Key Takeaways
- A research scientist at Google DeepMind identifies model scaling, paradigm shifts, agent coordination, and recursive self-improvement as potential pathways to artificial general superintelligence.
- Falling inference costs and expanding compute capacity could industrialize cognitive work even without a decisive breakthrough in frontier models.
- Businesses and governments may need broader evaluation systems because compute restrictions alone would address only one route to more capable AI.
A Google DeepMind researcher has mapped multiple technological pathways that could take artificial intelligence beyond human-level performance, arguing that the transition from artificial general intelligence to artificial general superintelligence is becoming an empirical research and governance issue rather than a purely philosophical one.
The researcher presented the framework in a November 19, 2026, essay published by the DeepMind Institute. The platform was started by researchers from Google and Google DeepMind, but notes that its essays represent their authors' ideas rather than Google's official position.
The researcher defines artificial general intelligence, or AGI, as systems that match humans across nearly all cognitive tasks. Artificial general superintelligence, or ASI, would sit much further along the spectrum: it would consistently outperform large groups of experts across virtually all tasks of human interest while using less time and fewer resources.
That distinction matters for enterprise planning. A single model that equals an individual employee represents one economic proposition. Billions of replicable agents operating faster than people, sharing information digitally, and moving between computing environments would present entirely different operational dynamics.
Continued scaling of models, training data, and test-time compute represents a primary pathway. Larger systems have repeatedly gained new capabilities, but the approach faces potential constraints. High-quality human-generated data is finite, larger models can be slower, and further expansion depends heavily on power generation, electrical grids, chip fabrication, and high-speed interconnects.
Synthetic data and simulated environments could ease the data shortage, although quality control remains a concern. Research published in Nature has shown that indiscriminate training on recursively generated data can degrade model performance. Sourcing larger volumes of data does not automatically yield better outputs.
A deeper technological shift offers an alternative trajectory. Current frontier systems largely rely on transformers trained to predict symbols, followed by instruction tuning and additional computation during inference. Yet they still exhibit limitations in continual learning, world modeling, decision-making, and sample efficiency. Researchers remain divided over whether incremental extensions can close those gaps or whether ASI would require a fundamentally different architecture.
Advancing agent coordination provides another distinct direction. Thousands or millions of AI agents could be organized into research groups, businesses, or virtual economies, much as human institutions combine specialized labor. Early work suggests multi-agent systems can improve performance on some tasks, but communication costs, duplicated work, misinformation, collusion, and diminishing returns complicate the picture.
However, adding agents does not automatically equate to adding useful intelligence. Enterprises already know that larger human teams can produce more meetings rather than proportionately more output. AI groups may reproduce that organizational problem at machine speed, making governance design a necessary layer of the technical stack.
Recursive self-improvement introduces an accelerating dynamic. AI could contribute to better model architectures, training methods, chip designs, and datasets, with each improvement potentially accelerating the next cycle. The underlying From AGI to ASI report divides this process into changes to code and design, improvements through reusable data, and group-level specialization.
Recursive improvement may advance fastest where outputs can be checked automatically, such as mathematical proofs or software tests. It could progress more slowly in strategy, policy, or scientific interpretation, where evaluation is contested and human experts may struggle to assess superhuman work.
Even without ASI, the researcher argues that "cognitive industrialization" could reshape business. The essay estimates that hardware improvement, algorithmic efficiency, and investment have together increased global effective compute stock by at least 10 times per year. It also cites falling inference costs that could reduce the expense of performing a fixed cognitive workload by about four times annually.
Those projections are extrapolations, not promises. Physical infrastructure, energy supply, capital availability, regulation, and diminishing research returns could interrupt them. Still, the business implication is clear: organizations may face rapidly declining prices for some forms of analysis, software development, content production, and technical support.
For policymakers, this multi-pathway model challenges strategies centered exclusively on limiting training compute. Progress could shift toward efficiency, coordination, new architectures, or AI-assisted research. Measuring frontier capability may therefore require evaluation systems covering cost, speed, autonomy, collective performance, and the ability to improve subsequent models. The central question is no longer simply when AGI arrives. It is how institutions will recognize and govern the potentially faster transition that follows.
⬇️