The arrival of AGI | Shane Legg (co-founder of DeepMind)
Shane Legg, Chief AGI Scientist at Google DeepMind, outlines his framework for AGI levels, predicts a 50% chance of minimal AGI by 2028, and discusses the profound societal and economic transformations that will follow.
Current AI systems exhibit “a lot more than sparks” of Artificial General Intelligence (AGI). While not yet consistently reliable, they demonstrate an uneven capability profile. In some areas, like translating languages or possessing vast general knowledge, they are already superhuman. However, they still fall short on tasks that are typical for humans, such as continual learning over extended periods and certain forms of reasoning, especially visual-spatial reasoning. These are not seen as fundamental blockers, and progress is being made through a combination of more targeted data, algorithmic advancements, and new architectures—not just by scaling up existing models.
A Framework for AGI: From Minimal to Superintelligence
AGI should be understood as a spectrum rather than a single, discrete threshold. A useful framework for categorizing this spectrum involves several levels:
- Minimal AGI: This is the point where an artificial agent can perform the cognitive tasks that people can typically do. It serves as a minimum bar because the system no longer fails in ways that would be surprising for a human. This is the level predicted to have a 50% chance of being reached by 2028.
- Full AGI: This level is achieved when an AI can match the full spectrum of human cognitive abilities, including the extraordinary feats of geniuses like Einstein or Mozart. This marks the point where we have truly “nailed” human-level intelligence. This is anticipated to arrive within a decade.
- Artificial Superintelligence (ASI): This refers to systems that go far beyond the cognitive capabilities of any human. The physical limitations of the human brain—power consumption (~20W), signal propagation speed (~30 m/s), and processing frequency (~100-200Hz)—are vastly inferior to what is possible with silicon. Data centers can have six to eight orders of magnitude more capacity across energy, space, bandwidth, and speed, making the idea of human intelligence as the upper limit “absolutely not” the case.
To operationalize the definition of Minimal AGI, a two-phase evaluation process is proposed. First, an AI would have to pass a comprehensive suite of tests covering a wide range of cognitive tasks where typical human performance is known. If it passes without failing any, it would enter a second, adversarial phase where a dedicated team has a set amount of time (e.g., a few months) to find any cognitive task a typical person can do that the AI cannot. If they fail to find such a weakness, the system would be considered a Minimal AGI for all practical purposes.
AI Ethics and System 2 Safety
A crucial aspect of developing capable AGI is ensuring it behaves ethically. The goal is to build systems that can reason robustly about ethics, potentially becoming more ethical than humans by consistently applying complex moral reasoning without emotional bias.
A promising approach is “System 2 Safety,” inspired by Daniel Kahneman’s model of human cognition:
- System 1 is fast, intuitive, gut-reaction thinking.
- System 2 is slow, deliberate, analytical reasoning.
Instead of relying on simple, pre-programmed rules (which can fail in nuanced situations), the AI would use a “chain of thought” process to reason through the complexities of an ethical dilemma. It would analyze the situation, possible actions, and likely consequences with respect to a given system of ethics. This process would be transparent, allowing for monitoring and interpretability to ensure the AI’s reasoning aligns with human values. This is critical because if superintelligence is inevitable, we must ensure it is also “super ethical.”
Societal Transformation: A New World Structure
The arrival of AGI will trigger a massive transformation, fundamentally restructuring the economy and society on a scale comparable to the Industrial Revolution. The current societal model, where most people contribute labor in exchange for access to resources, may no longer be viable.
The key implications include:
- Future of Work: Jobs that are purely cognitive and can be performed remotely with a standard computer interface are most vulnerable. Fields like software engineering, finance, and law will be significantly impacted. For instance, a team of 100 software engineers might soon be replaced by 20 engineers equipped with advanced AI tools. Work that is not purely cognitive (e.g., a plumber) or relies heavily on a specific human connection (e.g., an influencer) will be more protected in the short-to-medium term.
- Economic Disruption: While AGI should vastly increase the total “pie” of wealth and resources by automating cognitive labor and accelerating scientific discovery, the distribution of this wealth is a critical challenge. A post-AGI society requires new systems to ensure widespread human flourishing.
- A Call to Action: There is an urgent need for experts in every field—law, economics, education, medicine, politics—to take AGI seriously and begin thinking about its implications for their domains. Currently, many experts are not internalizing the exponential pace of progress, much like the public’s slow reaction to the pandemic warnings in early 2020. This is a pivotal moment where proactive, careful navigation is required to realize the immense benefits of AGI while mitigating its risks.