Google DeepMind Publishes Roadmap From AGI to Superintelligence

Google DeepMind has published an extensive report titled "From AGI to ASI," in which fourteen of the company's researchers describe how artificial intelligence could evolve after reaching human-comparable capability. The authors, who include DeepMind co-founder Shane Legg, argue that reaching AGI within the next decade has become a realistic goal for leading AI labs rather than a distant hypothesis, so it's time to seriously examine what comes next.

The report defines AGI as artificial general intelligence at the human level, and ASI as a system whose cognitive abilities exceed not a single person but entire human organizations. This distinction matters, because the authors do not treat the transition from AGI to ASI as one dramatic breakthrough moment, but as a process stretched over time and broken into many smaller, cumulative jumps in capability.

Four paths to superintelligence

The first path, scaling AGI, assumes that continued growth in compute power, training data volume and model size will on its own produce capabilities that surpass human organizations. The second, a paradigm shift, refers to fundamental architectural breakthroughs that go beyond today's approach built on transformers and training on massive datasets.

The third path is recursive self-improvement, meaning systems capable of improving their own code, architecture or training process without constant human involvement. The fourth, and least intuitive, holds that superintelligence won't appear in a single model but will emerge from the coordination of many AI agents operating together at scale, much as the capabilities of human organizations exceed those of individual workers.

Friction and bottlenecks

The authors stress that none of these paths is guaranteed or easy. The report examines barriers that could slow or block each of them, from the physical limits of available compute, to a shortage of high-quality data, to the difficulty of maintaining control over systems that modify their own behavior. The document also introduces the concept of "Universal AI" as a theoretical, far-off reference point for the entire continuum of machine intelligence, drawn from earlier theoretical work by co-author Marcus Hutter.

Rather than a single, dramatic "AGI moment," we should expect a series of accelerating, AI-assisted transformations - the report's authors, Google DeepMind

Why DeepMind is writing this now

The report's publication coincides with a string of statements from Google DeepMind CEO Demis Hassabis, who in recent weeks has called for an independent, international body to test and certify the most powerful AI models, modeled on financial oversight institutions. "From AGI to ASI" can be read as the technical case for those calls: if the paths to superintelligence are real and multiple, and control mechanisms remain untested, the authors argue that meeting this challenge requires "a massively interdisciplinary undertaking of global reach and significance."

That phrasing is no accident. The report's co-authors include not just engineers and machine learning theorists but also researchers focused on AI ethics and safety policy, such as Iason Gabriel and Allan Dafoe. This suggests DeepMind treats the question of ASI not as a purely technical one, but as a problem requiring the involvement of governments, scientific institutions and society well before such systems emerge.

What it means for the industry and Poland

For companies and investors tracking the race for ever more powerful models, DeepMind's report signals that one of the world's leading labs treats the transition from AGI to superintelligence as a scenario requiring preparation now, not distant science fiction. For Polish institutions working on AI regulation and safety, including the newly forming Centrum Implementacji Sztucznej Inteligencji (Center for AI Implementation) at the Polish Armed Forces and the Rada Przyszłości (Council for the Future) attached to the Prime Minister's office, the document offers a concrete, if speculative, vocabulary for discussing what oversight mechanisms will be needed before systems emerge that outperform entire organizations.

The report gives no specific date for the arrival of ASI, nor does it claim that any of the four paths is more likely than the others. The authors caution that uncertainty about the pace of progress is great enough that transformative change may arrive through a series of smaller breakthroughs spread over years rather than one sudden leap. That measured approach sets the document apart from more alarmist forecasts published in recent months by other AI safety research teams.

The full 57-page report is publicly available both on Google DeepMind's website and in the arXiv repository, allowing independent researchers to scrutinize its assumptions and critique the proposed development paths.