In April 2023, Google folded DeepMind and the Google Brain team into a single unit and handed it to Demis Hassabis. It looked like corporate tidying. It was actually a declaration: AGI had moved from a research curiosity to something with a delivery date attached.
Two years on, Google DeepMind is the most quietly consequential lab in the field. It publishes more real science than its rivals, ships through a distribution channel of billions of users, and still talks about artificial general intelligence as a decade-scale problem. Those three facts sit awkwardly together, which is exactly why the hype and the reality are so easy to confuse.
What Google DeepMind actually means by AGI
Hassabis has been consistent on this for years. His version of AGI is a system that can do essentially anything a human can do cognitively, including the messy parts: transferring a skill learned in one domain to a problem in another, holding a goal over days, and figuring out what to do when the instructions run out.
That is a much bigger claim than the one most labs put in writing. OpenAI’s working definition runs to fourteen words and has been picked apart for years, as this breakdown of what OpenAI’s AGI definition actually commits them to makes clear. DeepMind’s bar is higher and vaguer, which cuts both ways. It gives researchers room to chase hard problems instead of benchmark scores. It also means nobody can be held to a missed deadline.
Why the merger mattered
Before 2023, Google had two world-class AI teams competing for the same TPUs and occasionally publishing against each other. Consolidation fixed the compute allocation problem and, more importantly, created one roadmap. Gemini became the flagship. Everything else, from protein folding to robotics, either fed into it or ran alongside it.
Gemini is the vehicle, not the destination
It is tempting to treat Gemini as Google DeepMind’s AGI attempt. It isn’t quite that. Gemini is the surface area, the thing that gets used by Gmail, Search, Android and whoever opens the app. It processes hundreds of trillions of tokens a month across Google’s products.
The interesting technical work sits underneath: multimodal training from the start, context windows that stretch into the millions of tokens, and a push toward agentic behaviour where the model plans and executes rather than answers a single question.
Where it still falls over
Long-horizon work is the sticking point. A model can summarise a 700-page document and then lose the thread three steps into a multi-day task, retry something it already did, or confidently invent a file path. Error rates compound. A task with a 95% success rate per step falls below 40% after eighteen steps, which is roughly the length of a modest real project.
This gap between a demo and a Tuesday afternoon at work is the single best indicator of how far we actually are, and it is worth reading more on why polished demos and reliable work are different problems before taking any AGI claim at face value.
Beyond the chatbot: the science that doesn’t trend
Google DeepMind’s strongest evidence for real progress is not in a chat window. It is in the papers.
- AlphaFold 3 (May 2024) predicted the structure of proteins alongside DNA, RNA and small molecules, not just proteins in isolation. Isomorphic Labs is now trying to turn that into drugs.
- AlphaProof and AlphaGeometry 2 solved four of six problems at the 2024 International Mathematical Olympiad, scoring 28 out of 42, one point short of the gold-medal threshold. The proofs were formally verified, not judged by a language model.
- Genie 2 generates playable 3D environments from a single image, which is a world model rather than a text generator. It matters because embodied reasoning needs somewhere to happen.
- SIMA follows natural-language instructions inside commercial video games, which is a small but genuine test of transferring intent across unfamiliar environments.
None of these are AGI. Each one is a capability that a general system would plausibly need, built by a team that treats generality as a portfolio problem rather than a single model to scale.
Safety is being treated as an engineering constraint
The Frontier Safety Framework, first published in 2024 and updated in early 2025, sets Critical Capability Levels across four domains: autonomous capabilities, biosecurity, cybersecurity and machine learning R&D. Cross a threshold and the model gets extra scrutiny before release.
It is a reasonable structure and an unusually concrete one for this industry. The obvious weakness is that thresholds are defined by the labs that might cross them, using evaluations those labs designed. That is not a reason to dismiss the framework, but it is a reason to read safety announcements as process documents rather than guarantees.
Hassabis’s timelines, decoded
Asked repeatedly when AGI arrives, Hassabis tends to land on five to ten years, sometimes a decade, and he is careful to say current systems are nowhere near it. He has listed the missing pieces more than once: continual learning, long-term memory, planning, and consistency.
That list is more useful than any date. Continual learning in particular is the thing no frontier model has solved. Every system in production today is frozen after training. It can hold facts in a context window, but it does not genuinely learn from an interaction and carry that forward. Solve that and the timeline compresses sharply. Fail to solve it and you get very good tools that plateau.
Public roadmaps are marketing as much as engineering, and the skill of reading them is worth developing separately. There is a decent primer on how to read AGI roadmap updates without getting fooled, and it applies just as well to Google as to anyone else. The same scepticism is useful when comparing the dates OpenAI has committed to over the years, most of which have quietly moved.
The constraint nobody talks about enough
Compute is the quiet variable in all of this. Alphabet guided roughly $75 billion of capital expenditure for 2025, the overwhelming majority of it for technical infrastructure. That buys TPUs, data centres and power contracts, and it is the reason Google can train several frontier models in parallel while smaller labs pick one.
Talent flows the other way. Researchers move between DeepMind, OpenAI, Anthropic and Meta with regularity, and every departure takes tacit knowledge with it. Google’s advantage is depth of bench, not individual genius. An organisation that can lose a dozen senior researchers and keep shipping has a structural edge.
Signals worth watching over the next eighteen months
Ignore the launch events. Watch these instead:
- Any credible demonstration of continual learning in a deployed model, not a research preview.
- Agent benchmarks that measure multi-day task completion with real tools, rather than puzzle suites a model can be trained against.
- Whether the Frontier Safety Framework’s Critical Capability Levels trigger publicly, before a capability ships rather than after.
- Progress at Isomorphic Labs, which is the clearest test of whether DeepMind’s scientific models translate into something with a pipeline and a timeline.
- How Gemini handles memory across sessions, since persistent memory is the closest proxy for the agentic behaviour everyone promises.
The real bottleneck is boring
Google DeepMind has the compute, the distribution and arguably the deepest research culture in the field. What it does not have, and neither does anyone else, is a system that reliably finishes things. Reliability is unglamorous. It means fewer retries, better calibration about what the model does not know, and the ability to hold context across a working week without decaying.
Hassabis says a handful of major breakthroughs stand between here and AGI. That framing is probably right, and the breakthroughs are more likely to come from memory architectures and training objectives than from another leap in parameter count. The lab that solves continual learning first will look, in retrospect, like it was years ahead. Until then, the honest position is that the field has built extraordinarily capable pattern engines and has not yet built a mind that remembers what it learned yesterday.

