AGI is not the end point. In a new paper from DeepMind: Moving toward ASI, the true progress of AI has only just begun.
If Artificial General Intelligence (AGI) were to be realized tomorrow, what would the next stage of AI look like?
The Google DeepMind team and its collaborators proposed in their latest research report that AGI is likely not the end. In their view, AI will not stop at a level close to that of humans but will continue to become stronger, surpassing the top human expert teams, and ultimately evolving into Artificial Super Intelligence (ASI).
As Alan Turing wrote in 1950: “We can only see a short distance ahead, but we can also see that there is a great deal of work still to be done.”
In this report, the research team outlined four potential paths for AI to transition from AGI to ASI, the possible key bottlenecks, and the most worthy research questions to pursue.
Paper link: https://arxiv.org/abs/2606.12683
The research team stated that due to the large uncertainties in predicting the progress of ASI, it is currently impossible to rule out the possibility that AI will continue to develop at an accelerating pace in the next few years. This may mean that the scenario of a “single transformative leap” caused by introducing human - level AGI into society may not be accurate.
A more appropriate prospect may be that AI - driven progress and breakthroughs will emerge successively in many fields of science and technology, leading to a series of transformative social changes.
To address this prospect, a large - scale interdisciplinary project with a global perspective and broad concerns is needed.
After AGI, Comes ASI
Before discussing how AI will continue to become stronger, the research team first distinguished three concepts that are easily confused: AGI, ASI, and UAI.
AGI (Artificial General Intelligence): A general intelligence system that reaches the median human level in most cognitive tasks. It corresponds to the general cognitive ability of an ordinary person, rather than the level of top experts. The research team also pointed out that the first - generation AGI may have surpassed humans in some tasks, but it does not yet have sufficient generality.
ASI (Artificial Super Intelligence): It does not just surpass humans in a few tasks but overall exceeds humans in almost all fields that humans care about; the corresponding reference object is not a single expert but a large - scale, well - coordinated collective of human experts.
UAI (Universal Artificial Intelligence): The theoretical upper bound of machine intelligence, formally described by the AIXI framework. AIXI corresponds to a theoretically optimal general intelligent agent. Real - world AI can only gradually approach this upper bound and cannot achieve it directly.
Meanwhile, the research team pointed out that there may be more than one path for AGI to evolve into ASI, and they proposed four possible parallel paths, as follows:
Path 1: Continue to expand computing, models, and data
This path continues the basic logic of AI progress in the past decade, including more powerful hardware, larger training runs, higher algorithm efficiency, larger models, and more data. The research team pointed out that the “effective computing power” in recent years has roughly increased by a factor of 10 per year. Along this path, the improvement of AI not only comes from the strengthening of a single model but also may come from the expansion of collective capabilities brought about by more instances, faster inference, and larger - scale collaboration.
Path 2: Algorithms continue to evolve, and even new paradigm shifts occur
The research team pointed out that longer context, continuous learning, retrieval enhancement, tool use, robust decision - making in environmental interaction, and world models all belong to the extension of the existing paradigm; while new architectures, training objectives, or learning mechanisms are closer to a real paradigm shift. The research team did not specifically predict what the next paradigm shift would be but believed that it could still be an important source of continuous AI progress after AGI.
Path 3: Recursive self - improvement
Stronger AI can help develop the next - generation stronger AI, forming a positive feedback loop. The research team mentioned that this mechanism can be reflected in the improvement of algorithms and code, hardware design, data generation and screening, and division - of - labor efficiency. An example is AlphaZero, which first improves the output through search and then distills the results back into the model. More importantly, it is unclear how far this positive feedback can develop in reality.
Path 4: Multi - agent coordination and swarm intelligence
This path focuses not on how strong a single model can become but on a large number of AGI systems through division of labor and collaboration to form collective intelligence that exceeds the upper limit of a single entity. The research team regards automated companies, research organizations, and virtual economic systems as possible forms that may emerge along this path. According to this path, ASI may not necessarily be an extremely powerful single model but may also be a highly coordinated AI collective.
The research team also reminded that for AGI to evolve into ASI, more computing power is not necessarily better. Although the expansion of computing power is important, it will soon hit the resource ceiling, and new algorithmic ideas or even new paradigms are needed. More notably, even if a single AGI is only close to the human level, once a large number of AGIs can efficiently divide labor and cooperate, their overall capabilities may exceed those of humans.
Where are the real difficulties?
After discussing the four potential paths, the research team also summarized six types of key bottlenecks that may affect the continued strengthening of AI, as follows:
1. Data wall
The research team pointed out that high - quality human - generated data is limited, and the human text data suitable for large - scale pre - training may approach the upper limit within this decade. Whether synthetic data, simulated environment data, and data generated by AI's interaction with the real world can fill this gap quickly enough, the research team did not draw a conclusion but listed it as one of the core uncertainties.
2. Economic and natural resource pressure
If the progress of AI continues to mainly rely on scale expansion, then energy, chips, data centers, supply chains, and capital investment must all increase synchronously. The research team believes this is a real constraint but also pointed out that AI itself may also increase economic output, improve algorithm and hardware efficiency, thereby alleviating these pressures.
3. The existing neural network paradigm may not be sufficient
The research team did not rule out the possibility of the current path leading to ASI but also reminded that this path may still have fundamental limitations in continuous learning, stable reasoning, interactive decision - making, uncertainty expression, as well as hallucinations and prompt injection.
4. Research itself will become more and more difficult
The research team pointed out that as the field matures, continuing to make progress often requires higher investment; whether AI can offset this trend through automated research remains to be studied.
5. Abstraction barrier
The research team believes that if today's AI mainly learns the concept and symbol systems formed by humans, it may be good at recombining existing concepts but may not be good at autonomously extracting new primitive concepts from the real world. For example, if a modern large - scale model is trained only based on pre - Newtonian knowledge, it is almost impossible for it to independently derive the theory of general relativity or quantum mechanics from these materials.
6. Regulation, governance, and social backlash
The research team believes that regulatory thresholds, licensing systems, event reporting requirements, and social reactions caused by accidents will all affect the pace of AI's ability expansion. Behind this is not only a technical issue but also involves policy, system, market, and public risk perception.
Deficiencies and future development
Finally, the research team raised a very practical question: If AI has already surpassed humans, how should we continue to evaluate its capabilities?
Today, many benchmarks are based on human levels. Once AI approaches or exceeds the top humans in exams, programming, mathematics, Q&A, and professional knowledge tests, the original evaluation indicators may lose their meaning. Therefore, in the future, a new evaluation and prediction system for the post - AGI era needs to be established, including multi - agent competition and cooperation tasks, automatically generated tests, general compression tasks, indirect indicators such as economic productivity, and an evaluation mechanism that can be continuously updated and does not saturate prematurely.
However, in terms of content, this is not an experimental paper but more like a technical report centered around the post - AGI era. The research team pointed out that the directions worthy of attention in the future include: continuing to expand existing AGI systems, exploring new AI paradigms, achieving recursive self - improvement of systems, and forming stronger overall capabilities through large - scale multi - agent collaboration.
Finally, the research team pointed out that ASI is not an all - knowing and all - powerful “magic system”; it is still constrained by physical laws, computational complexity, data, resources, experiment time, and real - world feedback speed. It is still highly uncertain which path AI will take and at what speed. In the future, a continuously updated benchmark, prediction, and research mechanism still need to be established to reduce the uncertainty in judgment.
This article is from the WeChat official account “Academic Headlines” (ID: SciTouTiao), author: Academic Headlines. It is published by 36Kr with authorization.