DeepMind essay: AI could move from AGI to ASI as effective compute grows at least 10x per year

TL;DR

In an essay from the DeepMind Institute, Tim Genewein argues that AI could continue advancing, possibly at an accelerating pace, from human-level AGI to artificial superintelligence (ASI). The essay puts effective compute growth at at least 10x per year and describes a coming "cognitive industrialization."

4 min read
0

A Google DeepMind essay published on the DeepMind Institute site argues that AI development is on pace to move beyond human-level AGI to artificial superintelligence (ASI). Author Tim Genewein writes that the outcome is "far from certain," but that AI could keep advancing at its current pace, and possibly faster, with a scale and speed of impact on science, the economy and society greater than AGI itself.

How the essay defines ASI

The essay treats general intelligence as a continuum, with human-level AGI as one point on it. ASI is defined as a system that consistently outperforms large groups of human experts on tasks that could take them years, using fewer resources and less time. The essay's examples are making the same scientific progress as an entire research field in a fraction of the time, or matching a year of large-company chip design progress.

ASI would have to meet that bar across virtually all tasks of human interest, not only in narrow domains. The essay notes that some systems already beat humans in narrow areas such as chess, Go, protein structure prediction and theorem proving.

The essay says the theoretical upper limits of general intelligence are well understood mathematically, and that those limits lie far beyond both today's AI systems and humans. The space between human-level AI and those limits is, in its words, "largely uncharted."

The compute argument

The central argument is that effective compute is growing through three compounding factors, based on the last ten years:

  • Hardware: compute per dollar has increased roughly 1.5x per year for six decades.
  • Algorithmic efficiency: training compute needed to reach a fixed performance level has fallen roughly 3-6x per year over the last decade. Inference efficiency has improved roughly 3x per year.
  • Investment: spending on AI hardware has grown roughly 2.5x per year. Training compute of the largest frontier runs grows by more than 4x per year.

Multiplying the factors, the essay puts growth in global effective compute stock at at least 10x per year, a factor of 10 billion over ten years. If the trend held, it says, a decade from now each AI agent that can run today could be matched by more agents of the same quality than there are people in the world.

On cost, the essay says efficiency gains alone would cut the compute needed for an AI workload by about 4x per year. A cognitive workload costing $1 billion in compute tokens today would cost $1000 in ten years. It also cites a recent Epoch study finding that the consumer price of reaching a fixed benchmark score has fallen by more than 13x per year since 2023.

The essay calls the resulting shift "cognitive industrialization," analogous to mechanical industrialization commoditizing physical labor. It says practical compute build-out ceilings and more speculative algorithmic ceilings might still allow the effective population of today's AI agents to grow more than a thousand-fold by 2030 and more than a billion-fold within ten years. It states this is "certainly not a given."

Scale versus capability

The essay concedes that scaling today's AIs in speed and number alone might not even yield full AGI, since some cognitive tasks may still need human help. It argues, however, that if broadly human-level AI is achieved, a world of billions of agents acting much faster than humans differs qualitatively from one with slow, expensive ones. That leap might be enough to bridge the gap from AGI to (weak) ASI.

It lists digital advantages over humans: processing speed, large working memory and memorization, substrate independence, lossless replication, and high-bandwidth sharing of experience. The essay then introduces four pathways from AGI to ASI, and says that some missing capabilities may be hard or impossible to patch over with sheer volume.

What this means

The essay frames ASI planning as a near-term question because AGI is, in the author's view, "within reach." Its argument does not rest on a single breakthrough: even with frontier progress stalled, continued efficiency gains in hardware and software would sharply raise the number of agents that can run and lower the cost of cognitive work.

The essay also separates two claims. Cost and scale trends are presented as extrapolations from the last decade, with explicit caveats about compute and algorithmic ceilings. Whether scale alone produces ASI is left open. For researchers and builders, the stated priority is mapping the possible roads from AGI to ASI and building rigorous, reliable ways to track progress along them.

Related Articles

analysis

Ethan Mollick: Anthropic's Claude Fable 5.2 is "almost" AGI, with gaps in taste and self-directed judgment

Ethan Mollick, writing in One Useful Thing, says Anthropic's Claude Fable 5.2 is "almost" AGI, which he calls Jagged AGI. After early access, he reports expert-level results on hardware hacking, business planning and research-heavy web projects, with persistent gaps in taste and judgment.

analysis

OpenAI report: median researcher's agent usage would cost $2,400 per day at API list prices in October

OpenAI's second internal research acceleration report says the median researcher would have consumed $2,400 of tokens per day at API list prices in October, up 62 percent from September. Agent success rates on longer tasks have risen, but the systems still fail at complex engineering work and need close human oversight.

analysis

Meta employees warn Muse AI may mix up conversations across accounts and expose private data, Reuters reports

Reuters reports that internal posts show Meta employees raising concerns that Muse could be mixing up conversations between different accounts and exposing users' private data. Muse has been downloaded by more than 10 million people, and Meta expects it to generate billions.

model release

Gemini 4 Neon scores 61 on Artificial Analysis Intelligence Index, 3rd overall, at $3/$15 per 1M tokens

Google DeepMind has released Gemini 4 Neon, its fourth model in seven weeks. Artificial Analysis measures it at 61 on the Intelligence Index (xhigh), 3rd overall, with token pricing of $3/$15 per 1M input/output tokens and a 1M-token context window.

Comments

Loading...