Most of the conversation about AI stops at AGI — artificial general intelligence, roughly human-level. But what happens after we get there? A new report from Google DeepMind, From AGI to ASI, written by Shane Legg and a dozen colleagues, takes that question seriously. Here’s a short tour of what it says.
Three points on one line
The report frames machine intelligence as a single continuum rather than a set of discrete stages:
- AGI — a system roughly as capable as a single human across most cognitive tasks.
- ASI (artificial superintelligence) — a system that surpasses not just an individual expert but large organizations of human experts, across virtually every domain.
- UAI (universal AI) — the theoretical endpoint of the line, a mathematically ideal agent (AIXI). It’s provably the most capable possible — and provably impossible to actually build.
The point of the continuum is that “AGI” and “ASI” aren’t natural boundaries; they’re useful markers on a scale that, in principle, runs all the way to UAI.
Four ways to get from AGI to ASI
The heart of the report is four possible paths beyond human level:
- Scaling. Just keep growing compute, data, and model size. Effective compute has been rising roughly 10× per year. Even if a single AGI plateaus, millions of fast copies running in parallel might already amount to ASI.
- Paradigm shifts. A genuinely new architecture or training method — the kind of breakthrough that, by definition, is hard to predict in advance.
- Recursive self-improvement. AI that helps build better AI, which then builds better AI still — a feedback loop that could, in principle, accelerate sharply.
- Multi-agent collectives. Superintelligence emerging not from one model but from large populations of coordinating agents, the way human institutions are smarter than any single person.
What might slow it down
The report is careful to list the frictions, too: the data wall (high-quality training text running out this decade), limits on energy and chips, the possibility that today’s neural paradigm simply isn’t enough, research getting harder as the easy wins are used up, an abstraction barrier (AI may struggle to invent genuinely new concepts rather than recombine human ones), and deliberate slowdown from regulation or backlash.
The deepest bottleneck: can AI flip the board?
Of these, the abstraction barrier is the one I keep thinking about. The paper’s worry — drawn from Neil Lawrence’s “embodiment factor” argument — is that because today’s models are trained on human-generated data, they inherit our concepts and may never forge genuinely new ones from raw experience.
A clean way to sharpen this — which I’m borrowing from a sharp video walkthrough of the paper — is Margaret Boden’s three kinds of creativity: combinational (recombine familiar ideas), exploratory (find new moves inside a fixed framework), and transformational (overturn the framework itself). AlphaGo’s famous “move 37” was exploratory — breathtaking, but it never rewrote the rules of Go. Einstein replacing Newton was transformational — he flipped the board. The open question, one Demis Hassabis has raised in public talks, is whether scaling today’s AI ever reaches that third level, or whether it stays a brilliant explorer of frameworks humans have already drawn.
If it stays an explorer, then no amount of speed or scale yields a true superintelligence — only a very fast, very thorough human-level one.
The takeaway
What I found most interesting is the report’s restraint. These are among the people with the most reason to expect a dramatic, sudden takeoff — yet they explicitly doubt the picture of “a single transformative step.” They think it’s at least as likely that we see a series of transformative changes across science and technology, and they stress that the uncertainty is enormous: forecasts of AI progress, they warn, “may not be very informative.”
It’s a measured, map-not-prediction kind of document — an attempt to lay out the terrain past AGI rather than to claim it knows what’s there.