Google officially announced a 30,000-word roadmap: AI with 100 million human-level capabilities is ASI!

When will AGI arrive?

Google DeepMind announced: AGI is already outdated!

Recently, Google DeepMind released a substantial 57 - page report with a title of just four words: "From AGI to ASI".

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

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

Throughout the 57 - page report, it deduces just one question:

Assuming that AGI is truly developed, where will the machines go next? How fast will they progress? What can stop them?

The team is led by Shane Legg, the co - founder of DeepMind and the chief AGI scientist, along with his doctoral supervisor, Marcus Hutter, the inventor of the AIXI theory, and a top - notch team of 14 people.

18 years ago, Legg's doctoral thesis was titled "Machine Super Intelligence". 18 years later, the master and the apprentice turned their assumptions into a roadmap.

The first chapter of a paper

Is not even written for humans

The most astonishing move here is that 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 to summarize this report, please make sure to clarify our definitions, do not compress our lists, and remember to judge whether these conclusions can stand the test of time.

This is the first time in the history of human papers that the author assumes there are AIs among the readers and presupposes that AIs will read it on behalf of humans.

The core judgment of the whole report can be summarized in one sentence: Even if the capabilities of the model remain at the human level forever, as long as the computing power continues to increase, super - intelligence will still be "squeezed" out!

The threshold of ASI

Tens of thousands of experts working for a decade

In the report, Google DeepMind clearly defines intelligence, dividing it into three levels in total -

AGI, ASI, and Universal AI.

AGI reaches the median human level in most cognitive tasks. As long as an AI system has an intelligence level roughly equivalent to that of an ordinary person, it is AGI.

ASI must stably exceed the output of "tens of thousands of top experts, well - coordinated, and continuously collaborating on a single problem for a decade" in almost all tasks.

The efforts of an entire professional research field or a large - scale company going all - in for a decade are just the starting score. One - hit - wonder models like AlphaFold and AlphaGo do not count.

The report also preemptively plugs a loophole. These tens of thousands of experts can only use the technological reserves of 2010, preventing anyone from saying "humans can first create ASI and then use it to solve problems". 2010 is also the year when DeepMind was founded.

Universal AI (UAI / AIXI) is the theoretical absolute ceiling of intelligence.

The AIXI framework proposed by Marcus Hutter mathematically proves that in all computable environments, there exists an ultimate intelligence that can maximize the expected cumulative reward. ASI is just a milestone on this intelligence continuum that continuously approaches UAI.

The six advantages of digital intelligence

Why will silicon - based intelligence surely crush carbon - based organisms?

The report mercilessly points out that as computing power grows, AI has innate advantages that biological intelligence cannot match.

Moreover, the more computing power, the greater the gap.

Input/Output speed: Today's LLMs can absorb several books in a few seconds, a bandwidth that humans can't even imagine.

Internal processing speed: Whether in serial depth or parallel breadth, the "thinking" speed can be accelerated by increasing computing power. Even with diminishing returns, this scaling advantage is not possessed by biological intelligence.

Substrate independence: AI can seamlessly migrate from an old computer to a more powerful and energy - efficient supercomputer at will, and even perform hardware distributed deployment during operation.

Lossless replication and experience sharing: It takes 20 years for humans to train a doctor, while AI only needs to copy and paste the "DNA" (code) and "life experience" (memory state) to instantly generate millions of perfect clones.

Four golden paths to ASI

So, how exactly can we cross AGI and reach ASI? DeepMind proposes four possible parallel paths.

Path 1: Brute force creates miracles (expanding computing, models, and data)

This is currently the most intuitive and ongoing path: continue to expand the effective computing power, data, and model scale.

The report is very confident in its wording: Even if the capabilities of a single model completely stagnate, within a few years, AGI will change from a laboratory luxury to an infrastructure.

There is a thought experiment in the report: Suppose that when AGI is first developed, it is extremely expensive, and only 1000 instances can run globally. With a 10 - fold annual growth rate, there will be 10,000 instances after one year and 100 million instances after five years.

If AGI is a machine that reaches the human level, then through the growth of computing power, after five or ten years, we can run 100 million AGI instances simultaneously, or speed up their thinking by 100 times. This quantitative change in scale is sufficient to give rise to group capabilities at the ASI level.

One hundred million AIs at the human level are equivalent to an ASI.

Why did DeepMind come to this conclusion?

The reason is that if AGI is a machine that reaches the level of an ordinary person, then 100 million AGIs are not just 100 million "silicon - based workers" fighting independently.

DeepMind points out that this quantitative change in scale is sufficient to cross the red line dividing AGI and ASI, and give rise to terrifying super - intelligence at the group level.

First of all, this is a lossless and infinite "cloning clone".

It takes 20 years to train a top - notch scientific research talent, but it only takes an instant to copy the experience and knowledge of an AGI. These 100 million instances can be deployed at zero marginal cost to all blind spots in human science.

Secondly, there will be zero - friction high - dimensional mental communication.

Collaboration among humans is limited by the low - bandwidth language and text, full of misunderstandings and losses. The AGI cluster with the same origin has the same underlying weights, and they can directly share memories and contexts through high - dimensional vectors and codes. As long as one node has an epiphany about a difficult problem, 100 million clones will synchronously complete the "cognitive evolution" within milliseconds.

Then, there will be a fully automated "cyber scientific research empire".

They can collaborate in a mode that surpasses the human social structure. Facing a giant project like controlled nuclear fusion or room - temperature superconductivity, they can instantly break it down into 100 million sub - tasks, and conduct a large number of parallel deductions and trials and errors simultaneously, demonstrating organizational - level wisdom that a single individual can never achieve.

In addition, even for single - line tasks that cannot be disassembled in parallel, abundant computing power can be used for "longitudinal acceleration". Increasing the thinking speed of an AGI by 100 times means that a theoretical physics problem that humans need to spend ten years struggling with is only about a month's worth of computing for an AGI in an accelerated state.

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

Even if there is no fundamental revolution in the algorithm paradigm, relying solely on this cluster of 100 million tireless, brain - sharing, and hundreds of times faster - thinking entities, the collective wisdom demonstrated by its computing power network has already firmly entered the realm of ASI!

Path 2: Paradigm shift

If the current approach of "pretraining large models, fine - tuning, and inference during testing" hits the ceiling, it may force the emergence of a new architecture or learning paradigm.

To break through the limit, we may need a real paradigm shift - such as a completely novel architecture, or a shift to spiking neural networks and neuromorphic hardware, or the popularization of a linear - time architecture with infinite working memory (such as Mamba) to solve the context window limitation.

Path 3: Multi - agent collaboration and group emergence

ASI may not be an isolated "super - brain" at all, but an extremely large and complex digital ecosystem. Millions of AGI experts can collaborate through the "market mechanism" or the "swarm mind".

Through extremely high - bandwidth communication, they can break down extremely complex problems, and each agent is only responsible for the field it is best at. This multi - agent synergy may give rise to super - group intelligence far exceeding the sum of all individuals.

People familiar with science fiction will immediately recognize that this is a bit like the Borg Collective in "Star Trek".

Path 4: Recursive self - improvement (RSI)

This is also the most powerful path.

This is the path that is most likely to trigger an "intelligence explosion" and exponential growth. AI can accelerate AI research and development in the following ways:

· Genetic evolution (modifying code and hardware): AI can write better neural network architectures by itself, and even design more energy - efficient AI chips (such as what AlphaEvolve and FunSearch are