The arrival of AGI | Shane Legg (co-founder of DeepMind)
Shane Legg, Chief AGI Scientist at Google DeepMind, outlines his framework for AGI, predicting 'minimal AGI' within years and 'full AGI' within a decade. He details a path to more reliable systems and introduces 'System 2 Safety' for building ethical AI. Legg issues an urgent call for society to prepare for the massive economic and structural transformations that advanced AI will inevitably bring.
Defining Artificial General Intelligence (AGI)
Current AI systems exhibit “a lot more than sparks” of AGI. The working definition for minimal AGI is an artificial agent capable of performing the cognitive tasks that people can typically do. This benchmark is useful because it avoids setting the standard at the level of extraordinary genius, while still requiring a broad range of general capabilities.
The term “AGI” was originally proposed to define a field of study focused on general, thinking machines, as opposed to the narrow, specialized AI that was prevalent at the time. It has since evolved to describe a category of artifact, which has led to widespread confusion as different groups use the term to mean different things. Many popular definitions, such as the ability to make $1 million from a starting fund or pass a specific exam, are too narrow and miss the crucial element of generality that is core to the concept.
The Levels of AGI: A Spectrum of Capability
AGI should not be viewed as a single, discrete threshold but as a spectrum of increasing capability and generality.
- Level 1: Minimal AGI: This is the point at which an AI is no longer surprisingly deficient in cognitive tasks that typical humans can perform. It can reliably do what you would expect from an average person. Shane Legg maintains his long-standing prediction of a 50% probability of reaching this level by 2028.
- Level 2: Full AGI: At this level, the AI can achieve the full spectrum of human cognitive abilities, including the extraordinary feats of geniuses like Einstein or Mozart. This is predicted to be achievable within the next decade.
- Level 3: Artificial Superintelligence (ASI): This refers to intelligence that is far beyond what is possible for any human. The transition to ASI is considered inevitable due to the fundamental physical and computational advantages of silicon over biological brains. A data center has orders-of-magnitude advantages over the human brain in energy consumption (megawatts vs. ~20W), signal speed (speed of light vs. ~30 m/s), and bandwidth. Human intelligence is not the upper limit of what is possible.
Operationalizing the Definition: A Two-Phase Test
To scientifically determine if a system has reached minimal AGI, a two-phase evaluation process is proposed:
- Standardized Testing: The AI must first pass a comprehensive suite of tests that cover a wide range of cognitive tasks for which typical human performance is a known baseline. It must not fail any of them.
- Adversarial Testing: Following success in the first phase, an expert adversarial team would be given an extended period (e.g., months) to probe the AI with full access. Their goal is to find any cognitive task a typical person can do that the AI cannot. If they fail to find such a weakness, the system can be considered a minimal AGI for all practical purposes.
Building More Capable and Reliable AI
Progress towards AGI is not merely a matter of scaling up models or data. It requires fundamental algorithmic and architectural innovations to address current weaknesses, such as:
- Continual Learning: Current models are largely static after training. Future systems will need processes for ongoing learning, akin to an episodic memory, allowing them to learn and adapt over time much like a human starting a new job.
- Reasoning: Key weaknesses remain in areas like visual reasoning (e.g., understanding 3D perspective in a 2D image) and abstract reasoning. Addressing these requires collecting specific types of data to train these specific abilities.
AI Ethics and “System 2 Safety”
A crucial aspect of developing AGI is embedding a robust understanding of ethics. The proposed approach is “System 2 Safety,” a concept inspired by Daniel Kahneman’s model of human cognition (System 1 “fast” thinking vs. System 2 “slow” thinking).
This framework moves beyond simple, instinctual responses. The AI would be designed to engage in a deliberate, logical “chain of thought” to analyze complex ethical situations. It would consider nuances, the likely consequences of different actions, and how those actions align with established ethical principles. The ultimate goal is for the AI to become more ethical than humans by consistently applying this rigorous reasoning at a potentially superhuman level, thereby scaling its ethical judgment with its capability.
The Imminent Societal and Economic Transformation
The arrival of capable AGI will trigger a “massive transformation,” fundamentally and structurally changing the global economy and society. The current economic model, where most people contribute labor (mental or physical) in exchange for access to resources, may no longer function effectively.
While this transformation will enable the production of immense wealth and a dramatic increase in goods and services, it creates an urgent need to rethink societal structures. New systems for wealth distribution and a new vision for a “post-AGI society” must be developed to ensure broad human flourishing.
The Future of Work
In the coming years, AI will shift from being a helpful tool to performing a significant fraction of all cognitive labor.
- High-Risk Jobs: Roles that are purely cognitive and can be performed remotely with a standard computer interface are most susceptible to automation. This includes elite cognitive work like software engineering, finance, and law. Productivity gains will be immense; an example given is a team of 100 software engineers being replaced by 20 engineers who are augmented by advanced AI.
- Lower-Risk Jobs: Roles that require physical embodiment (e.g., plumbers, electricians) or have a core human-to-human interaction component (e.g., therapists, influencers) will be more protected in the short to medium term.
A Call to Action: The “March 2020” Moment
The current situation is compared to the early days of the COVID-19 pandemic, where experts could see an exponential curve coming while the general public and many institutions were slow to react. AGI must be taken seriously now by experts in every field—education, law, economics, medicine, city planning, and more. Every part of society needs to begin thinking deeply about the profound implications of cheap, abundant, and capable machine intelligence to successfully navigate the risks and harness the unprecedented opportunities for progress.