Google's Post-AGI Paper: What It Actually Says

On 10 June 2026, Google DeepMind researchers posted a paper titled From AGI to ASI to arXiv, where it carries the identifier 2606.12683. It is credited to a team that includes Tim Genewein, Matija Franklin, Laurent Orseau, Samuel Albanie, Marcus Hutter, Thore Graepel and DeepMind co-founder Shane Legg. Coverage from outlets and commentators, including the YouTube analyst Wes Roth, quickly framed it as DeepMind's view of "what comes after AGI".

The paper's purpose is narrow and explicit. It assumes, for the sake of argument, that human-level AGI arrives, and then asks a single question: by what technological routes could intelligence keep climbing past that point, and what could stop it. The paper does not claim AGI has been built, does not announce a product, and does not predict a date for superintelligence. It is a survey and a roadmap, written in the cautious register of a research lab rather than the breathless register of its coverage.

• The thesis: the path from AGI to ASI is plural, not a single jump. Four overlapping pathways could each contribute.
• The four pathways: scaling, AI paradigm shifts, recursive improvement, and multi-agent collectives.
• The definition: ASI is set above large expert organisations, not just above a single expert.
• The brakes: seven bottlenecks, from a data wall to a deliberate, safety-driven slowdown.
• The framing: a series of transformative societal changes is judged more likely than a single "singularity" event.

The paper situates itself as the third instalment in a deliberate DeepMind sequence. An earlier paper set out to define AGI, a 2025 paper addressed making AGI safe, and this one maps the territory beyond it. Taken together, the sequence signals that DeepMind is treating superintelligence as something to be planned for rather than dismissed.

Its definitions matter, because they set the bar deliberately high. The paper describes AGI as a system with roughly median human-level intelligence across most cognitive tasks, already superhuman in some narrow domains but not yet general enough to count. It then defines artificial superintelligence as a system "far surpassing human-level AGI in a broad sense", one that exceeds "the performance of large human-expert collectives on virtually all tasks and domains". The reference point is not a single genius but an entire skilled organisation.

The paper places both AGI and ASI on a single continuum of machine intelligence, drawing on the Legg-Hutter framework that scores an agent by its average performance across all computable tasks, with simpler tasks weighted more heavily. At the theoretical top of that continuum sits Universal AI, formalised through the AIXI agent. AIXI is, by construction, incomputable, but the paper notes it can be approximated from below, and treats it as the formal upper bound that real systems can only move towards. Recent work on combining pretraining with planning is cited as a sign that the dominant deep-learning paradigm may, in principle, edge in that direction.

Alongside the definitions, the paper catalogues the structural advantages that digital intelligence holds over biological intelligence, advantages that tend to intensify with scale. These include high-bandwidth input and output, fast internal processing, large and reliable working memory, substrate independence, lossless copying, and the ability to share learned experience across instances at high bandwidth. The paper is careful to balance this with a counter-argument it calls the "abstraction barrier": because AI systems can communicate at enormous bandwidth, they may never be forced to compress the world into the deep abstractions that humans develop precisely because our own communication is so limited.

The heart of the paper is the four-pathway framework. The paper presents these as routes that could individually or jointly carry capability past the AGI threshold, and stresses that they overlap rather than compete.

1. Scaling

The first pathway is the familiar one: keep increasing effective compute, model size and data. The paper sketches how effective compute has been growing on the order of ten times a year, decomposing that into hardware gains, rising investment and improvements in algorithmic efficiency. The open question it raises is whether continued quantitative scaling produces qualitative jumps in capability or merely smooth, incremental progress. It flags data exhaustion, the depletion of high-quality training text, as the most concrete near-term limit, with synthetic and self-generated data offered as possible mitigations.

2. AI paradigm shifts

The second pathway is the arrival of a genuinely new approach that replaces or extends the current recipe of transformer pretraining, reinforcement learning tuning and test-time compute. The paper discusses candidate directions such as unlimited context through retrieval or recurrence, continual learning, and explicit world models, while conceding that "true paradigm shifts are, by their nature, difficult to predict". This is the pathway the paper is most openly humble about, since forecasting an unknown breakthrough is close to impossible by definition.

3. Recursive improvement

The third pathway is AI improving AI. The paper breaks this into several mechanisms: changes to code and architecture, data-driven improvement such as synthetic generation and distillation, division of labour across specialised systems, and gains in hardware and infrastructure. It notes that weak recursive loops already exist in the form of neural architecture search and automated tuning, and that stronger forms are emerging in systems such as FunSearch and AlphaEvolve. If such loops became fully autonomous, the paper says, growth could in principle turn hyperbolic, but it is candid that the dynamics are poorly understood and the loop could equally "fizzle out".

4. Multi-agent collectives

The fourth pathway is the least discussed elsewhere and arguably the most distinctive. Rather than one giant model, ASI might emerge from large numbers of AGI-level agents coordinating into "group agents", for example fully automated corporations. The paper considers both centralised collectives sharing a single goal and decentralised agent economies coordinated by something like price signals. The key uncertainty it raises is whether collective intelligence scales super-linearly with the number and interaction density of agents, and how such collectives can be steered.

The seven bottlenecks

Crucially, the paper devotes as much attention to what could stop progress as to what could drive it. It lists seven bottlenecks: a data wall, resource and infrastructure limits, the possibility that pretrained neural networks are simply insufficient, rising research difficulty as low-hanging fruit is picked, the abstraction barrier, a deliberate slowdown driven by safety concerns and regulation, and plain "unknown unknowns". For each it offers possible counters, but it repeatedly stresses that the real-world impact of these brakes remains an open research question.

The paper is also frank about limits that no amount of intelligence can remove. It tabulates fundamental constraints on any ASI, including the laws of physics, real-time bounds, the friction of acting in the physical world, irreducible uncertainty about the world, complexity-theoretic hardness, and logical limits such as Gödel's results and the halting problem. In its own words, "ASI is certainly bound by some fundamental physical and complexity-theoretic limitations".

The significance of this paper is as much about who wrote it and when as about its specific claims. When a frontier lab with Shane Legg and Marcus Hutter on the author list publishes a formal roadmap past AGI, it moves "post-AGI" out of the realm of speculation and into the realm of corporate planning. That shift in framing is the news.

The redefinition of ASI is the second reason it matters. By setting the bar above large organisations of experts rather than above a single expert, the paper quietly raises the threshold for what would count as superintelligence. This matters for how progress is measured and how claims are tested, and it pushes back against the habit of declaring each new model a step towards superintelligence.

Third, the paper's "no single moment" framing reshapes the popular narrative. Instead of one dramatic takeoff, it argues for "a series of transformative societal changes" spread across domains and years. That has practical consequences. A gradual transition offers more time to adapt and govern, but it also makes the dangerous moments harder to spot, because there is no single bright line to watch for.

Finally, the paper lands in the middle of an intensely competitive moment for frontier AI, the subject of our June 2026 model wars roundup. Read in that context, it functions partly as DeepMind staking out the intellectual high ground on the long-term trajectory, while rivals fight over this quarter's benchmarks.

The paper spread quickly through AI commentary channels. Wes Roth covered it on YouTube and on X, where he summarised it as a report "examining how artificial intelligence could progress beyond human-level intelligence", and several explainer sites published walkthroughs of the four pathways within days. Much of the coverage was respectful, treating the paper as a serious attempt to structure an under-examined question.

The praise tends to centre on three things: that the paper takes bottlenecks as seriously as accelerants, that it resists the temptation to put a date on superintelligence, and that it gives a vocabulary, the four pathways and seven bottlenecks, for a debate that has often been vague. The inclusion of the AIXI framework also won approval from readers who wanted the discussion grounded in a formal theory of intelligence rather than vibes.

The criticism, fairly summarised, runs in a few directions. Some readers note that the paper is light on safety and alignment specifics: it lists a "deliberate slowdown" for safety reasons as one bottleneck among seven, and gestures at ideas such as iterated amplification, but it does not offer a detailed alignment roadmap, focusing instead on capability trajectories. Others are sceptical of the AIXI anchor, pointing out that AIXI is incomputable and that the gap between that theory and working systems is vast, so its role here is more philosophical than predictive. A third strand argues that the whole exercise is premature, since AGI by the paper's own definition does not yet exist, and that publishing a roadmap beyond it risks normalising assumptions that have not been demonstrated. Finally, some critics read the timing through a commercial lens, as a way for DeepMind to shape the long-horizon narrative.

It is worth being clear that the paper itself anticipates several of these objections. It repeatedly foregrounds uncertainty, it admits paradigm shifts are unforecastable, and it stresses that whether the bottlenecks bite is unresolved. The most defensible reading is that this is a structured map of possibilities, not a forecast, and much of the friction comes from coverage that reads it as the latter.

For most people building with or buying AI tools today, the paper changes nothing about which model to pick this week. It is a long-horizon document, not a buyer's guide. If you are weighing the current frontier models against each other, our Claude versus ChatGPT, Gemini and Grok comparison is the more relevant starting point.

That said, three takeaways are useful to carry forward. First, treat "steps towards superintelligence" marketing with caution: by this paper's definition, ASI sits above whole expert organisations, a bar no current product is near. Second, watch the multi-agent pathway, because the idea of capability emerging from coordinated fleets of agents maps directly onto where many products are already heading, and it implies the interesting risks may be at the system level rather than inside any one model. Third, the "gradual, multi-step" framing is the practical heart of it: rather than waiting for one big moment, expect a steady accumulation of capability that rewards organisations which build the habit of evaluating, governing and adapting continuously.

The paper's own closing recommendation is in the same spirit. It calls for the work of mapping the road past AGI to be a "massively interdisciplinary endeavour of global scope", spanning measurement of AI progress, the study of multi-agent dynamics, and the bridging of universal-intelligence theory with empirical deep learning. The practical message is less "prepare for a singularity" and more "build the measurement and governance muscles now, because the changes will be cumulative".

What is Google's "post-AGI" paper?

It is a June 2026 Google DeepMind paper titled From AGI to ASI, posted to arXiv as 2606.12683 on 10 June 2026. Covered widely as "what comes after AGI", it maps the possible transition from human-level AGI to artificial superintelligence and the bottlenecks that could slow it. It is a survey and roadmap of roughly 57 to 60 pages, not a model release or a claim that ASI has arrived.

What are the four pathways from AGI to ASI?

Scaling (more compute, models and data), AI paradigm shifts (new architectures or training methods), recursive improvement (AI building better AI), and multi-agent collectives (groups of AGI-level agents coordinating into something more capable than any one of them). The paper treats them as overlapping rather than mutually exclusive.

How does the paper define ASI?

As a system that far surpasses human-level AGI in a broad sense, exceeding the performance of large human-expert collectives on virtually all tasks and domains. The bar is set above whole organisations of experts, and the paper uses Universal AI, formalised via the AIXI agent, as the theoretical upper limit.

Does the paper say superintelligence is coming soon?

No. It is cautious about timelines, stressing that uncertainty margins are wide, that paradigm shifts cannot be forecast, and that bottlenecks such as a data wall and rising research difficulty could slow things. It argues a series of transformative changes is more likely than a single watershed moment.

Who wrote the From AGI to ASI paper?

A team of Google DeepMind researchers, including Tim Genewein, Matija Franklin, Alexander Lerchner, Laurent Orseau, Samuel Albanie, Marcus Hutter, Thore Graepel and co-founder Shane Legg, among others. Hutter is the theorist behind the AIXI model that anchors the paper.

From AGI to ASI is best understood as a map, not a prophecy. It assumes AGI for the sake of argument and then asks, with unusual rigour, how intelligence might climb further and what could stand in the way. Its lasting contribution is vocabulary: four pathways, seven bottlenecks, a high bar for what ASI even means, and a formal anchor in universal intelligence.

The reaction has been loud because the framing is new. A leading lab is now treating life after AGI as a planning problem rather than a thought experiment. The paper's own posture is more measured than its coverage: it leans on uncertainty, it takes the brakes as seriously as the accelerator, and it expects a gradual accumulation of change rather than one dramatic leap. Read it as a structured set of possibilities, attribute its claims carefully, and it is a genuinely useful guide to how the people building these systems are thinking about where the road goes next.