DeepMind Report: AGI Is Not the End; The Path to ASI Begins

If artificial general intelligence (AGI) were achieved tomorrow, what would the next stage of AI look like?

The Google DeepMind team and their collaborators propose in their latest research report that AGI is unlikely to be the endpoint. In their view, AI will not stop at a level approaching human capability but will continue to grow stronger, surpassing even the most elite human expert teams, ultimately leading to artificial superintelligence (ASI).

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

In this report, the research team outlines four potential pathways for AI's evolution from AGI to ASI, key bottlenecks that may arise, and the most promising research questions to pursue.

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

The research team stated that, due to significant uncertainty in predicting ASI progress, it cannot be ruled out that AI will continue to accelerate over the coming years. This may suggest that the scenario of a single transformative leap brought about by introducing human-level AGI into society might not be accurate.

A more realistic prospect is that AI-driven advancements and breakthroughs will emerge across numerous fields of science and technology, triggering a series of transformative societal changes.

To address this prospect, a large-scale, interdisciplinary effort with a global perspective and broad concern is required.

Before discussing how AI will continue to grow stronger, the research team first distinguished among three commonly confused concepts: AGI, ASI, and UAI.

AGI (Artificial General Intelligence): A general intelligent system that reaches the median human level across most cognitive tasks, corresponding to the average cognitive abilities of ordinary people rather than the expertise of top specialists. The research team also noted that the first generation of AGI may already surpass humans in some tasks, but has not yet achieved sufficient broad generality.

ASI (Artificial Super Intelligence): It surpasses humans not just in a few specific tasks, but overall in virtually all domains that matter to humans; its benchmark is not a single expert, but a large, well-coordinated collective of human experts.

UAI (Universal Artificial Intelligence): The theoretical upper bound of machine intelligence, formally described by the AIXI framework. AIXI represents a theoretical optimal universal agent. Real-world AI can only gradually approach this upper bound but cannot directly achieve it.

At the same time, the research team noted that the transition from AGI to ASI may involve more than one pathway, and they proposed four potential parallel pathways:

Path One: Continue Expanding Computing, Models, and Data

This path follows the fundamental logic of AI progress over the past decade, including more powerful hardware, larger training runs, higher algorithmic efficiency, larger models, and more data. The research team notes that “effective compute” in recent years has grown by approximately a factor of 10 per year. Along this trajectory, AI advancements arise not only from individual models becoming stronger, but also from collective capability expansion through more instances, faster inference, and larger-scale collaboration.

Path Two: The algorithm continues to evolve, even giving rise to new paradigm shifts.

The research team notes that longer contexts, continual learning, retrieval augmentation, tool use, robust decision-making in environment interaction, and world models all represent extensions of existing paradigms; whereas new architectures, training objectives, or learning mechanisms are closer to genuine paradigm shifts. The team does not specifically predict what the next paradigm shift will be, but believes it could still be a significant source of continued AI progress beyond AGI.

Path Three: Recursive Self-Improvement

Stronger AI can help develop the next generation of even stronger AI, creating a positive feedback loop. The research team notes that this mechanism can manifest in improvements to algorithms and code, hardware design, data generation and filtering, and efficiency in task allocation. A relevant example is the approach used by AlphaZero, which first refines outputs through search and then distills the results back into the model. More importantly, the extent to which this positive feedback loop can realistically evolve remains to be seen.

Path Four: Multi-Agent Coordination and Swarm Intelligence

This path focuses not on how strong a single model can become, but on how a large number of AGI systems, through division of labor and collaboration, form a collective intelligence that exceeds the limits of any individual system. The research team views automated companies, research organizations, and virtual economic systems as potential manifestations of this path. Under this approach, ASI may not be a single, extremely powerful model, but rather a highly coordinated collective of AIs.

The research team also cautions that the transition from AGI to ASI is not necessarily about having more computing power alone. While expanding computing power is important, it will soon hit resource limits and will require new algorithmic approaches, or even entirely new paradigms. More notably, even if each individual AGI is only close to human-level capability, a large number of AGIs working together efficiently with specialized roles could collectively surpass human abilities.

After discussing four potential pathways, the research team also identified six key bottlenecks that could hinder AI’s continued advancement, as follows:

1. Order Book

The research team noted that high-quality, human-generated data suitable for large-scale pretraining may approach its upper limit within this decade. The team did not conclude whether synthetic data, simulation environment data, and data generated from AI interactions with the real world could fill this gap quickly enough, instead listing it as one of the key uncertainties.

2. Economic and natural resource pressures

If AI progress continues to rely primarily on scaling up, then energy, chips, data centers, supply chains, and capital investment must all grow in tandem. The research team views this as a practical constraint but also notes that AI itself could increase economic output and improve the efficiency of algorithms and hardware, thereby alleviating these pressures.

3. Existing neural network paradigms may not be sufficient

The research team has not ruled out the possibility that the current path leads to ASI, but also cautions that this approach may still have fundamental limitations in areas such as continuous learning, stable reasoning, interactive decision-making, expressing uncertainty, and issues like hallucinations and prompt injection.

4. The research itself will become increasingly difficult.

The research team noted that as the field matures, further progress often requires greater investment; whether AI can offset this trend by automating research remains to be determined by future studies.

5. Abstract barriers

The research team believes that if today’s AI primarily learns from human-established concepts and symbolic systems, it may excel at recombining existing concepts but may not be adept at autonomously deriving new conceptual primitives from raw reality. For example, if a modern large model is trained solely on pre-Newtonian knowledge, it would be nearly impossible for it to independently derive general relativity or quantum mechanics from those materials alone.

6. Regulation, Governance, and Social Backlash

The research team believes that regulatory thresholds, licensing systems, incident reporting requirements, and societal reactions to accidents will all influence the pace at which AI capabilities scale. This involves not only technical factors but also policies, institutions, markets, and public perceptions of risk.

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 calibrated against human performance; once AI approaches or surpasses top human abilities in exams, programming, mathematics, question answering, and domain-specific tests, existing evaluation metrics may lose their relevance. Therefore, future evaluation and prediction systems must be designed for the post-AGI era, incorporating indirect metrics such as multi-agent competition and collaboration tasks, automatically generated tests, general compression tasks, and economic productivity, along with assessment mechanisms that can continuously evolve without saturating prematurely.

However, in terms of content, this is not an experimental paper but rather a technical report focused on the post-AGI era. The research team highlights key areas for future attention: further expanding existing AGI systems, exploring new AI paradigms, achieving recursive self-improvement of systems, and enhancing overall capabilities through large-scale multi-agent collaboration.

Finally, the research team emphasized that ASI is not an omniscient or omnipotent "magic system"—it remains constrained by physical laws, computational complexity, data availability, resources, experimental time, and the speed of real-world feedback. The path AI will take and the pace of its advancement are still highly uncertain. In the future, ongoing benchmarks, predictions, and research mechanisms will be needed to reduce uncertainty in assessments.

This article is from the WeChat public account "Academic Headline" (ID: SciTouTiao), authored by Academic Headline.