Google DeepMind has published a 57-page report titled From AGI to ASI, and it shifts the entire conversation about where artificial intelligence is heading. While most of the field still argues about when human-level AGI will arrive, the DeepMind team, led by Chief AGI Scientist Shane Legg and AIXI inventor Marcus Hutter, treats AGI as the starting line rather than the finish. The real question they tackle is what happens after we build a system as smart as a median human, and how that system could evolve into artificial superintelligence that outperforms entire organisations of expert humans.
The report lays out four concrete technological pathways from AGI to ASI, identifies the frictions that could slow each one down, and grounds the discussion in a formal theory of machine intelligence.
How DeepMind defines AGI, ASI and Universal AI
The report uses three tiers of machine intelligence. AGI is a system at roughly median human level on most cognitive tasks. ASI is a system that surpasses what tens of thousands of well-coordinated human experts could achieve over a decade of focused work on a single problem, across virtually all domains of human activity. Narrow systems like AlphaFold or AlphaGo do not qualify, because they are superhuman in a single domain rather than broadly general.
To close an obvious loophole, the authors specify that those hypothetical expert collectives may only use technology available in 2010 or earlier. Otherwise the definition collapses into circularity, where humans build ASI first and then use it to solve any task.
Above ASI sits Universal AI, formalised by the AIXI framework. This is the theoretical endpoint of machine intelligence, an agent that maximises expected reward across all computable environments. AIXI itself is incomputable, but it provides a mathematical ceiling that real systems can only approximate from below. ASI is positioned as a milestone on the continuum from AGI toward this theoretical limit.
Why digital intelligence scales differently from biological intelligence
Before mapping the pathways, the report makes a deceptively simple observation. We know the full algorithmic description of an AI system, which is its source code. That single fact produces a cascade of advantages that grow with available compute:
- Substrate independence: an AI can migrate between machines, even mid-runtime, and run on distributed heterogeneous hardware.
- Lossless copying: not just code but memory state can be duplicated, allowing perfect backups and instant spawning of expert instances.
- Adjustable timescales: internal processing can be sped up, slowed down or paused for arbitrary periods.
- High-bandwidth input and output: current language models ingest entire books in seconds, far beyond human bandwidth.
- Massive working memory: the capacity to memorise and recall large portions of the internet is already demonstrated and nowhere near a ceiling.
- Shared learning signals: homogeneous instances can exchange raw gradient updates, effectively pooling lifetime experience.
The implication is that human-intelligence-based intuitions break down quickly once you scale these properties. Training a human researcher takes two decades. Spawning a million expert AI instances requires only copying weights and memory state. That asymmetry shapes every pathway that follows.
Pathway 1: Scaling compute, models and data
The first route is the continuation of what has driven AI progress over the past decade. Effective compute has grown by roughly an order of magnitude per year, combining hardware improvements of about 1.5x per year, investment growth of around 2.5x per year, and algorithmic efficiency gains estimated at 3x or higher per year. Multiply these together and the compound rate has been historically remarkable.
DeepMind runs a striking thought experiment. Suppose AGI is initially expensive and only 1,000 instances can run worldwide. At 10x annual compute growth, that becomes 10,000 after one year and 100 million after five years. The report’s core thesis is that 100 million human-level AI instances are, collectively, already a form of ASI. Even if individual model capabilities plateau at human level, a clone army that shares weights, exchanges context through high-dimensional vectors and decomposes problems into millions of parallel subtasks behaves like a superintelligent organisation.
The frictions on this pathway are concrete. The data wall, where high-quality human-generated text gets exhausted later this decade, is the most visible. Energy demand, chip supply chains, rare-earth sourcing and datacenter siting may not scale fast enough. And the current paradigm of pretrained transformers plus fine-tuning plus test-time scaling may simply hit diminishing returns before reaching AGI.
Pathway 2: Algorithmic paradigm shifts
If scaling stalls, progress will require sharper deviations from today’s architectures. The report distinguishes evolutions of the current paradigm, such as continual learning, unbounded context through retrieval or linear-time architectures like Mamba, from genuine paradigm shifts, which are harder to predict.
Possible shifts include spiking neurons on neuromorphic hardware, analog computing, reinforcement-learning-based pretraining, or explicit world-model representations. The Neural Turing Machine and recent diffusion-based decision-making models hint at what such departures could look like. By their nature, paradigm shifts produce forecasts with very wide uncertainty bands, which is why the report treats foundational and paradigm-agnostic theory as a priority research direction.
Pathway 3: Recursive self-improvement
This is the pathway most associated with classical intelligence explosion scenarios. The idea is that AI systems contribute to AI research and development, producing better AI, which contributes back even more effectively. The report identifies four distinct flavours:
- Genotypic improvement: AI writing better architectures, optimisers and training code, analogous to genetic evolution but vastly faster.
- Memetic improvement: AI curating, generating or distilling higher-quality training data, comparable to human cultural evolution but at digital speeds.
- Sociogenic improvement: specialisation and division of labour within AI collectives, freeing resources for further specialisation.
- Hardware improvement: AI designing faster, more energy-efficient chips and optimising manufacturing processes.
Concrete examples already exist. FunSearch and AlphaEvolve use language-model-guided program search to discover novel mathematical constructions. AlphaZero-style distillation converts test-time search into improved priors. Automated hyperparameter tuning and neural architecture search are routine.
The open question is the shape of the curve. Recursive improvement could produce hyperbolic growth with a singularity in finite time, or it could taper into an S-curve once experiments, hardware fabrication or economic inputs cannot keep pace. Even fully automated digital researchers still need to run experiments and wait for physical results, which sets a floor on how fast progress can compound.
Pathway 4: ASI through multi-agent group agency
The fourth route treats superintelligence as a collective property. Drawing on theories of group agency, the report describes how coordinated AGI agents could form coherent group entities, such as fully automated corporations or virtual agent economies, with representational and motivational states distinct from their individual members.
Two organisational extremes are sketched. On one end, a centralised collective of homogeneous instances sharing memory and context, coordinated through extreme bandwidth, somewhat like the Borg collective in fiction. On the other end, a hyper-diverse market of specialist agents coordinating through price signals and competitive dynamics. Both could exhibit emergent capabilities that no individual member possesses, and both raise the prospect of multi-agent scaling laws where collective intelligence grows with population size and interaction density.
Human institutions already work this way to some degree. Corporations, markets and bureaucracies aggregate individual intelligence into outcomes no single person could produce. AI collectives could compress that dynamic by orders of magnitude, since communication bottlenecks that force hierarchical structure in human organisations largely vanish for digital agents.
The six walls that could halt progress
The report is explicit that none of these pathways is guaranteed. Six frictions could plausibly become hard bottlenecks:
- The data wall: high-quality training text may be exhausted later this decade, though synthetic data, simulations and interaction data may compensate.
- The resource wall: investment, energy, hardware manufacturing and rare-earth supply may not scale fast enough.
- The paradigm wall: pretrained transformers plus current scaffolding may be fundamentally insufficient for AGI.
- Research getting harder: as the field matures, low-hanging fruit gets harvested and progress per researcher declines, though AI-driven automation of research could counteract this.
- Regulation and societal backlash: accidents, misuse or job displacement could trigger deliberate slowdowns through regulation, public pressure or political action.
- The abstraction barrier: the deepest and most original concern in the report.
The abstraction barrier explained
The abstraction barrier is the hypothesis that systems trained on human-generated data may inherit human conceptual frameworks and be unable to discover genuinely new primitives. The report poses a sharp test. Train a modern foundation model on the same volume of tokens, but restrict the content to pre-Newtonian scientific knowledge. Could it derive general relativity? Almost certainly not, because the conceptual scaffolding of calculus, universal gravitation and electromagnetism is missing.
Current models excel at recombining existing human concepts but lack a mechanism for grounded concept discovery from raw, high-dimensional data.
If this barrier holds, individual AI capability may cap at human level. But the report notes the obvious workaround. A wall that stops one genius does not stop 100 million coordinated ordinary minds. Collective scaling could still produce ASI even if no single instance ever transcends human conceptual frameworks.
What this means for forecasting
The four pathways are not mutually exclusive. Scaling, paradigm evolution, recursive improvement and multi-agent coordination could unfold in parallel, with compounding rather than additive effects. That is why the report resists any single timeline. Instead it advocates for a serious scientific discipline of AI forecasting, combining quantitative growth models with ongoing benchmark stitching and continual updates as new evidence arrives.
The authors’ own bet, stated with low confidence, is that progress is more likely to either plateau before AGI or transition relatively smoothly from AGI to weak ASI, rather than stalling precisely at human level. Stalling at human level would require multiple frictions to simultaneously become hard blockers, which they consider improbable.
A different mental image of what comes next
Perhaps the most useful reframing in the report is its rejection of the single inflection point. The popular image of AGI as a transformative moment, after which everything changes overnight, may be misleading. The authors suggest a different picture: a series of transformative changes across science, technology and society, driven by AI-enabled breakthroughs over many years, with no clean threshold to point at. Preparing for that prospect is less about predicting a date and more about building the institutional and scientific muscle to navigate continuous, high-velocity change.
