces, and a coding tool — Antigravity — that gives Gemini models an immediate proving ground. If Gemini 4 ships early and performs, the product-shop framing will look like discipline rather than decline.

The stakes extend past Google. If the lab that most explicitly claimed the AGI mantle no longer treats AGI as its organizing conversation, the industry's frontier narrative loses one of its two or three canonical anchors. OpenAI and Anthropic still frame their work in AGI terms. Google's flagship AI organization now frames its work in release terms. That divergence — mission versus product, research horizon versus shipping quarter — will shape where researchers, capital, and enterprise trust flow next, and the pace of Gemini 4's arrival will be the first hard evidence of whether the new posture pays off.