The people who defined AGI just mapped what comes next: DeepMind's four pathways "From AGI to ASI"

The people who defined AGI have, for the first time, formally mapped what comes after it. On June 10, 2026, DeepMind published a 57-page paper, "From AGI to ASI." It is led by Shane Legg (DeepMind co-founder) and Marcus Hutter (inventor of AIXI), and it lays out four pathways from AGI to artificial superintelligence (ASI). Yet even ASI, it argues, is bound by fundamental physical and computational limits. ASAP summarizes the roadmap from the primary source.

Who wrote it is the signal

"From AGI to ASI" is a 57-page paper written by the researchers who defined the concept of AGI. DeepMind published it on June 10, 2026, led by Shane Legg (DeepMind co-founder and Chief AGI Scientist) and his PhD supervisor Marcus Hutter (inventor of AIXI). The authorship itself is the signal. The person who gave the term AGI its weight, and the person who pinned an ideal intelligent agent to an equation in AIXI, put their names on one document. This is not a startup marketing whitepaper or an outsider's forecast; it is the camp that owns the definitions charting the next phase itself, which gives the roadmap a standing ordinary predictions do not have.

Why the four pathways should be read together

The paper's pathways are continued scaling of effective compute (which the authors estimate at about 10x a year), non-linear paradigm shifts in architecture, recursive self-improvement, and multi-agent collectives. The easily missed point is the claim that the four are not mutually exclusive and can compound. This is not a story about clearing a single bottleneck; it envisions the result when several forces push at once. Even the 10x figure reads better as a coordinate for the slope of the scaling axis than as a lone prophecy. Misreading the pathways as exclusive scenarios loses the compounding effect that sits at the paper's core.

Why "smarter than organizations" is the benchmark

The paper defines ASI as a system more capable than large organizations of humans, not one person or one team. It means cognition that exceeds an entire organization with its division of labor, specialized expertise, and collective memory. Declining to use a single genius as the yardstick is interpretively loaded. If the ceiling of intelligence is set at a collaborating collective rather than one individual, the debate moves from "can AI surpass Einstein" to "can AI replace an entire company or lab wholesale." Change the tick marks on the benchmark and the picture of the threat model and the regulatory target shifts with it.

What is left for practitioners

This is less a new product or benchmark score than a document that moved the coordinates of the debate. It will not, on its own, decide deployment or adoption in the near term. But in a climate thick with overheated stories selling superintelligence, the fact that the concept's own creators named physical and computational limits acts as ballast for practical judgment. Simply remembering that the roadmap has physical ceilings built in is enough to make one question a plan drawn on the assumption of unbounded growth.

Stated limits, and the questions that remain

The paper holds that even ASI is bound by fundamental physical and computational limits. The examples include the speed of light, thermodynamic limits on computation, complexity theory, and Godel's incompleteness results. It does not get infinitely smart; there are walls it cannot cross. The caveat worth noting is that these limits draw a ceiling but do not guarantee the timing or feasibility of reaching it. Saying a wall exists and saying we actually arrive at the base of that wall are different claims. What the paper answers is closer to the former; the latter remains an open question.

Source: ASAP summary of DeepMind's "From AGI to ASI" (arXiv 2606.12683, June 10, 2026; Shane Legg and Marcus Hutter, 57 pages, four pathways to ASI, physical and computational limits).