DeepMind publishes « From AGI to ASI »: four paths to superintelligence, and why AGI is only the starting point

DeepMind publishes "From AGI to ASI": four paths to superintelligence, and why AGI is just the starting point

🔎 Why is the world's most powerful lab mapping the post-AGI era now?

On June 10, 2026, Google DeepMind published a 57-page report on arXiv titled "From AGI to ASI". Not a blog post. Not a keynote. A technical document signed by 14 researchers, including Shane Legg (co-founder of DeepMind), Marcus Hutter (creator of the AIXI theoretical framework) and Allan Dafoe (head of long-term strategy).

This document does not ask whether AGI will arrive. It assumes it is an intermediate step and maps four parallel paths to ASI — Artificial Superintelligence, defined as a system that exceeds the cognitive capabilities of large coordinated human organizations.

The institutional signal is strong. When the engineers who built AlphaGo, AlphaFold and Gemini treat the AGI→ASI transition as a tractable engineering problem, it is no longer speculation. It is a roadmap.

The key points

• DeepMind defines a four-level continuum: Current AI → AGI → ASI → Universal AI (UAI), the optimal AIXI agent on any computable task.
• Four non-exclusive pathways are identified: continuous scaling, algorithmic breakthroughs, recursive self-improvement, and multi-agent collective intelligence.
• "Effective compute" is growing by approximately 10x per year (1.5x hardware × 2.5x investment × 3x algorithmic efficiency), according to the analysis by Tech Times.
• The identified frictions are as concrete as physical walls (compute, energy), alignment bottlenecks, and real-world verification latencies.
• AGI is no longer the destination. It is the starting point.

Recommended tools

| Gemini 3.1 Pro | Long document research and analysis | Variable pricing (June 2026, check on deepmind.google) | Overall benchmark #1 (score 92) |
| GPT-5.5 | Agentic reasoning and autonomous tasks | Variable pricing (June 2026, check on openai.com) | Agentic benchmark #1 (score 98.2) |
| Claude Opus 4.7 Adaptive | Critical analysis and safety | Variable pricing (June 2026, check on anthropic.com) | Agentic score 94.3, overall 90 |
| Hostinger | AI project hosting / demos | Starting from 2.99€ (June 2026, check on hostinger.com) | Fast and affordable deployment |

The four levels: from current AI to universal AI

The report doesn't beat around the bush. It establishes a precise four-tier reference framework.

Current AI: today's systems, high-performing in targeted domains but without generality. Think of Claude Sonnet 4.6 (score 83 overall, 81.4 agentic) or DeepSeek V4 Pro (88 overall). Effective, but limited.

AGI: the median human level across all cognitive tasks. Not an über-lie, just parity with a typical human professional. GPT-5.5 (91 overall, 98.2 agentic) and Gemini 3.1 Pro (92 overall) are approaching it, without being able to confidently claim it has been crossed.

ASI: the system exceeds what large coordinated human organizations can accomplish. It's not "a little better than a human." It's "better than a hundred engineers working together for a year."

Universal AI (UAI): the theoretical optimal agent — Hutter's AIXI framework. Optimal on any computable task. It's a mathematical limit, not an engineering goal.

The report's key insight, echoed by The AGI Clock: "AGI is no longer the destination — it is a waypoint."

Path 1 — AGI Scaling: continuing to floor it

The first path is the most obvious. Take an AGI and scale it: more data, more compute, more parameters.

"Effective compute" — the actual power available for training — is growing by about 10x per year. This isn't just a matter of chips. It's the product of three factors: hardware improvement (~1.5x/year), increased investment (~2.5x/year), and algorithmic efficiency gains (~3x/year).

But this pathway runs into concrete walls. First, the data wall: there isn't an infinite amount of quality human text on the internet. Next, the compute wall: a 5 GW datacenter campus has become the standard planning unit for frontier labs in 2026, according to the analysis of The AGI Clock.

And above all, scaling produces emergent discontinuities. Capabilities that appear abruptly at certain scales, with no possible prediction. This is both what makes scaling exciting and what makes it dangerous: you don't know which capability will "turn on" at the next order of magnitude.

Pathway 2 — Paradigm shifts: when AI invents a new way to learn

The second pathway does not depend on size. It depends on elegance.

A paradigm shift is moving from one architecture to another that is fundamentally more efficient. From statistical pattern recognition to true compositional generalization. From a model that reproduces to a model that understands.

Explainx.ai emphasizes that this pathway stands out from scaling due to its non-linear nature. Scaling yields gradual and relatively predictable improvements. A paradigm shift can produce a qualitative leap overnight.

The DeepMind report identifies two key metrics for measuring these shifts: sample efficiency (how much data is needed to learn a skill) and compositional generalization (the ability to combine acquired skills to create new ones).

A thought experiment cited in Tech Times illustrates the current limit: an AI trained on all pre-Newtonian physics could not produce general relativity. Transformative creativity — Margaret Boden's ability to create a new space of possibilities — remains out of reach. Current AI masters combinatorial creativity (mixing existing ideas) and exploratory creativity (searching within a known space), but not transformative creativity.

This pathway is the most unpredictable. No one can plan a paradigm shift. But when it happens, it can render all other pathways obsolete overnight.

Pathway 3 — Recursive improvement: AI that improves itself

This is the pathway that fascinates the most — and worries the most the most. Recursive self-improvement, or RSI (Recursive Self-Improvement), is the mechanism by which an AI becomes capable of improving its own capabilities, triggering a loop of exponential progress.

This concept is not new — it is central in the work of RSI est le nouveau AGI : pourquoi l'auto-amélioration récursive obsède la Silicon Valley. But the DeepMind report brings a crucial distinction: AI-assisted improvement vs AI-autonomous improvement.

In the first case, AI is a tool in the hands of human researchers. It generates architectures, writes training curricula, proposes optimizations. Humans validate and integrate. This is already underway — Anthropic publishes its own analyses on this subject.

In the second case, AI modifies its own code without significant human supervision. This is where the loop can theoretically run away.

But the report identifies a physical bottleneck often underestimated: the speed of RSI is bounded by the slowest verification loop. An AI can generate thousands of modifications per second, but if each modification must be tested on a benchmark that takes hours, the loop slows down drastically. In the real world, you cannot simulate reality instantaneously.

The concrete mechanisms of RSI listed by MindStudio include: automated architecture search, AI-generated training curricula, self-modification of code, and AI-assisted research.

Pathway 4 — Collective intelligence: coordinated swarms of AGI agents

The fourth pathway is perhaps the most underestimated by the general public. It is not about making a single AI smarter. It is about coordinating dozens, hundreds, or even thousands of AGI agents into a cognitive collective.

MindStudio classifies this pathway as "active and growing rapidly" in its status table. This is consistent with what we observe: the agentic benchmarks from June 2026 show GPT-5.5 at 98.2, Gemini 3 Pro Deep Think at 95.4, Claude Opus 4.7 at 94.3. These models are no longer simple chatbots — they are agents capable of planning, executing, and iterating.

The distinction between a single agent and an avatar IA vs chatbot becomes crucial here. A chatbot responds. An agent acts. A collective of agents coordinates.

The identified mechanisms: specialized agents working in parallel, oversight and cross-critique between agents, shared persistent memory, and coordinated tool use. The analogy with a human company is illuminating: no single individual builds an airplane alone, but a coordinated company does. A collective of AGI agents could do the same thing, but at digital speed.

The major friction: coordination costs. As the collective grows, communication, goal alignment, and conflict resolution between agents become increasingly expensive. It is not a free lunch.

Concrete barriers: why ASI is not guaranteed

The report is remarkable for its honesty about the obstacles. DeepMind is not selling a deterministic roadmap. It maps out frictions just as serious as the pathways.

Physical walls

Compute and energy are the most trivial but most real constraints. A 5 GW datacenter is the consumption of a city of several million inhabitants. The electrical infrastructure, cooling, chip supply chain — all of this has monumental inertia.

Alignment bottlenecks

The report treats alignment as a working assumption, not a solved problem. Three sub-problems are identified: scalable oversight (how to supervise a system smarter than you), safe self-modification (how an AI can modify itself without breaking its own alignment constraints), and collective alignment (how to align a swarm of agents with each other, not just a single system).

This is where the report intersects with institutional concerns. When Meta Muse Spark betrayed open-source with its first closed model, it was partially motivated by safety concerns related to the diffusion of advanced capabilities. The tension between openness and control will intensify on the path to ASI.

Interpretability walls

The more capable a system becomes, the more opaque it becomes. If we don't understand how a model makes its decisions, how can we guarantee that it remains aligned during a self-improvement loop?

Institutional friction

Faster Please! recalls the political context: the report was published one day after the launch of Anthropic Fable 5, two days before a US government ban. Governments are not passive spectators. Regulation, export controls, moratoriums — all of this can slow down, or even halt, certain pathways.

The report also names a structural risk: "military-economic adaptationism," this competitive pressure between nations that pushes for deployment before validation. Exactly the type of dynamic that turns a technological race into an arms race.

The AGI-to-ASI gap: how long?

The inevitable question. The DeepMind report does not give a date. But analysts have compiled the estimates circulating in the community.

These figures, reported by The AGI Clock, show massive dispersion. But note the bias: no credible person is saying "never". The question is no longer "will it happen" but "how much time between the clear AGI signal and the ASI tipping point".

Hassabis himself stated that AGI would arrive "maybe 2030, plus or minus a year", and that it was missing "one or two big breakthroughs" in continual learning, memory, context windows and long-term reasoning, according to Faster Please!.

"Digital minds": why the substrate matters

A passage from the report, detailed by Tech Times, lists the structural advantages of digital minds over biological minds. This is not science fiction. It is engineering.

First, input-output speed: a model like GPT-5.5 can process millions of tokens per hour. A human reads perhaps 250 words per minute. Next, working memory: no limit of 7±2 items. Substrate independence: no sleep, no fatigue, no aging. Lossless copying: an AGI can be duplicated in a few hours. High-bandwidth sharing: two agents can synchronize their states in seconds, whereas it takes two humans years to transfer tacit knowledge.

These advantages are not theoretical. They are inherent to the very fact of being a computational process. And they accumulate: each advantage accelerates the others. An agent that doesn't sleep has more time to improve. An agent that copies itself can work in parallel. An agent that shares perfectly experiences no loss of knowledge between iterations.

Implications: what this changes concretely

For research

The report redefines the research agenda. If AGI is a waypoint, researchers must think beyond current benchmarks. MMLU or HumanEval-type metrics become as relevant as measuring a marathon runner's speed to predict their Formula 1 performance.

Research in interpretability, scalable oversight, and collective alignment becomes just as much of a priority as research in architecture or compute efficiency. Otherwise, we are building an engine without brakes.

For governance

This is perhaps the most urgent implication. Current regulatory frameworks are designed for "current" AI. They are not designed for systems that self-improve or operate in coordinated collectives.

The distinction between oracle systems (which advise) and agentic systems (which act) becomes a public policy issue. An oracle that says "here is the optimal molecule" is fundamentally different from an agent that synthesizes the molecule, tests the results, and iterates without supervision.

For industry

Companies that treat AGI as a definitive competitive advantage are mistaken in their framing. If the AGI→ASI gap is measured in months or years, the advantage goes to whoever starts building the transition infrastructure earliest. It's not the model that matters, it's the feedback loop between the model, the compute, and improvement.

❌ Common mistakes

Mistake 1 : Confusing AGI and ASI

AGI is parity with the median human. ASI is surpassing coordinated human organizations. This is not a difference in degree. It is a difference in nature. Saying "we have AGI, so we have ASI in five minutes" ignores the physical, alignment, and coordination frictions that the report details exhaustively.

Mistake 2 : Believing that scaling alone is enough

Pathway 1 (scaling) is the most visible, but it is also the one with the most predictable walls. Data, compute, energy — these are finite resources. Pathways 2 (paradigms) and 3 (RSI) are the ones that can produce non-linear jumps. Ignoring them means having an incomplete vision of the landscape.

Mistake 3 : Treating the four paths as mutually exclusive

DeepMind insists: the pathways are parallel and potentially synergistic. An algorithmic breakthrough (pathway 2) can unlock scaling (pathway 1). RSI (pathway 3) can accelerate paradigm innovation. Collectives (pathway 4) can amplify each of the other three.

Mistake 4 : Ignoring institutional frictions

ASI will not be built in a political vacuum. The report explicitly mentions "military-economic adaptationism" and regulation as possible frictions. The history of technology shows that social and political constraints are often more determinant than technical constraints.

❓ Frequently Asked Questions

Who are the authors of this report?

14 researchers from Google DeepMind, including Shane Legg (co-founder), Marcus Hutter (creator of AIXI), Allan Dafoe (long-term strategy), Tim Genewein, Laurent Orseau, Adam Bales, Iason Gabriel, and Joel Z. Leibo. Published on June 10, 2026, on arXiv.

Is ASI imminent?

The report does not give a date. External estimates range from a few months (Aschenbrenner) to a decade (Hassabis). The report maps out the pathways and frictions; it does not publish a timeline.

Which pathway is the most likely?

None are exclusive. MindStudio classifies scaling as "active but constrained", algorithmic innovation as "active", RSI as "early stage", and multi-agent as "active and rapidly growing". The combination of multiple pathways is the most coherent scenario.

Is this report different from previous publications on superintelligence?

Yes, in its institutionalization. This is not a philosophical essay. It is a Google DeepMind research document, with authors who have built the most capable AI systems in the world. The framing is one of engineering, not speculation.

Is alignment solved according to the report?

No. The report treats it as a necessary working assumption, not a solved problem. The three sub-problems (scalable oversight, safe self-modification, collective alignment) are identified as potential bottlenecks.

✅ Conclusion

The "From AGI to ASI" report from Google DeepMind marks a turning point: superintelligence is no longer a science fiction concept discussed in essays, it is an engineering problem that the world's best researchers are methodically mapping out — with four parallel paths, concrete frictions, and no guarantee of success. AGI is not the finish line. It is the starting gun.