· via Hacker News – Front Page (native)
DeepMind essay maps the road from AGI to superintelligence via cognitive industrialization
An essay from DeepMind that reached Hacker News's front page argues that compounding compute trends could push AI past human level toward superintelligence, and urges the field to build rigorous ways to track progress beyond AGI.
DeepMind asks what follows human-level AI
An essay published by DeepMind, titled "What Comes After AGI?" and picked up on Hacker News's front page, argues that the field can no longer defer thinking about the step beyond artificial general intelligence. With AGI — machines matching humans across nearly all cognitive tasks — now plausibly within reach, the essay contends that AI could keep advancing at its current pace, or even accelerate, toward artificial general superintelligence (ASI), with an impact on science, the economy and society at greater scale and speed than AGI itself. Preparing for that shift, DeepMind writes, requires mapping the possible roads from AGI to ASI and building rigorous, reliable ways to track progress.
ASI as a point on a continuum
The essay treats general intelligence, biological or artificial, as a continuum on which human-level AGI is just one point. ASI sits further along: an advanced system that consistently outperforms large groups of human experts on tasks that would take them years, while using fewer resources and less time. The examples given include matching the scientific output of an entire research field in a fraction of the time, or achieving chip-design advances on par with what large companies deliver in a year. The bar must be met across virtually all tasks of human interest, not merely in narrow domains — where superhuman AI already exists in chess and Go, protein-structure prediction and theorem proving.
Known ceiling, uncharted middle
According to the essay, the very top of the intelligence continuum is well understood mathematically, along with hard physical, logical and complexity-theoretic limits — arbitrarily precise prediction of complex systems such as climates, immune systems or societies is out of reach. What those limits do not reveal is what lies just beneath them. Perfect chess play would demand more computing power than exists on Earth, yet a small dedicated chip can defeat any human. The vast region between human-level AI and the theoretical limits — spanning expert AGI and progressively stronger superintelligent systems — remains largely uncharted.
Cognitive industrialization
Even those skeptical that frontier capabilities will keep improving should weigh efficiency trends, the essay argues. It cites three compounding drivers from the past decade: compute per dollar has risen roughly 1.5x per year for six decades; the training compute needed to hit a fixed performance level has fallen roughly 3–6x per year, with inference efficiency improving about 3x per year; and spending on AI hardware has grown about 2.5x per year, with the largest frontier training runs growing more than 4x per year. Multiplied together, the global effective compute stock has grown at least 10x per year — a factor of ten billion over a decade. A decade from now, for every agent runnable today, there could hypothetically be more agents of equal quality than there are humans alive.
Efficiency gains alone would cut the compute demand of a given workload by about 4x per year, meaning cognitive work costing $1 billion in compute tokens today would cost roughly $1,000 in ten years. A recent Epoch study cited in the essay finds that consumer prices for achieving a fixed benchmark score have fallen more than 13x per year since 2023. DeepMind labels the outcome "cognitive industrialization": the standardization and automation of cognitive work across sectors, industrial-scale infrastructure build-out, and abundance of certain cognitive outputs — echoing how mechanical industrialization commoditized routine physical labor. Even with practical ceilings on compute build-out and more speculative ones on algorithmic efficiency, the effective population of today's agents might scale more than a thousand-fold by 2030 and more than a billion-fold within ten years.
When quantity becomes quality
Scaling alone might not even yield full AGI — some tasks could still require human help. Yet scale can be transformative in itself, the essay suggests: there is a major difference between human-level AIs that are slow and costly to run and a world where billions of agents operate faster than humans, and that quantitative leap might itself bridge the qualitative gap from AGI to weak ASI. Digital intelligence gains from compute in ways biology cannot: faster processing and data ingestion; working memory and memorization beyond human limits; substrate independence across machines; lossless, near-instant replication of both source code and accumulated experience; and high-bandwidth sharing of raw experience across a population. The essay then turns to four pathways from AGI to ASI, cautioning that some missing capabilities may be hard — or impossible — to patch over with sheer volume of agents.
Why it matters
If even a fraction of these compounding trends holds, the economics of cognitive work change within a decade rather than a generation: billion-dollar workloads becoming thousand-dollar line items would commoditize whole categories of professional and scientific labor. The essay also compresses the policy timeline — if AGI is near, the window to think seriously about superintelligence may be short, and the field currently lacks rigorous measures for progress beyond human level. Charting the uncharted middle between AGI and the theoretical limits of intelligence becomes, on DeepMind's reading, a research priority in its own right.
- #artificial-intelligence
- #deepmind
- #agi
- #superintelligence
- #compute