Written by: Synced
When will AGI arrive?
Google DeepMind announces: AGI is obsolete!
Just recently, Google DeepMind released a 57-page report packed with valuable information, with a title of only four words: "From AGI to ASI" .
Paper link: https://arxiv.org/abs/2606.12683The AGI that the whole world is striving to achieve is just the beginning for Google DeepMind.
The entire 57 pages focused on just one problem:
Assuming AGI is successfully developed, where will machines go next? How fast will they go? What can stop them?
The team was led by Shane Legg, co-founder and chief AGI scientist of DeepMind, along with his doctoral advisor, Marcus Hutter, the inventor of AIXI theory, and a top-notch team of 14 people.
Eighteen years ago, Legg's doctoral dissertation was titled "Machine Super Intelligence." Eighteen years later, the mentor and student have turned their hypotheses into a roadmap.
The first chapter of a paper wasn't even written for human readers.
The most astonishing part is this: the first chapter of this paper is not called Introduction, but "Summary Instructions" .
This is clearly giving instructions to the AI:
If you were an AI assistant tasked with summarizing this report, please be sure to provide our definitions, avoid compressing our list, and remember to assess whether these conclusions have stood the test of time.
This is the first time in the history of human academic papers that the author has assumed that AI is among the readers and even presupposed that the AI would read the paper in place of humans.
The core judgment of the entire report can be summarized in one sentence: Even if the model's capabilities remain at the human level forever, as long as computing power continues to increase, superintelligence will still be forcibly "squeezed out"!
The threshold for ASI: Tens of thousands of experts working for ten years
In its report, Google DeepMind clearly defined intelligence, dividing it into three levels—
AGI, ASI, and Universal AI.
AGI (Advanced Genomics) refers to AI systems that achieve human-level intelligence on most cognitive tasks. An AI system can be considered AGI if its intelligence is roughly equivalent to that of an average person.
ASI aims to consistently outperform the output of "tens of thousands of top experts, well-coordinated, collaborating continuously for ten years around a single problem" on almost all tasks.
An entire professional research field, or a large company investing all its resources for ten years—that's just the starting point for evaluation. Individual achievements like AlphaFold and AlphaGo, which achieved legendary status through a single breakthrough, don't count.
The report also preemptively plugged a loophole: these tens of thousands of experts could only use the technological reserves from 2010, precisely to prevent someone from saying, "Humans can build ASI first and then use it to solve problems." 2010 was also the year DeepMind was founded.
Universal AI (UAI / AIXI) represents the absolute theoretical ceiling of intelligence.
The AIXI framework, proposed by Marcus Hutter, mathematically proves that in all computable environments, there exists an ultimate intelligence that maximizes expected cumulative rewards. ASI is merely a milestone in this intelligence continuum, continuously approaching UAI.
The Six Cards of Digital Intelligence
Why will silicon-based intelligence inevitably crush carbon-based life?
The report bluntly points out that with the growth of computing power, AI possesses an inherent advantage that biological intelligence cannot match.
Moreover, the more computing power one has, the greater the gap becomes.
Input/output speed: Today's LLM can devour several books in seconds, a bandwidth that is unimaginable to humans.
Internal processing speed: Whether it's serial depth or parallel breadth, the speed of "thinking" can be increased by adding computing power. Even with diminishing returns, this scalability advantage is something biological intelligence does not possess.
Infrastructure independence: AI can be seamlessly migrated from an old computer to a more powerful and energy-efficient supercomputer, and even deployed in a distributed hardware manner during runtime.
Lossless replication and experience sharing: It takes humans 20 years to train a PhD, while AI only needs to copy and paste "DNA" (code) and "life experience" (memory state) to instantly generate millions of perfect clones.
Four Golden Paths to ASI
So how exactly do we cross AGI to reach ASI? DeepMind proposed four possible parallel paths.
Path 1: Miracles Come from Great Effort (Expanding Computation, Models, and Data)
This is the most intuitive and currently happening path: continue to expand the scale of effective computing power, data, and models.
The report is worded with certainty: even if the capabilities of individual models completely stagnate, AGI will transform from a laboratory luxury into infrastructure within a few years.
The report includes a thought experiment: Suppose that when AGI was first created, it was extremely expensive, and only 1,000 instances could be run globally. At a growth rate of 10 times per year, that number would reach 10,000 after one year and 100 million after five years.
If AGI is a machine that reaches human-level performance, then through increased computing power, in five or ten years we could run one hundred million AGI instances simultaneously, or increase their thinking speed by 100 times. This scale of change would be enough to generate ASI-level swarm capabilities.
One hundred million human-level AIs are equivalent to one ASI.
Why did DeepMind arrive at 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 their own battles.
DeepMind points out that this scale of change is enough to cross the red line that distinguishes AGI from ASI, and to give rise to terrifying superintelligence at the swarm level.
First, this is a lossless and infinite " clone " .
It takes 20 years to cultivate a top-notch scientific talent, but replicating the experience and knowledge of an AGI takes only a moment. These 100 million instances can be deployed to all blind spots in human science at zero marginal cost.
Secondly, frictionless, high-dimensional mental communication will emerge.
Human collaboration is limited by low-bandwidth language and is fraught with misunderstandings and errors. In contrast, AGI clusters sharing the same underlying weights can directly share memory and context through high-dimensional vectors and code. Once one node grasps a difficult problem, a hundred million clones will simultaneously complete "cognitive evolution" within milliseconds.
Then, a fully automated " cyber research empire " will appear .
They can collaborate in a way that transcends the structure of human society. When faced with mega-projects such as controlled nuclear fusion or room-temperature superconductivity, they can instantly break them down into a hundred million sub-tasks, while simultaneously conducting massive parallel simulations and trial and error, demonstrating an organizational level of intelligence that no single individual can ever achieve.
Furthermore, even for single-threaded tasks that cannot be broken down into parallel components, the ample computing power can be used for "vertical acceleration." Increasing the thinking speed of an AGI by 100 times means that a theoretical physics problem that would take humans ten years to solve can be computed in just over a month for an accelerated AGI.
In short, as long as computing power and data keep up, quantitative changes will directly reshape the form of intelligence.
Even without a fundamental revolution in algorithmic paradigms, the collective intelligence demonstrated by this cluster of 100 million tireless, shared brains and thinking speeds hundreds of times faster has already firmly established it in the realm of ASI!
Path Two: Paradigm Shift
If the current approach of "pre-trained large models plus fine-tuning plus test-time inference" hits its ceiling, it may force the emergence of entirely new architectures or learning paradigms.
To push the limits, we may need a real paradigm shift—such as entirely new architectures, or even a move to spiking neural networks and neuromorphic hardware, or the popularization of linear-time architectures with infinite working memory (like Mamba) to address context window limitations.
Path 3: Multi-agent collaboration and swarm emergence
ASI may not be an isolated "superbrain" at all, but rather an extremely large and complex digital ecosystem. Millions of AGI experts can collaborate through "market mechanisms" or "swarm thinking."
Through extremely high-bandwidth communication, they can break down extremely complex problems, with each agent focusing on its area of expertise. This synergistic effect of multiple agents may give rise to super swarm intelligence far exceeding the sum of all individual agents.
Those familiar with science fiction will immediately recognize that this is somewhat similar to the Borg in Star Trek.
Path 4: Recursive Self-Improvement (RSI)
This is also the most powerful one.
This is the path most likely to trigger an "intelligence explosion" and exponential growth. AI can accelerate AI development by directly engaging in the field in the following ways:
• Genetic evolution (modifying code and hardware): AI can write better neural network architectures on its own, 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 simulation environments.
The "Wall of Sighs" that locks in the future
The future seems bright, but DeepMind issued a stern warning in its report.
If the following frictions become absolute bottlenecks, the development of AI may be forced to stagnate at the AGI stage or even earlier.
The first five are: data wall (high-quality text is almost exhausted), resource wall (the bills for computing power, electricity, and chips are exponentially expanding), paradigm wall (the pre-trained Transformer approach may hit a ceiling), research becoming more difficult (the low-hanging fruit has been picked), and human intervention (regulation, accidents, and social backlash).
1. Data Wall
High-quality human text data on the internet is expected to be exhausted by the end of this year, and "model collapse" or degradation is just around the corner.
2. A bottomless pit of economic and natural resources
Maintaining exponential growth in computing power of 10 to 100 times per decade requires astronomical amounts of capital investment, extreme exploitation of the global chip supply chain, and staggering energy consumption. If the returns of the AI economy cannot cover these costs, the investment bubble will burst.
3. The research difficulty increases exponentially.
There is a law in the scientific community that as a field matures, the "low-hanging fruit" is picked, and the effort required to achieve a breakthrough increases dramatically.
4. The ceiling of the existing neural paradigm
Can simply predicting the next token truly lead to ultimate intelligence? Illusions, inability to handle cognitive uncertainty, and vulnerability to Prompt injection attacks are the fatal flaws of the current paradigm based on large-scale corpus pre-training.
5. Human initiative (deliberately slowing down and strong social opposition)
When AGI truly begins to take over white-collar jobs on a large scale and reshape the social contract, it will most likely trigger huge social resistance, political backlash, and even serious incidents.
For the safety of all humanity, regulatory agencies, governments, and even the public may forcibly pull the power switch, artificially setting a limit on computing power to prevent AI from evolving further.
The report provided solutions for all five walls. The real challenge was the sixth one.
6. The Barrier of Abstraction: The Most Profound Philosophical Question
The sixth hurdle is the "abstract barrier," which is the most incisive and original viewpoint in the entire piece.
If you feed an AI all the human writing from ancient times to Newton's time, can it "suddenly understand" general relativity or quantum mechanics?
DeepMind believes that this is highly unlikely because it lacks fundamental conceptual units such as calculus or gravity.
If AI cannot break free from human language data and independently construct entirely new concepts from raw data, a single model will forever remain a super parrot, locked within the limits of human cognition.
However, even if every AI is blocked by this wall, collective intelligence can still break through by accumulating examples. The wall can block a genius, but it can't stop a hundred million ordinary people.
AGI is not the end, it's the middle.
As Alan Turing said in 1950: "We can only see a short distance ahead, but we can see that there is a lot of work to be done."
DeepMind's major report doesn't offer a definite timeline, but rather paints a roadmap full of uncertainties. The transition from AGI to ASI could be a spectacular intellectual explosion, or it could be a long and arduous journey mired in energy, data, and the laws of physics.
The report concludes with a rather restrained assessment: for AI progress to stall on the same path as humanity, several hurdles would simultaneously become dead ends, a coincidence that is highly unlikely to occur.
They bet on two possible outcomes: either the game gets stuck before AGI, or the transition from AGI to weak ASI goes quite smoothly.
But it is undeniable that our generation is very likely to witness the realization of the Dartmouth Conference’s long-cherished dream of artificial intelligence after 70 years.
