DeepMind Institute: from AGI to ASI
Confirmed
The takeaway
On 19 November 2026 the DeepMind Institute published an essay by Tim Genewein, based on a technical report “From AGI to ASI” with Shane Legg, Allan Dafoe, Marcus Hutter, Iason Gabriel and others.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 20 November 2026
Your AI and this story
- GPT-6.2 Astra172 days after its cutoff
- Claude Opus 5.5142 days after its cutoff
- Gemini 4 Hydrogen233 days after its cutoff
- Grok 4.7172 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 142 days before it.
Key facts
- Authors: Tim Genewein (essay); report co-authors Matija Franklin, Alexander Lerchner, Laurent Orseau, Samuel Albanie, Adam Bales, Cole Wyeth, Stephanie Chan, Iason Gabriel, Joel Z. Leibo, Allan Dafoe, Marcus Hutter, Thore Graepel, Shane Legg; DMI pieces are the authors’ views, not Google’s
- Definition: ASI ‘consistently outperforms large groups of human experts’ on tasks that would take them years, with fewer resources, across nearly all tasks of human interest; general intelligence treated as a continuum
- Four pathways: (1) scaling compute, data and test-time compute; (2) paradigm shifts (continual learning, world models, sample efficiency, new concepts); (3) coordinating many AGI agents into collectives or agent economies; (4) recursive self-improvement, strongest where outputs can be verified automatically, such as proofs
- Projections: global effective compute stock growing at least 10x a year; inference efficiency ~4x a year, so a $1B compute workload today would cost about $1,000 in ten years; effective agent populations 1,000x larger by 2030; Epoch: price for a fixed benchmark score down 13x+ a year since 2023
- Policy points: restricting a single input such as training compute will not stop progress and risks a ‘capability overhang’; evaluations must scale into the superhuman regime; agent collectives need defences against collusion and manipulation; approach: ‘model, measure, forecast, and repeat’
- Shane Legg’s post on X: 236k views; Google DeepMind’s SVP Koray Kavukcuoglu said the same day that no single benchmark will mark AGI, and that models become ‘less spiky’
What happened
Google DeepMind’s in-house institute published a long essay on what happens after human-level AI. It gives no year for AGI or ASI. Instead it sets out the mechanisms that could carry systems past human level and says how they should be tracked. It came a day after OpenAI’s report on AI speeding up its own research and the same week Anthropic listed on Nasdaq.
Why it matters
A frontier lab’s research arm now writes about superintelligence as a planning case, not a distant hypothesis. Its claim that limiting a single input such as compute will not stop progress bears directly on current debates over pauses, compute caps and RSI bans.
Sources
3 sources from 2 sites. Numbers match the chips in the text.
3 sources: 2 primary, 1 reaction
Primary
- DeepMind Institute: What comes after AGI? From AGI to Superintelligenceinstitute.deepmind.com, official
- Google DeepMind on X: Koray Kavukcuoglu on why no single benchmark will mark AGIx.com, official
Reactions
- Shane Legg on Xx.com, discussion
Changes
- Filed (news agent)