DeepMind’s original website carried one line of mission text: solve intelligence, then use it to solve everything else. That was 2010, when the company was a small team of researchers in a London office. Fifteen years on, the phrasing survives nearly intact, but the stakes have changed beyond recognition. Artificial general intelligence is now the stated goal of Google’s largest research organisation, and the company has put a timeline, a safety framework and a substantial marketing budget behind it.
In 2023 DeepMind merged with Google Brain to form Google DeepMind, with Demis Hassabis as chief executive. The two labs had spent years circling the same problems. Combining them pooled compute and headcount behind a single roadmap, and that roadmap has AGI written at the top rather than buried in a footnote.
What DeepMind Actually Means by AGI
Shane Legg, one of the three co-founders, spent his PhD years trying to pin the term down. His working definition treats intelligence as the ability to achieve goals across a wide range of environments. The weight sits on “wide range”. A system that plays Go better than any human alive is not intelligent by that standard if Go is the only thing it can do.
Judged that way, what artificial general intelligence actually means in research practice is a spectrum rather than a switch. Nobody flips from narrow to general overnight, and the honest assessments tend to be uncomfortable for whoever is doing the assessing.
Two axes, five levels
In November 2023 a group at DeepMind published a paper called Levels of AGI. It scored systems on two dimensions, performance and generality, with a separate scale for autonomy running from zero to five. The authors placed the best models of the day at “emerging AGI” at most, roughly comparable to an unskilled adult human, and noted that nothing had cleared the bar for competent or expert general intelligence. Their own framework, in other words, did not flatter their own products.
The Results DeepMind Keeps Pointing At
The case for progress rests on a short list of results that genuinely changed how the field thinks.
- AlphaGo (2016) beat Lee Sedol 4-1 in Seoul. Move 37 in game two was strange enough that commentators assumed the broadcast had glitched.
- AlphaZero (2017) learned chess from the rules alone in about four hours, then beat Stockfish 25 games to 3 with 72 draws across 100.
- AlphaFold 2 (2020) effectively solved protein structure prediction at CASP14. The database built on it now holds predicted structures for more than 200 million proteins.
- AlphaProof and AlphaGeometry 2 (2024) scored 28 out of 42 at the International Mathematical Olympiad for a silver medal, and AlphaProof solved a problem that only five human contestants managed.
- Gemini models handle text, images, audio and video inside one system, with context windows measured in millions of tokens.
- SIMA and Genie point at embodied agents and generated worlds: one follows instructions inside 3D environments, the other turns a single image into something playable.
It is a remarkable list, and each entry deserved the attention it got. It is also worth asking what the list proves. A survey of artificial general intelligence examples tends to deflate the excitement quickly, because almost none of them survive contact with the word “general”.
Why a Protein Folder Is Not a General Learner
AlphaFold is one of the most useful scientific instruments of the decade. It is also narrow in a way that press coverage tends to skip. It predicts structures. It cannot design an experiment, plan a route across a city, or hold a conversation. Its training consumed enormous volumes of domain-specific data and compute, and a genuinely new class of protein with different folding behaviour would likely require retraining.
Generality shows up in transfer. A system that learns one skill should pick up the next one faster. Gemini’s cross-modal performance is DeepMind’s strongest current evidence of that, since a model trained heavily on text can reason about a diagram it has never seen. Continual learning is the remaining gap. Models forget. Fine-tuning on new data degrades old abilities, and nothing DeepMind ships today absorbs a conversation and keeps that knowledge permanently without weight updates. The traits separating general systems from impressive narrow ones deserve proper attention, and five traits that actually matter is a clearer guide than any launch event.
The Timeline Has Moved, and Not in One Direction
Hassabis said in early 2024 that AGI could be a decade away, with a real chance of arriving sooner. Legg’s long-standing public estimate put 50% odds on 2028, with the 90% mark out around 2085. Those numbers are not identical, and they come from two people with desks in the same building.
Hassabis has since softened the message, telling interviewers that the field is probably missing one or two conceptual breakthroughs on the scale of AlphaGo before systems can plan, reason and remember at anything like a human level. Follow the DeepMind AGI timeline closely and a pattern emerges: confidence about direction, caution about dates.
Google’s Compute Changes the Math, and the Scrutiny
Inside Google, DeepMind sits on TPU capacity that no startup can rent at any price. That matters for a research agenda where scale has repeatedly produced step changes rather than diminishing returns. The plan inside Google DeepMind rests on three legs: compute, distribution through products used by billions of people, and a safety apparatus that keeps pace with the first two.
That apparatus is the Frontier Safety Framework, introduced in May 2024 and revised since. It defines Critical Capability Levels across areas such as autonomy, cybersecurity, biosecurity and machine learning research, and commits the company to evaluations before deployment. Reasonable on paper. The awkward part is structural: the lab sets the thresholds, runs the tests and decides what counts as a pass. External red-teaming exists, but the final judgement stays in-house.
Signals That Would Justify the Language
Ignore the demo videos. A handful of measurable behaviours would tell you far more about real progress than any keynote.
- Persistence. A model that remembers your project on Friday without you re-pasting the context on Monday.
- Long-horizon work. METR’s research suggests the length of task an AI can finish unaided has been doubling roughly every seven months. Watch whether that holds when tasks are measured in weeks rather than minutes.
- Sample efficiency. A human learns to drive in about 15 hours of lessons. Compare that with the fleet mileage behind any autonomous driving model.
- Novelty. ARC-AGI was designed to resist memorisation. OpenAI’s o3 reached 75.7% on its semi-private set in December 2024, though reporting put the cost near $20 per task, which tells you how far brute force stretches before it snaps.
What to Watch Over the Next Eighteen Months
The next Gemini releases will say more than any roadmap post. Look for memory that survives between sessions, agents that complete multi-day workflows instead of single tickets, and whether DeepMind publishes results that failed. Negative findings are a stronger credibility signal than another benchmark record, particularly from a lab whose evaluations are self-administered.
Watch the company’s own autonomy scale as well. When its researchers move a system up a level in their own published framework, methodology attached and limitations spelled out, that will mean considerably more than a phrase in a press release. Until then the most defensible position is roughly the one Legg took years ago: the direction is clear, several of the pieces are in place, and the hardest work has not happened yet.

