In July 2024, a DeepMind system called AlphaProof solved four of six problems at the International Mathematical Olympiad and scored 28 out of 42. Silver medal territory. It needed days of compute per problem, and nobody outside the lab could try it. Still, that result says more about Google’s AGI research than any keynote slide.
Google doesn’t publish a tidy AGI roadmap you can hold it to. What it publishes is papers, model cards, safety frameworks and a steady drip of capability demos. The closest thing to a public statement of intent is DeepMind’s AGI plan, and even that reads like a research agenda with dates bolted on afterwards.
Here’s what’s actually inside the programme, what’s shipping, and where the honest uncertainty sits.
What Google Means by “AGI”
Shane Legg, DeepMind’s co-founder, has been using the phrase since the early 2000s, long before it became a marketing term. His working definition leans on generality: a system that can do anything a human can, including the things nobody has thought to test yet. Demis Hassabis tends to add scientific creativity to the list, which is why maths and biology demos matter to him more than chatbot leaderboards.
Compare that with OpenAI, which compressed the whole idea into fourteen words about highly autonomous systems outperforming humans on most economically valuable work. Same acronym, very different testable claim. Google’s version is harder to falsify, which is convenient for a company that will eventually be judged against it.
The Four Pillars of Google’s AGI Research
Frontier models
Gemini 3 Pro arrived in November 2025 and took the top spot on LMArena, with strong showings on reasoning-heavy tests like ARC-AGI-2 and GPQA. Gemini 2.5 had already pushed adjustable thinking budgets and long context into Google’s product stack. The pattern repeats: each generation adds something that looks like a piece of the AGI puzzle, then gets productised through Search, Workspace and Cloud within months.
Reasoning with verifiable rewards
Maths and code are the training ground because answers can be checked automatically. That makes AlphaProof’s IMO run less of a stunt and more of a data-generation strategy. When correctness is machine-verifiable, you can generate training signal at scale without human labelling. It’s arguably the most important technical thread in the entire effort.
World models and embodiment
Genie 3 produces interactive, playable environments at 720p and roughly 24 frames per second, with a few minutes of visual memory. Gemini Robotics handles manipulation across different robot bodies. Stack those and you get a route to training agents inside simulated worlds, which is exactly the scaffolding AGI researchers have wanted for decades.
Evaluation and safety
DeepMind has taken an unusually public posture here, publishing its Frontier Safety Framework and revising it three times since May 2024. More on that below.
The Narrow-to-General Playbook
Google’s most persuasive AGI argument isn’t a chatbot. It’s a repeating pattern: take a hard, narrow problem, solve it with a general method, then move to the next one.
- AlphaFold predicted structures for roughly 200 million proteins and shared a Nobel Prize in Chemistry in 2024. The technique was problem-specific. The lesson was that the right architecture plus the right data can crack domains people assumed were years away.
- AlphaProof formalised olympiad problems into Lean and searched for proofs using reinforcement learning. Four of six IMO problems, silver-medal level, 2024.
- AlphaEvolve found a way to reclaim about 0.7% of Google’s fleet-wide compute, sped up a TPU kernel by 23%, and produced a better 4×4 complex matrix multiplication algorithm, the first improvement to that problem in 56 years.
Each result is narrow. But narrow in the way a ladder rung is narrow. The research bet is that enough rungs, climbed with the same underlying model family, eventually produces generality rather than a pile of party tricks.
Compute Is the Real Strategy
You can’t separate Google’s AGI research from its silicon. The seventh-generation TPU, Ironwood, reached general availability in November 2025 with roughly ten times the peak compute of TPU v5p, and it’s tuned for inference-heavy workloads rather than training alone. Google owns the chip design, the data centres, the networking fabric and the model. No other lab in the race has all four.
That vertical stack is why Gemini models can be priced as aggressively as they are, and why Google can afford training runs other labs would write off as uneconomical. It also explains the energy headlines. Data centre power has become a binding constraint rather than a footnote in a sustainability report.
Safety Thresholds With Actual Numbers In Them
The Frontier Safety Framework is built around Critical Capability Levels, and the most consequential one concerns machine learning R&D. Level 1 is defined as a model that can fully automate the ML research lifecycle at roughly ten times cost-effectiveness. Google evaluates new Gemini releases against those thresholds before deployment. Whether the evaluations are strict enough is a fair debate, but the bar is at least written down in specific terms.
Tracked capabilities also include cyber offence, biological misuse, harmful manipulation and instrumental self-preservation. That last one is the one worth watching, because it describes the point where a tool starts resisting being switched off.
Timelines: Confident in Public, Fuzzy in Private
Hassabis has said AGI is probably five to ten years out and has put roughly even odds on reaching it by 2030. Legg, who has been forecasting this since the 2000s, lands in similar territory. Google has never published a date it will be held to.
Rivals are less shy. OpenAI has produced a string of dates, and stacking them side by side is revealing, as this breakdown of every AGI timeline OpenAI has given makes clear. Reading the OpenAI AGI roadmap on its own terms is a useful exercise in separating product milestones from research claims. Google’s equivalent exercise mostly involves noticing how rarely the company commits to anything at all.
What to Watch Over the Next 18 Months
Ignore the press releases. These are the signals that will actually move the needle:
- Gemini 4 and the CCLs. If a launch slips because capability evaluations flagged something, the framework is being used for real.
- AlphaEvolve-style results compounding. Automated algorithm discovery finding genuine wins inside Google’s own infrastructure would mean the AI-improving-AI loop is closer than sceptics assume.
- Agent reliability, not benchmarks. Multi-step tasks in Workspace that succeed 80% of the time are a product. Ninety-nine percent is a different category of system.
- Robotics transfer. Gemini Robotics operating hardware it was never trained on would be the clearest evidence of generalisation to date.
- Power and chip supply. Compute is the rate limiter. Watch energy contracts as closely as model cards.
Nobody at Google is going to stand up and declare the race won, and that’s part of why the work is readable. A research programme spread across papers instead of press events leaves a trail: AlphaProof, AlphaEvolve, Genie 3, Ironwood, a safety framework with thresholds in it. All public, all checkable. That is more than most of the field offers, and considerably more useful than a countdown clock.

