DeepMind essay asks what comes after AGI and sketches four paths to ASI

DeepMind essay asks what comes after AGI and sketches four paths to ASI

An essay titled "What comes after AGI?", published on institute.deepmind.com, argues that the regime beyond artificial general intelligence is moving from philosophical speculation to empirical research and governance questions. For decades AGI, meaning machines that match humans across nearly all cognitive tasks, was the milestone, and looking past it felt premature. Now that AGI is "coming within reach", the essay says, what comes next can no longer be put off. While "far from certain", AI could keep advancing at its current pace, and possibly accelerate, toward artificial general superintelligence (ASI), with an impact on science, the economy and society at a scale and speed much greater than AGI itself. Preparing for that means mapping the possible roads from AGI to ASI and building rigorous, reliable ways to track progress.

The essay treats general intelligence as a continuum, with human-level AGI as one point on it. It defines ASI as an advanced system that consistently outperforms large groups of human experts on tasks that could take them years, using fewer resources and less time. Its examples: making the same scientific progress as an entire research field in a fraction of the time, or matching in chip design what large companies can do in a year. ASI would have to meet this bar across virtually all tasks of human interest, not just narrow domains, where some AI systems already beat any human at Chess and Go, predict protein structures or prove theorems that resisted mathematicians for decades. The theoretical upper limits of intelligence are well understood mathematically, and so are hard physical, logical and complexity-theoretic limits, such as predicting climate or immune systems with arbitrary precision. But those limits say little about what lies before them: solving chess perfectly needs more computing power than exists on Earth, yet a thumbnail-sized chip plays well enough to beat any human. The vast space between human-level AI and the theoretical limits is, in the essay's words, largely uncharted.

The second section, "cognitive industrialization", argues that even a pessimist about frontier capabilities should consider what happens if AI hardware and software keep getting more efficient. Effective compute has been driven by three compounding factors. Hardware compute per dollar has risen roughly 1.5x per year for six decades, so the same investment 18 months from now buys twice the compute. Training compute needed to reach a fixed performance level has fallen roughly 3 to 6x per year over the last decade, while algorithmic efficiency at inference improved roughly 3x per year. Spending on AI hardware has grown roughly 2.5x per year over the past decade, and the training compute of the largest frontier runs grows by more than 4x per year. Multiplied together, these give global effective compute stock growth of at least 10x per year, a factor of 10 billion over ten years. If that held, a decade from now each AI agent you can run today could be matched by more agents of the same quality than there are people in the world.

On cost, hardware and inference efficiency gains alone would cut the compute needed for an AI workload by about 4x per year. A cognitive workload that today's AI can do for $1 billion in compute tokens would cost just $1000 in ten years. A recent Epoch study finds the consumer price of reaching a fixed score on a set of benchmarks has fallen by more than 13x per year since 2023. The essay concludes that many cognitive tasks may soon be commoditised, much as mechanical industrialization turned routine physical labor into a commodity, bringing standardization and automation of cognitive work across sectors, industrial-scale infrastructure build-out, and an abundance of certain cognitive outputs. It adds that this is "certainly not a given": practical ceilings on compute build-out and more speculative ceilings on algorithmic efficiency together might still allow today's effective AI agent population to grow by more than a thousand-fold by 2030 and more than a billion-fold within ten years.

Would scale alone reach ASI? Scaling only today's AIs in speed and number might not even count as full AGI, since some cognitive tasks may still need human help. But if broadly human-level AI exists, there is a major difference between slow, expensive systems and billions of agents acting much faster than humans, and that quantitative leap might itself bridge the gap from AGI to (weak) ASI. The essay lists ways digital intelligence benefits from more compute: processing speed; working memory and memorization; substrate independence (the same system can move between computers and hardware); lossless replication of both code and memory state; and high-bandwidth sharing of experience across a population of AIs.

The final section, "Four pathways from AGI to ASI", starts from the point that some missing capabilities may be hard or impossible to patch with volume. The essay's illustration: Einstein did figure out general relativity, but a conference of one hundred thousand Ptolemaic astronomers probably wouldn't have. It says AI could follow four technological pathways to unlock previously unavailable cognitive capabilities, and points to a separate "From AGI to ASI" report for depth.

Pathway one is scaling models and data. Capabilities have so far improved consistently with scale, captured by scaling laws covering training compute, data and test-time compute. But these laws are extrapolations that hold only within familiar boundaries (an architecture, a type of data, a family of evaluations) and cannot confidently predict the scale needed for AGI or ASI. The essay names three bottlenecks. Data exhaustion: demand for high-quality data outpaces its supply, synthetic data helps only partly, and researchers could turn to high-fidelity simulations, reinforcement-learning environments and test-time compute to create better training data. Diminishing returns: each linear step in capability typically needs multiplied training compute and model size, and bigger models are slower, which squeezes methods like chain-of-thought reasoning. Physical and economic limits: power plants, grids, chip fabs and low-latency interconnects, with unclear answers on whether efficiency gains can compensate.

Pathway two is technological paradigm shifts. Today's paradigm trains large transformers to predict the next symbol, refines them to reason and follow instructions, and adds test-time compute. Recent research suggests it can in principle yield models that run complex algorithms, including powerful learning algorithms, through their prompts without retraining, a key ingredient for systems at the theoretical upper limit of intelligence. Yet today's models fall short of humans on continual learning, certain forms of world modeling, decision-making and sample efficiency, and may lack the ability to invent genuinely new concepts rather than recombining existing ones. Researchers are divided on whether simple extensions of the current paradigm can fix this or whether deep paradigm shifts are required.

Pathway three is improving agent coordination. Corporations, markets and research communities solve problems no individual could, and coordinating thousands or millions of AGI instances into "group agents" or virtual agent economies could produce collective superintelligence. What is unclear today is how AI group intelligence scales as agents and compute are added, depending on how agents are organized. The available text of the essay breaks off in this section, so the fourth pathway is not covered here.

Key facts

  • The essay on institute.deepmind.com says AGI is coming within reach, so the regime beyond it (ASI) needs empirical study and tools to track progress. It calls continued progress toward ASI "far from certain".
  • It defines ASI as a system that consistently beats large groups of human experts on tasks that could take them years, using fewer resources and less time, across virtually all tasks of human interest.
  • Multiplying hardware (about 1.5x per year), algorithmic efficiency and investment (about 2.5x per year) gives effective compute growth of at least 10x per year, or a factor of 10 billion over ten years.
  • A workload costing $1 billion in compute tokens today would cost $1000 in ten years if efficiency trends hold; a recent Epoch study puts the benchmark price decline at more than 13x per year since 2023.
  • The essay names four technological pathways from AGI to ASI and describes three: scaling models and data, technological paradigm shifts, and improving agent coordination.

Why it matters

Here a major lab's institute argues that superintelligence is no longer too distant to analyse. The essay does not claim anyone has built AGI or ASI. Its case is that if present efficiency and investment trends continue, cheap and abundant cognitive work could arrive even before any new capability breakthrough, and that a quantitative leap in speed and scale might itself carry AI from AGI to weak ASI. It also frames the work as building rigorous, reliable ways to track progress along the path.

Who it affects

The essay speaks of impact on science, the economy and society, and of standardization and automation of cognitive work across sectors. Its direct audience is AI researchers and those thinking about governance, since the subtitle ties the post-AGI regime to empirical research and governance questions. The text gives no specific policy recommendations.

How to use it

It is a framework rather than a tool. Readers can use the continuum of general intelligence, the ASI definition and the four-pathway structure as a checklist for judging claims about progress: does a given advance come from more scale, a new paradigm, or better coordination of many agents? The essay points to a separate "From AGI to ASI" report for in-depth discussion.

How solid is it

The compute and cost figures are trend extrapolations stated in the essay, with one external reference, an Epoch study on falling benchmark prices. The essay itself says scaling laws hold only within familiar boundaries and cannot confidently predict the scale needed for AGI or ASI. The thousand-fold growth in effective agent population by 2030, and billion-fold within ten years, is presented as a possibility, not a forecast. No author names or publication date are given in the visible text, and the text available covers three of the four pathways.

Risks and caveats

The essay is open about its limits: such a proliferation of intelligence is "certainly not a given", and it is unclear whether efficiency gains can offset limits on power plants, grids, chip fabs and interconnects. Researchers are divided on whether today's paradigm can reach ASI with simple extensions. Scaling only today's AIs might not even count as full AGI. It is a conceptual piece with no new experimental results, and it offers no timescale or probability for reaching AGI or ASI.

“While Einstein did figure out general relativity, a conference of one hundred thousand Ptolemaic astronomers probably wouldn’t have.”

— DeepMind essay, "What comes after AGI?"