DeepMind Report Maps Four Paths from AGI to Superintelligence (ASI)

In February, DeepMind CEO Demis Hassabis warned at an AI summit in India that AGI could trigger an industrial‑revolution‑scale impact within a decade, and he suggested AGI might arrive within five years. Beside him, co‑founder Shane Legg has publicly maintained for 17 years a 50 % chance that AGI will appear by 2028.

The newly published 57‑page report titled From AGI to ASI shifts the focus from "when" AGI will arrive to "what" will happen after we build systems with human‑level cognition. The authors define ASI as a system that, across virtually all tasks and domains, outperforms the collective expertise of tens of thousands of trained human specialists.

The report first highlights the asymmetry between digital and biological intelligence: human brains are carbon‑based, speed‑limited, and non‑replicable, whereas AI can ingest information at ever‑increasing bandwidth, read an entire book in seconds, scale processing speed with compute, and be perfectly duplicated—including its weights—into millions of instances within hours. Effective AI compute, which combines hardware progress, investment growth, and algorithmic efficiency, is estimated to increase roughly tenfold each year, a rate that has been stable for the past decade.

Four paths to ASI are described:

Continuous scaling : larger models, more data, and more compute. If "more compute equals higher intelligence" holds (as with chess engines), a quantitative increase could trigger a qualitative leap. The report sketches an extreme scenario where a human‑level AGI initially runs only 1,000 instances due to cost, but ten‑fold annual compute growth would allow 100 million instances or a 100× speedup within five years.

Algorithmic paradigm shift : the current paradigm of massive pre‑trained Transformers with fine‑tuning may not suffice for ASI. Researchers are exploring continual learning without forgetting, agents that make reliable decisions in open environments, and entirely new training regimes based on neuromorphic hardware or reinforcement learning.

Recursive self‑improvement : AI assists in improving AI, creating a feedback loop that accelerates progress. The authors liken this to simultaneous gene, cultural, and division‑of‑labor evolution. AlphaZero’s self‑play loop that refines its intuition is cited as a prototype, and newer "AI scientist" systems such as AlphaEvolve that discover novel algorithms suggest the loop could become an engineering reality.

Multi‑agent collaborative emergence : thousands of specialized AGI instances could form a cooperative network that decomposes complex tasks, exchanges results at high bandwidth, and collectively exceeds any single agent. The report introduces the notion of "cognitive division of labor" and even jokes about an "AGI CEO" that can directly communicate with every employee instance, eliminating bureaucratic friction.

Six possible slowdown bottlenecks are also listed:

Data wall : high‑quality web text is nearing exhaustion; synthetic data without careful curation may cause model collapse.

Economic and resource limits : scaling demands ever‑greater investment, chips, energy, and data‑center capacity; speculative solutions like orbital data centers raise ecological and safety concerns.

Neural‑network paradigm ceiling : even massive pre‑trained models with post‑processing may fail to reach AGI due to architectural limits.

Diminishing research returns : studies by Bloom and others show per‑researcher output declining over time; however, AI‑driven automation could offset this by replacing many researchers with compute.

Abstract barrier : current AI operates within human‑derived abstraction frameworks and may lack the ability to invent fundamentally new concepts without direct interaction with the physical world.

Active slowdown : regulatory actions, public opinion, major accidents, and geopolitical competition could deliberately or unintentionally curb AI progress.

The report concludes with a Turing quote: "We can only see a short distance ahead, but we can see that there is a great amount of work to be done," emphasizing that when AI approaches human‑level capability, preparation—not prediction—is the priority.