From AGI to ASI: What DeepMind's New Research Actually Says

The research paper titled „From AGI to ASI" was published on June 10, 2026, on the arXiv server and was signed by 14 scientists from Google DeepMind, including Shane Legg — co-founder of the lab, who popularized the term AGI himself 20 years ago. The document is 284 kilobytes and maps what has been missing in AI research so far: a systematic overview of possible paths from artificial general intelligence to superintelligence, including realistic obstacles on each of them.

Key definition from the document: AGI means a system that achieves at least median human performance across a wide range of cognitive tasks. ASI, on the other hand, is a system with general superhuman intelligence — that is, one that surpasses large groups (thousands) of human experts working over an extended period (years).

The authors openly admit that predicting the future is difficult. „Even if progress in AI continues far beyond AGI, it does not mean that ASI will be omnipotent," they write. According to them, superintelligence will not automatically mean the end of aging, brain uploading, or Dyson spheres — even though DeepMind's head, Demis Hassabis, speaks enthusiastically about these possibilities.

Four Ways AI Can Go Beyond Human Limits

The document identifies four possible routes from AGI to ASI, which can also complement each other:

1. Scaling Computational Power

The most straightforward path: simply add more chips, more electricity, and more money. Effective compute ("effective compute") can continue to grow roughly tenfold annually — thanks to better chips, larger training budgets, and software that gets more out of each chip. This is essentially a continuation of the current trend, just pushed beyond the horizon of human capabilities.

2. Swarm Intelligence

Even if individual models stop somewhere around the human level, we can run a huge number of them simultaneously. Coordinated multi-agent systems working 24/7, without interruption — this alone could create an entity that surpasses human organizations. The authors jokingly refer to the Borg from Star Trek: „They knew what they were doing."

3. Progress Flywheel

AI that starts helping to build better AI. This self-reinforcing cycle — a model designs a better architecture, which allows training a better model, which designs an even better architecture — could theoretically lead to an "intelligence explosion." This is a scenario described years ago by I. J. Good, and one that a part of the research community fears the most.

4. Fundamental Breakthrough

Not everything is about scaling. A new architecture or training method could emerge that changes the rules of the game, similar to what the Transformer did in 2017. The authors point out that relying on breakthroughs is not a strategy — but they cannot be ruled out either.

What Stands in the Way: Physics, Money, and Politics

The document is valuable precisely because it is neither optimistic nor catastrophic — it is realistic. It names specific frictions:

• Scaling Limits: There is only a limited number of chips, electricity costs money, and data centers cannot be built indefinitely.
• Exhaustion of Easy Gains: The self-improvement loop may slow down once the "low-hanging fruit" are gone.
• Coordination Problems: An army of AI agents may encounter bureaucratic chaos — something like an "agent organizational swamp."
• Regulation: Governments, pressured by a concerned public, can simply slow down development.

Also interesting is the thought experiment with physics. Imagine an AI trained on everything humanity knew about physics before Newton. Could it derive general relativity? The authors are skeptical: „It seems highly unlikely that a system could arrive at the laws of general relativity, let alone quantum mechanics, without the conceptual primitives of calculus, universal gravitation, or electromagnetism." In other words — today's models inherit concepts from humans. Whether a machine can extract a truly new idea from raw reality remains an open question, according to the authors.

Hassabis: AGI Will Arrive Around 2030. What Then?

In early June 2026, Google DeepMind CEO Demis Hassabis stated at the Stanford Graduate School of Business that AGI is a matter of "perhaps 2030, give or take a year." "I think it will be such a hugely transformative technology that it will effectively be a new human era," he added. At the same time, he admitted that "one or two major breakthroughs" are still needed — specifically in continuous learning, better memory, more efficient context windows, and long-term planning.

Hassabis's vision for the future is bold: he speaks of a "post-scarcity world," asteroid mining, nuclear fusion, and hydrogen extraction from seawater. According to Sebastian Mallaby's book The Infinity Machine, Hassabis believes that people „don't think ambitiously enough about what the world will look like after AGI."

At the same time, he warns — there is little time left for preparation. „Society needs to hear this because we don't have long to prepare for what this means," he said at Stanford. The future, according to him, „is still waiting to be written, but the next few years will be critical."

What This Means for Czechia and Europe

DeepMind's research comes at a time when the European Union is launching 35 new AI supercomputers with a total performance of 800 exaflops from Nvidia, intended to serve 3 million scientists. The Czech Republic plays a role in this infrastructure thanks to the Czech AI Factory in Ostrava, which has become one of the hubs of European artificial intelligence.

For Czech companies and developers, a simple lesson follows: the debate about superintelligence might sound like science fiction, but infrastructure, regulation, and research are already being prepared. The EU AI Act is already in force, the Czech National Bank is building its own AI center on Nvidia chips, and Czech startups like Ecomail or Poke are increasingly integrating with large language models. Whoever doesn't know what AGI means today will be left behind in five years — whether superintelligence arrives or not.

Honesty Instead of Hype: Why This Paper Is Different

Unlike many other documents from AI labs, which often serve as marketing tools, "From AGI to ASI" is exceptionally honest. It doesn't promise, threaten, or speculate about a date. Instead, it says: here are four paths, here are the obstacles, and here is a list of questions to which we don't yet know the answer. It is precisely this approach — "we can only see a little way ahead, but there's a lot of work to be done even there" — that makes the document essential reading for anyone who wants to understand where artificial intelligence is truly heading.