Google DeepMind has released a 57-page report titled "From AGI to ASI," stating that AGI is merely a starting point. The report defines ASI as intelligence surpassing the collective output of tens of thousands of top experts working together for a decade, and outlines six inherent advantages of digital intelligence over biological intelligence, including processing speed, lossless replication, and shared experience.

Author、Source: 36Kr

When will AGI arrive?

Google DeepMind announces: AGI is already outdated!

Recently, Google DeepMind released a comprehensive 57-page report titled simply: “From AGI to ASI.”

Paper URL: https://arxiv.org/abs/2606.12683

AGI, which the whole world is desperately trying to achieve, is just a starting point for Google DeepMind.

Fifty-seven pages entirely devoted to exploring one question:

If AGI is truly developed, where will machines go next? How fast will they progress? What could stop them?

Led by Shane Legg, co-founder and Chief AGI Scientist at DeepMind, along with his PhD advisor and AIXI theory creator Marcus Hutter, plus a top-tier team of 14 members.

18 years ago, Legg’s doctoral thesis was titled “Machine Super Intelligence.” 18 years later, the mentor and student turned the hypothesis into a roadmap.

The first chapter of a paper is not written for humans.

The most astonishing thing is here: the first chapter of this paper is not called "Introduction," but "Summary Instructions."

This is clearly giving instructions to AI:

If you are an AI assistant called upon to summarize this report, be sure to state our definitions, do not compress our list, and remember to evaluate whether these conclusions have stood the test of time.

This is the first time in the history of human academic papers that the author assumes the readers include AI and presumes AI will read it on behalf of humans.

The core judgment of the entire report can be summarized in one sentence: Even if model capabilities remain forever at the human level, as long as computing power continues to grow, superintelligence will still be forcibly "squeezed" out!

The threshold for ASI requires tens of thousands of experts working for ten years.

In the report, Google DeepMind provided a clear definition of intelligence, divided into three levels—

AGI, ASI, and Universal AI.

AGI, achieving median human performance on most cognitive tasks. An AI system is AGI if its intelligence is roughly equivalent to that of an average person.

ASI must consistently outperform the output of tens of thousands of top experts, working together in coordinated, continuous collaboration on a single problem for a decade.

An entire professional research field and a large company betting everything for a decade—this is just the baseline. Even single-point breakthroughs like AlphaFold or AlphaGo don’t count.

The report also preemptively closed a loophole: the tens of thousands of experts could only rely on technological capabilities from 2010, specifically to counter the claim that “humans could first build an ASI and then use it to solve the problem.” 2010 was also the year DeepMind was founded.

Universal AI (UAI / AIXI) represents the theoretical absolute peak of intelligence.

The AIXI framework, proposed by Marcus Hutter, mathematically proves that there exists an ultimate intelligence capable of maximizing expected cumulative reward in all computable environments. ASI is merely a milestone along the continuum of intelligence, continuously approaching UAI.

Six cards of digital intelligence

Why will silicon-based intelligence inevitably surpass carbon-based life?

The report bluntly points out that, as computing power grows, AI possesses inherent advantages that biological intelligence cannot match.

Moreover, the greater the hash power, the larger the difference.

Input/output speed: Today’s LLMs can ingest several books in seconds—a bandwidth unimaginable to humans.

Internal processing speed: Whether in terms of serial depth or parallel breadth, the speed of "thinking" can be accelerated by increasing computational power. Even with diminishing returns, this scalability advantage is beyond the reach of biological intelligence.

Hardware independence: AI can seamlessly migrate from an old computer to a more powerful, energy-efficient supercomputer, and even be distributed across hardware during operation.

Lossless replication and experience sharing: It takes humans 20 years to cultivate a PhD, but AI can instantly generate millions of perfect clones by simply copying and pasting the "DNA" (code) and "lifetime experience" (memory state).

Four Golden Paths to ASI

So, how exactly do we cross AGI to reach ASI? DeepMind has proposed four potential pathways that could occur in parallel.

Path One: Brute Force Yields Miracles (Scaling Computation, Models, and Data)

This is currently the most intuitive and actively unfolding path: continue expanding effective computing power, data, and model scale.

The wording of the report is confident: even if the capabilities of individual models plateau, AGI will transition from a laboratory luxury to infrastructure within a few years.

The report includes a thought experiment: suppose that when AGI is first created, it is extremely expensive, and only 1,000 instances can be run globally. With an annual growth rate of 10 times, there would be 10,000 after one year and 100 million after five years.

If AGI is a machine that reaches human-level intelligence, then through increased computational power, within five or ten years we could run a hundred million AGI instances simultaneously, or accelerate their thinking speed by 100 times. This scale of quantitative change alone would be sufficient to give rise to ASI-level collective capabilities.

One hundred million human-level AIs are themselves equivalent to an ASI.

Why did DeepMind arrive at this conclusion?

The reason is that if AGI is a machine that reaches the level of an average human, then a hundred million AGIs are far more than just a hundred million individual "silicon-based workers" working in isolation.

DeepMind points out that this scale of quantitative change is sufficient to cross the line separating AGI from ASI, giving rise to a terrifyingly powerful superintelligence at the collective level.

First, this is a lossless and infinite "clone duplicate".

It takes 20 years to cultivate a top-tier scientific researcher, but replicating the experience and knowledge of an AGI takes just an instant. These hundred million instances can be deployed at zero marginal cost to every blind spot in human science.

Second, there will be frictionless high-dimensional mental communication.

Human collaboration is limited by low-bandwidth language and text, riddled with misunderstandings and losses. In contrast, an AGI cluster with shared origins has identical underlying weights and can directly share memories and context through high-dimensional vectors and code. When one node achieves insight into a difficult problem, a hundred million clones will synchronize their “cognitive evolution” within milliseconds.

Then, an entirely automated "cyber science empire" will appear.

They can collaborate in a manner that transcends human social structures. Faced with massive engineering challenges such as controlled nuclear fusion or room-temperature superconductivity, they can instantly break them down into a hundred million subtasks, simultaneously conducting massive parallel simulations and trials, demonstrating organizational intelligence far beyond the reach of any single individual.

Additionally, even for single-threaded tasks that cannot be parallelized, abundant computing power can be used to achieve "vertical acceleration." Increasing an AGI's thinking speed by 100 times means that theoretical physics problems requiring humans a decade of intense effort would only amount to a few months of computation for the accelerated AGI.

In short, as computing power and data keep up, "quantitative change" will directly reshape the form of intelligence.

Even without a fundamental revolution in algorithmic paradigms, the collective intelligence demonstrated by this network of a hundred million tireless, brain-sharing, and hundred-times-faster-thinking agents is already firmly entering the realm of ASI!

Path Two: Paradigm Shift

If the current approach of "pre-trained large models plus fine-tuning plus inference at test time" hits a ceiling, it may force the emergence of entirely new architectures or learning paradigms.

To push beyond limits, we may need true paradigm shifts—such as entirely novel architectures, a shift toward spiking neural networks and neuromorphic hardware, or the adoption of linear-time architectures with infinite working memory (like Mamba) to overcome context window limitations.

Path Three: Multi-Agent Collaboration and Collective Emergence

ASI may not be an isolated "superbrain" at all, but rather an extremely large and complex digital ecosystem. Millions of AGI experts could collaborate through "market mechanisms" or "swarm intelligence".

Through high-bandwidth communication, they can break down extremely complex problems, with each agent responsible only for its own area of expertise. This collaborative synergy among multiple agents may give rise to a supergroup intelligence far exceeding the sum of all individual capabilities.

Fans of science fiction will immediately recognize this as reminiscent of the Borg Collective from Star Trek.

Path Four: Recursive Self-Improvement (RSI)

This is also the most powerful one.

This is the most straightforward path to trigger a "smart explosion" and exponential growth. AI can accelerate AI research directly in several ways:

· Genetic evolution (modifying code and hardware): AI can autonomously write better neural network architectures and even design more energy-efficient AI chips (as AlphaEvolve and FunSearch are already doing).

· Cultural evolution (data-driven self-improvement): Similar to AlphaZero, AI can generate, filter, and refine higher-quality training data through self-play and testing in simulated environments.

The Wall of Sighs locking the future

The future seems bright, but DeepMind issued a stern warning in its report.

If these friction points become absolute bottlenecks, AI development may be forced to stall even before reaching the AGI stage.

The first five barriers are: the data wall (high-quality text is running out), the resource wall (bills for computing power, electricity, and chips are exploding exponentially), the paradigm wall (the pre-trained Transformer approach may have hit its limits), research becoming harder (the low-hanging fruit has been picked), and human-induced brakes (regulation, accidents, and social backlash).

1. Order Book

High-quality human-generated text data on the internet is expected to be exhausted by the end of this decade, with model collapse or degradation just around the corner.

2. Economic and natural resource black hole

Sustaining an exponential growth in computing power by a factor of 10 or even 100 per year requires astronomical capital investment, extreme strain on the global semiconductor supply chain, and staggering energy consumption. The economic returns from AI cannot cover these costs, and the investment bubble will burst.

3. The difficulty of research increases exponentially

In the scientific community, there is a law that as a field matures, the "low-hanging fruit" are picked, and the effort required to achieve breakthroughs increases sharply.

4. The ceiling of existing neural paradigms

Can truly reaching ultimate intelligence be achieved merely by predicting the next token? Hallucinations, inability to handle epistemic uncertainty, and susceptibility to prompt injection attacks are fatal genetic flaws of the current large-scale corpus pre-training paradigm.

5. Human active decision-making (intentionally slowing down and strong social opposition)

When AGI truly begins to大规模 take over white-collar jobs and reshape the social contract, it will very likely trigger significant social resistance, political backlash, or even catastrophic incidents.

For the safety of all humanity, regulators, governments, and even the public may forcibly shut down the power, artificially cap computational power, and prohibit further AI evolution.

All five walls have solutions provided in the report. The real challenge is the sixth.

6. Abstract Barrier: The Deepest Philosophical Question

The sixth challenge is the "Abstract Barrier." It is the most incisive original insight in the entire piece.

If all human writings from ancient times up to Newton’s era were fed into an AI, could it independently "realize" general relativity or quantum mechanics?

DeepMind believes: it is extremely unlikely, as it lacks fundamental conceptual primitives such as calculus or gravity.

If AI cannot脱离 human corpora and independently construct entirely new concepts from raw data, a single model will forever be a super parrot, trapped within the upper limits of human cognition.

However, even if each AI is held back by this wall, collective intelligence can still push through by stacking examples. A wall can stop one genius, but not a hundred million ordinary people.

AGI is not the end, but the midpoint

As Alan Turing said in 1950: "We can only see a short distance ahead, but we can see that there is much to be done."

DeepMind's landmark report does not provide a definitive timeline, but instead paints a roadmap filled with uncertainty. The journey from AGI to ASI could be a spectacular intellectual explosion—or a long, arduous trek bogged down by energy, data, and the laws of physics.

At the end of the report, a notably restrained judgment was left: for AI progress to halt at the human level, multiple checkpoints would need to become dead ends simultaneously—a coincidence unlikely to occur.

They are betting on one of two outcomes: either they stall before AGI, or they progress smoothly from AGI to weak ASI.

But undeniably, our generation is very likely the one to witness the realization of artificial intelligence’s long-held aspiration since the Dartmouth Conference 70 years ago.