Researchers from Google DeepMind and partner universities have presented the report "From AGI to ASI" (arXiv:2606.12683), which analyzes possible scenarios for the transition from Artificial General Intelligence (AGI) to Artificial Superintelligence (ASI).
What Happened
The work identifies four key vectors of technological progress: continued scaling of compute and data, shifts in algorithmic paradigms, recursive self-improvement of AI systems, and the development of multi-agent structures. The authors also point to barriers such as the quality data deficit ("data wall"), rising research costs, and physical limitations of logical computation.
Context
Current growth in compute power demonstrates rates of approximately 10x per year according to Epoch AI data. However, the industry faces the necessity of transitioning from simple parameter scaling to searching for new architectures capable of working efficiently under conditions of limited training data.
Why It Matters for the Industry
The report forms a theoretical foundation for long-term R&D planning. The industry needs to shift focus from a Transformer-centric approach toward developing data synthesis methods, agentic frameworks, and algorithms that support self-improving loops.
Why It Matters for Users
For readers and developers, this analysis moves the discussion of superintelligence from the realm of science fiction into the realm of technological forecasting. This allows for a better understanding of the specific technological and regulatory barriers that will shape the face of AI in the coming decades.
What Remains Unknown / Limitations
There is a difference in the assessment of the significance of development vectors: technical specialists emphasize architectural shifts, while the business community focuses on the economics of development and the efficiency of multi-agent systems.
Sources
Author
Look at AI, Editorial Staff