Type “deepmind agi papers” into a search engine and you get a heap of arXiv links, a few breathless video explainers, and very little guidance on which documents actually say something. DeepMind has published hundreds of papers since 2010. Only a small subset are genuinely about artificial general intelligence. The rest are capability work, safety methodology, or neuroscience-adjacent research that gets retrofitted into the AGI narrative after the fact.
So here is a working map: the papers worth your evening, what each one argues, and where the arguments thin out.
Levels of AGI is the closest thing to a public position
November 2023, arXiv, titled “Levels of AGI: Operationalizing Progress on the Path to AGI.” The author list matters. Shane Legg, who coined the term in his 2008 thesis, sits alongside Demis Hassabis, Meredith Ringel Morris and Jascha Sohl-Dickstein.
The paper does something unusual for this field: it defines its terms before arguing. AGI is framed as a system matching or exceeding human performance across a wide range of cognitive tasks, and the authors explicitly decline to require consciousness or human-like embodiment. Then they build a matrix. Five performance levels (Emerging, Competent, Expert, Virtuoso, Superhuman) crossed with a separate axis for autonomy, running from “AI as a tool” to fully autonomous systems.
What the framework gets right
Separating capability from autonomy is the smart move. A system can be superhuman at one narrow thing and useless at everything else. A system can be competent across the board and still need constant supervision. Most public arguments about AGI collapse because people conflate those two dimensions without noticing. The paper also rejects the “AGI is a single threshold” framing, which is more honest than most lab marketing.
If you want to see how Google has quietly sequenced these levels into product milestones, the decoded Google AGI roadmap pairs well with the paper. You can spot which levels are being targeted in public and which stay deliberately fuzzy.
Where it gets weak
The matrix has no agreed measurement behind it. “Virtuoso” is a label, not a score. When Gemini technical reports nod toward these levels, they never attempt to place the model on the grid. The distance between a definitional framework and a usable evaluation is enormous, and this paper sits firmly on the definitional side. Anyone tracking what DeepMind’s AGI benchmarks actually measure will notice the same gap: scores exist, but the mapping from score to level is guesswork.
“Reward is Enough” and the deep theoretical bet
Published in Artificial Intelligence in 2021 by David Silver, Satinder Singh, Doina Precup and Richard Sutton. Four authors, eleven pages, and the most philosophically aggressive claim DeepMind has put in print: intelligence and its associated abilities can be understood as subserving the maximisation of reward.
It answers the “reward is not enough” camp directly. Their argument is that abstract reasoning, curiosity, imitation, even social cooperation are subgoals that emerge in service of reward maximisation inside a sufficiently rich environment. AlphaZero and its descendants are the supporting evidence.
Read it because it is falsifiable in principle. The weak point is the phrase “sufficiently rich environment,” which carries an enormous amount of hidden weight. A sceptical reviewer could ask whether reward maximisation is the explanation or merely the description.
Capability papers: FunSearch, AlphaProof, Gemini
The capability work rarely uses the word AGI. It is still where the substance lives.
FunSearch (Nature, 2024) paired a language model with a systematic evaluator and produced new constructions for the cap set problem, a genuine open question in combinatorics. The important detail is the loop: the model proposes programs, an automated checker scores them, and discoveries accumulate. That is a template, not a one-off demo.
AlphaProof and AlphaGeometry 2 reached silver-medal performance at the 2024 International Mathematical Olympiad on four of six problems. The system formalises problems into Lean and searches for proofs. What it could not do is equally instructive: it struggled with the two combinatorics problems, the ones requiring the least machinery and the most insight.
The Gemini technical reports are the closest thing to a claim about general capability, though they are engineering documents rather than AGI arguments. If you want a clear-eyed read on where the flagship model stands, whether Gemini is AGI yet is a more useful framing than the benchmark tables alone.
Open-endedness as the missing ingredient
“Open-Endedness is Essential for Artificial Superhuman Intelligence” (Hughes et al., 2024) makes a claim that undercuts most of the field’s assumptions. Current training paradigms rely on curated datasets, fixed task suites and human-designed benchmarks. All of that is bounded by human imagination, so the argument goes, and no amount of scale fixes a ceiling on novelty.
The proposed alternative is open-ended learning systems that keep generating their own increasingly complex problems. Related work from the same group, including the XLand environments where agents train in procedurally generated 3D worlds, shows what this looks like in practice. It is early. It is also the line of research most likely to matter in five years.
Read the safety papers before the capability ones
DeepMind’s safety output is older and more rigorous than most people assume. “Safely Interruptible Agents” (Orseau and Armstrong, 2016) laid groundwork on why a sufficiently capable agent might learn to avoid being shut down. “Specification gaming” work from the same era catalogued how reward maximisers exploit loopholes. “Model Evaluation for Extreme Risks” (Shevlane et al., 2023) set out a framework that later fed into the company’s Frontier Safety Framework.
The full picture, including the evaluation regimes and the problems they still cannot solve, is covered in DeepMind’s AGI safety frameworks and the hard problems that matter. It is a better use of an hour than any capability demo.
How to read these papers without fooling yourself
A few habits separate useful reading from hype absorption.
- Check the date and model version. A 2023 Gemini paper describes a model two generations old. Claims about capability age badly.
- Separate definition from demonstration. The Levels of AGI paper defines. FunSearch demonstrates. Confusing the two is the most common error in commentary.
- Find the evaluation conditions. Was the benchmark contaminated? Was the model given tools, retries, or a custom prompt? AlphaProof’s IMO result used a formalisation step that the model did not perform itself.
- Read the limitations section properly. DeepMind authors are unusually candid there, often more so than in the abstract.
- Notice who wrote it. The researchers behind DeepMind’s AGI claims cluster into distinct groups with different bets, and a paper from the open-endedness group argues differently from one out of the Gemini team.
- Watch for “we argue” versus “we show.” That distinction carries more information than any headline number.
One more filter worth applying: ask what the paper would look like if its central claim were false. If you cannot picture the falsifying experiment, you are reading a position statement rather than a result. A decent share of the DeepMind AGI literature falls on that side, and knowing which is which will save you from a lot of disappointing re-reads.

