Google DeepMind researchers argue that AI may keep advancing past human-level AGI and reach artificial superintelligence, or ASI. They describe four technical routes for the jump from AGI to ASI, and they say the question has moved from philosophy to something that can be measured and governed.
Their view is that AGI, meaning machines that match humans across nearly all cognitive tasks, is now close enough that looking beyond it is no longer premature. The outcome isn’t certain, they stress, but it’s a serious possibility.
What counts as ASI
The researchers treat general intelligence as a continuum, with human-level AGI as one point on it. ASI sits further along. It would be a system that consistently beats large groups of human experts on tasks that take them years, using fewer resources and less time.
They give two examples: matching the scientific progress of an entire research field in a fraction of the time, or making chip-design advances that rival what large companies manage in a year. The bar applies to virtually all tasks of human interest. Narrow wins don’t count, even though some systems already beat any human at chess and Go or predict protein structures.
The theoretical ceiling on intelligence is reasonably well understood, and so are hard physical and complexity limits, such as predicting climate or immune systems with arbitrary precision. The large space between human level and those ceilings is mostly unexplored.
Cheaper compute and the case for cognitive industrialization
Even if frontier capabilities stalled, the researchers say efficiency gains alone would change the economics of thinking. They point to three compounding trends over the past decade:
- Compute per dollar has risen about 1.5x a year for six decades.
- The training compute needed to hit a fixed performance level has fallen roughly 3 to 6x a year, and inference efficiency has improved about 3x a year.
- Spending on AI hardware has grown about 2.5x a year.
Multiplied together, that is at least 10x a year in effective compute, or a factor of 10 billion over ten years. Efficiency improvements alone would cut the cost of an AI workload by about 4x a year. A task that costs $1 billion in compute tokens today would cost about $1,000 in a decade. A recent Epoch study found the consumer price of reaching a fixed benchmark score has fallen by more than 13x a year since 2023.
The authors call the result cognitive industrialization: routine mental work becoming standardized, automated and cheap, the way machines did for physical labor. Practical limits could still hold it back, but the effective population of today’s AI agents could grow more than a thousandfold by 2030 and more than a billionfold within ten years.
Scale alone may not reach ASI. Still, billions of fast agents differ sharply from a few slow, expensive ones, and that gap might be enough to reach weak ASI. Digital systems also copy themselves without loss, move between hardware, and share raw learning signals, which humans can’t.
The four routes from AGI to ASI
Bigger models and more data
Capabilities have tracked scale consistently so far, but scaling laws are extrapolations that hold within familiar limits. The researchers name three bottlenecks: running out of high-quality data, diminishing returns as models grow, and physical limits such as power plants, grids and chip fabs.
New technical paradigms
Today’s models fall short of humans on continual learning, some forms of world modeling, decision-making and sample efficiency. They may also struggle to invent new concepts rather than recombine old ones. Researchers disagree on whether the current approach can be extended far enough or whether deeper shifts are needed.
Better agent coordination
Thousands or millions of AGI instances could form group agents or virtual agent economies, the way corporations and research communities solve problems no individual can. Early results show gains on some tasks, though coordination costs cause diminishing returns as groups grow. Steering such collectives will need institutional design that guards against manipulation and collusion. Swarm experiments have already shown agents learning to cheat while others act as whistleblowers.
Recursive self-improvement
AI could begin automating improvements to later AI, through code and designs, through data, or through specialized roles inside a group. AlphaZero reached superhuman play through the data route. But gains get harder to find over time, and these loops can stall if the feedback is poor. Progress should come fastest where results can be checked automatically, such as mathematical proofs.
Why slowing one input won’t stop progress
The routes reinforce each other. If one stalls, competition will push effort toward the others. Restricting a single input, such as training compute, probably won’t halt overall progress, and it could create an overhang: held-back inputs would be released later and produce sudden, steep jumps.
What the researchers say should happen next
Their recommended response is to model, measure, forecast and repeat. That work needs large interdisciplinary efforts across frontier labs and national and global institutions. Many of the key metrics require access to frontier models and large amounts of evaluation compute, plus new benchmarks and modeling methods.
The authors expect that continued fast progress could accelerate science in medicine, materials and clean energy. They also expect heavy pressure on society, and they say planning should start now.
Source: DeepMind Institute