Google DeepMind Maps the Path From AGI to Superintelligence

Google DeepMind published a 60-page technical memo this week mapping the route from artificial intelligence that matches human performance to something categorically beyond it. The document, titled "From AGI to ASI," reads less like a research paper and more like a structured organizational forecast, written by people who treat the question as when, not whether.

The opening section is the best preview of the argument. Section 1 is labeled "Summary Instructions" and contains two paragraphs: one addressed to the human reader, one to the AI assistant they will likely use to process the document. It specifies what to cover, what to preserve verbatim, and what to flag if content changes after publication. Digit observed that this detail captures the paper's thesis before the argument even begins: artificial intelligence is no longer just the subject of these documents.

Shane Legg, co-founder of DeepMind and one of the researchers who helped popularize "AGI" as a working term in the field, is among the named authors. The definitions are accordingly precise. AGI means a system reaching median human performance across most cognitive tasks, roughly what a competent but unremarkable person can do. ASI is defined more demandingly: a system that outperforms not individual experts but large, well-coordinated teams of domain specialists, consistently, across almost any task, over extended periods of time.

Pathways and precedents

The authors outline four potential routes to ASI. Scaling is the most established: adding compute, data, and parameters has driven most capability gains in artificial intelligence over the past decade, and the paper treats it as a baseline, not a ceiling. Available excerpts do not enumerate all four pathways completely; the full breakdown requires reading the complete document, and Digit's coverage excerpts several key sections.

Notably, the paper does not predict when AGI will arrive. That choice is deliberate and stated: the authors treat AGI onset as a starting condition rather than a forecast target. Their questions are downstream from that threshold: how quickly do capabilities compound, where does human oversight remain meaningful, and which research directions accelerate the transition to ASI rather than stall before it.

For practitioners, the framing has a concrete implication. Most current evaluation frameworks, including red-teaming, capability benchmarks, and alignment audits, are calibrated against individual human performance. If ASI is defined as defeating coordinated expert teams, those benchmarks are measuring the wrong reference class. The paper's definition raises the effective bar for what alignment would need to mean once that threshold is crossed.

As ZDNET's model release tracker shows, 2026 has been relentless on the product side. The definitional groundwork in the DeepMind paper may matter more to safety researchers than any single benchmark released this year, precisely because definitions determine what gets measured and funded.

Reading the source

There is a longer history of labs publishing forward-looking memos on AI trajectories, and those documents have shaped funding priorities and policy for years afterward. DeepMind's paper carries more institutional weight than most, given Legg's role in establishing the vocabulary now used across the field; Humanity Redefined has noted Google's broader push alongside releases like Gemini 3 Flash and AlphaEvolve, and this memo signals that the theoretical scaffolding is being built in parallel with the product pipeline.

Such prominence is also a reason for calibrated reading. A 60-page document from a Google subsidiary on the road to ASI is not a neutral artifact. It reflects the lab's own research agenda, funding relationships, and competitive position. None of that invalidates the analysis, but it should inform how its definitions get adopted elsewhere.

The real question is not whether DeepMind's four pathways are exhaustive. It is whether their ASI definition, outperforming coordinated expert teams rather than individuals, becomes the working threshold the broader field measures against. Definitions shape what gets funded, what counts as progress, and where safety research focuses. This paper is an opening bid on those terms.

FAQ

What is the difference between AGI and ASI in DeepMind's paper?
AGI is defined as AI matching median human performance across most cognitive tasks. ASI must consistently outperform large, coordinated teams of domain experts over extended periods on nearly any task, a significantly higher bar than individual human comparison.

Why does the paper include instructions for AI assistants in Section 1?
The authors added a paragraph addressed to AI summarization tools specifying what to cover and what to flag if the document is updated. It functions as both practical documentation and a demonstration of the paper's central claim about how technical content is now consumed.

What are the four pathways to superintelligence?
Scaling is the one explicitly named in widely available excerpts: adding compute, data, and parameters. The complete set of four requires reading the full 60-page document.

Does this paper change how capability evaluations should be run?
If ASI is defined relative to expert teams rather than individuals, single-human-level benchmarks are the wrong reference class. That is a practical implication worth taking seriously now, regardless of where one stands on timeline predictions.