DeepMind's Succession and the AGI Debate: How Google Is Restructuring for the Long Game

DeepMind's Succession and the AGI Debate: How Google Is Restructuring for the Long Game

*By Lukas Hoffmann, Europe & Frontier Correspondent — Berlin, September 20, 2026*

The week of September 15–20 produced no single model release from Google DeepMind that rewrote the benchmark tables. What it produced instead was something arguably more consequential: a deliberate restructuring of how the world's most resource-rich AI laboratory intends to govern itself, communicate with the outside world, and position its founder for the transition to whatever comes after the current generation of large language models.

Two announcements, separated by three days, define the moment. On September 16, Google DeepMind launched the DeepMind Institute (DMI), a think tank and publication platform dedicated to the societal, economic, and safety implications of artificial general intelligence. On September 19, Sundar Pichai announced that Demis Hassabis would step back from the CEO role at Google DeepMind to become Chair of the division and Chief Scientist of Alphabet, while Koray Kavukcuoglu — a 13-year veteran who previously served as Chief Technology Officer and Chief AI Architect — would assume the role of Senior Vice President, taking operational control of model development, research, and the Gemini product ecosystem.

Taken together, the moves signal that Google is no longer treating the frontier AI race as a sprint in which the same person can simultaneously run the lab, manage the product, and represent the company's long-term scientific vision. It is, in effect, a division of labour that acknowledges the scale of what is now at stake.

The Succession: What Changes and What Doesn't

Hassabis's transition is not a departure. He remains at the centre of Google's AGI strategy, retaining his role at Isomorphic Labs, the drug-discovery subsidiary that grew out of AlphaFold's success. His new title — Chair of Google DeepMind and Chief Scientist of Alphabet — positions him as the long-horizon thinker, freed from the quarterly cadence of model releases and product reviews that increasingly define the competitive landscape.

Kavukcuoglu's promotion is a statement of technical continuity. His fingerprints are on some of DeepMind's most durable contributions: WaveNet, the neural text-to-speech architecture that reshaped voice synthesis, and Deep Q-Networks (DQN), the reinforcement-learning system that first demonstrated superhuman performance on Atari games and laid conceptual groundwork for much of what followed. He now reports directly to Pichai and oversees the full Gemini model family, the Gemini application, and the developer ecosystem built around it.

The departure of Jeff Dean — who spent 27 years at Google and co-authored foundational papers on distributed systems, neural architecture search, and large-scale machine learning — adds a further layer of significance. Dean is leaving to launch an independent public benefit corporation focused on machine learning and scientific discovery, with Google serving as a founding investor and cloud partner. His co-founder is Sanjay Ghemawat, another Google veteran. The arrangement is unusual: a major technology company funding a spin-out that will operate independently, presumably competing for talent and research attention in the same domains. It suggests that Google sees value in maintaining proximity to Dean's work even outside a direct employment relationship.

"The next chapter of AI momentum requires us to separate the work of building the frontier from the work of understanding what the frontier means," Pichai wrote in his announcement. "Demis is uniquely positioned to do the latter."

The market's initial reaction was sceptical. Alphabet's stock dipped on the news, reflecting investor uncertainty about whether a leadership transition at this moment — with competitors releasing capable models at an accelerating pace — introduces unnecessary risk. That reaction is understandable but probably misreads the situation. Kavukcuoglu has been running significant portions of DeepMind's technical operations for years. The change in title formalises an existing division of responsibility more than it creates a new one.

The DeepMind Institute: Governance as Strategy

The DeepMind Institute↗, launched three days before the leadership announcement, is harder to categorise. It is simultaneously a research publication platform, a policy advocacy vehicle, and a reputational positioning exercise — and it is worth being precise about which of those functions it serves in each of its stated activities.

The institute is led by Hassabis, James Manyika, and Shane Legg, with Legg serving as managing editor. Its inaugural collection of essays covers four domains: reasoning transparency, economic policy, governance frameworks, and the definition of AGI itself. The essays are explicitly attributed to their individual authors and are described as separate from official Google policy — a structurally important distinction that gives the institute room to publish views that the company might not want to endorse corporately.

The inaugural essays address a range of concrete questions:

• Reasoning transparency: Rohin Shah and Anca Dragan argue that declining interpretability in advanced models is not inevitable, and propose that developers treat monitoring of sequential computation as a design requirement rather than a post-hoc aspiration.
• Economic policy: Researchers used 51 AI agent raters modelled on data from 51 economists to evaluate 11 policy interventions for AGI-driven labour market disruption, from expanded unemployment insurance to Universal Basic Capital.
• Governance: Hassabis proposes a U.S.-led frontier AI standards body with voluntary model submission at least 30 days before release, with the possibility of transitioning to mandatory assessments if safety risks warrant it.
• AGI definition: The institute’s directors maintain that current AI systems lack the consistency and creativity required for full AGI, but that the gap is expected to close in the near future — a position that carries significant implications for how urgently governance frameworks need to be in place.

The most technically substantive contribution at launch came from Rohin Shah and Anca Dragan, who argued that declining transparency in advanced AI models is not an inevitable consequence of scale. Their proposal centres on the ability to monitor sequential computation and what they call "opaque serial depth" — the degree to which a model's reasoning steps are inaccessible to external inspection. The argument is that developers and regulators should treat interpretability as a design requirement rather than a post-hoc aspiration.

The governance essay, attributed to Hassabis himself, proposes a U.S.-led frontier AI standards body. The mechanism is a voluntary submission system in which developers would submit models for independent evaluation at least 30 days prior to release, with the possibility of transitioning to mandatory assessments and coordinated development slowdowns if safety risks warranted such measures. The proposal is notable for what it concedes: that voluntary frameworks may be insufficient and that some form of binding coordination could become necessary.

"AI capability development must not outpace safety controls," Legg said at the institute's launch, adding that he personally assigns a 50% probability to the arrival of "minimal" AGI by 2028 — a personal forecast, he was careful to note, not a corporate timeline.

The economic policy contribution is methodologically interesting. Researchers used 51 AI agent raters modelled on data from 51 economists to evaluate 11 policy interventions for AGI-driven labour market disruption, ranging from expanded unemployment insurance and earned income tax credits to Universal Basic Capital. The use of AI agents as proxies for expert opinion is itself a methodological choice that deserves scrutiny — it is not obvious that a model trained on economists' published views will accurately represent the distribution of their considered judgements on novel scenarios — but the exercise at least attempts to ground policy analysis in something more structured than intuition.

What the Institute Is Not

It is worth being direct about the limits of the DeepMind Institute as a governance mechanism. It is a publication platform operated by one of the companies whose models it is, in part, discussing. Its essays represent individual authors' views, not binding commitments. Its governance proposals are voluntary. None of this makes the institute worthless — high-quality public analysis of AGI risks and policy options is genuinely scarce, and the institute's authors have access to technical knowledge that most external commentators lack. But the structural independence that would make its safety assessments most credible — the kind that the EU AI Act's third-party audit requirements are designed to create — is not present here.

The contrast with the EU's approach is instructive. The EU AI Act↗, which reached general application on August 2, 2026, establishes mandatory conformity assessments for high-risk AI systems and systemic risk evaluations for general-purpose AI models above a compute threshold. Those evaluations are conducted by notified bodies with defined independence requirements, not by institutes affiliated with the developers themselves. The DeepMind Institute's governance proposals, if adopted, would represent a significant step toward accountability — but they would still fall short of the structural separation that European regulators have concluded is necessary.

The Technical Picture: Gemini, AlphaGenome, and WeatherNext

The leadership transition and institute launch should not obscure the fact that Google DeepMind has had a technically productive September by any reasonable measure.

Gemini 3.8 Flash↗, released on September 2, is the third update to the Flash tier in six weeks. It maintains introductory pricing at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, after which rates are scheduled to double. Performance gains over Gemini 3.7 Flash are measurable across published benchmarks:

• DeepSWE v1.1: 73.7% (up from 65.3% for 3.7 Flash), a meaningful improvement on a software engineering evaluation that tests multi-step code generation and debugging
• OSWorld-2.0: 59.0% (up from 50.6%), measuring performance on computer-use tasks in realistic operating system environments
• HLE-Verified: 54.9% (up from 53.6%), a harder-to-game evaluation of general reasoning that is more resistant to benchmark contamination than standard academic tests

Google attributes the gains partly to a design choice where the model executes additional reasoning steps and iterative tool calls — which means higher token usage for complex tasks, a trade-off that developers integrating the model into cost-sensitive pipelines will need to account for.

A specialised variant, Gemini 3.8 Flash Cyber, replaces the previous 3.5 version and is restricted to trusted testers, government authorities, and critical infrastructure operators through Google's Fairwind Program. Internal testing showed a 2.6x increase in patch accuracy for Chrome vulnerabilities compared to larger commercial models — a specific, verifiable claim that is more useful than generic capability descriptions.

On the scientific side, the AlphaGenome Atlas↗, released September 8, is a 1-petabyte database of precomputed molecular effect predictions for all 9 billion possible single-nucleotide variants in the human genome. The scale is striking: the database is 30 times larger than the AlphaFold Database. The AlphaGenome Variant Impact (AVI) score condenses predictions from both AlphaGenome and AlphaMissense into a single ranking metric, making the resource accessible to clinical researchers without extensive computational infrastructure. Collaborations with the GREGoR Consortium have already demonstrated utility in identifying causal variants linked to epileptic encephalopathy that were previously overlooked.

WeatherNext 3↗, released in early September, represents a different kind of technical ambition. Built on a Functional Generative Network mesh transformer architecture, it ingests live geostationary satellite mosaics directly — bypassing the six-hour data lag inherent in traditional numerical weather prediction pipelines. The model delivers multi-resolution outputs from a single forward pass, from 5 km resolution for temperature and dew point down to 25 km for 3D atmospheric pressure levels. It achieves up to a 60% reduction in Continuous Ranked Probability Score for precipitation forecasting compared to traditional baselines. The model is now integrated into Google Search, Maps, and the Gemini app.

The Broader Competitive Context

These releases do not occur in isolation. The week of September 15–20 also saw OpenAI extend its GPT-6 Astra platform into the legal sector with Astra for Law↗, a configuration that integrates a legal search index covering over 99.9% of published U.S. precedential case law and demonstrated a 40% relative improvement in correctness on the Vals AI Legal Research Bench compared to standard GPT-6 Astra with web search. The move into professional verticals — law, medicine, engineering — is a pattern across the frontier labs, reflecting the recognition that general-purpose capability alone is insufficient for enterprise adoption in regulated sectors.

xAI shipped Grok Voice Transcribe 2.0↗ on September 17, claiming twice the accuracy of version 1.0 with support for multilingual transcription, speaker diarization, word-level timestamps, and multichannel audio up to 8 channels. The release is incremental rather than transformative, but it fills a gap in xAI's product stack for developers building voice-first applications.

Mistral continues to consolidate its European position following its €3 billion Series D↗ announced September 8, which brought its post-money valuation above €21 billion. A September 16 partnership with Mozilla will bring Mistral's open-weight models into the Firefox Smart Window — a distribution channel that reaches hundreds of millions of users without requiring them to interact with a dedicated AI product. The strategic logic is clear: open-weight models that ship inside existing software can accumulate usage at a scale that no standalone AI application can match.

What the Restructuring Signals

The Google DeepMind leadership transition is best understood not as a crisis response but as an acknowledgement that the organisation has outgrown the structure that built it. Hassabis founded DeepMind in 2010 with a specific scientific mission — solve intelligence, use it to solve everything else — and led it through its acquisition by Google, the AlphaGo breakthrough, AlphaFold, and the construction of the Gemini model family. That is an extraordinary run. But the organisation he now chairs employs thousands of researchers, ships consumer products used by hundreds of millions of people, and operates under regulatory scrutiny from multiple jurisdictions simultaneously.

The DeepMind Institute is, in part, an attempt to give Hassabis a platform commensurate with his new role: not the person responsible for the next model release, but the person responsible for articulating what the entire enterprise is for. Whether that platform produces genuine intellectual accountability or functions primarily as a reputational asset will depend on whether its governance proposals are adopted, whether its safety analyses are acted upon, and whether its independence from Google's commercial interests is maintained as the stakes increase.

Those are open questions. What is not open is that the restructuring represents a considered bet on a particular theory of how to navigate the next phase of AI development — one in which the scientific vision and the operational execution are separated, and in which the organisation that builds the most capable systems also takes responsibility for explaining, and constraining, what those systems can do.

---

*Lukas Hoffmann covers European AI labs and frontier research for Neuron. He is based in Berlin.*