The Arrival of AGI: DeepMind's Co-Founder on When It Happens

01 Defining AGI and Why the Definition Matters

Artificial general intelligence is one of the most discussed and least precisely defined concepts in technology. Shane Legg, who coined the term in a 2004 paper with Ben Goertzel, defined it as a system capable of performing any intellectual task that a human can perform. This definition, while elegant, raises immediate questions: which humans, at what skill level, and across what range of tasks? A more practical definition, often used in the AI industry, describes AGI as a system that can be given any human job and perform it at or above human level without requiring task-specific training.

The definition matters because it determines when AGI has been achieved. A loose definition might be met by systems that exist today, while a strict definition might require capabilities that are decades away. Companies have incentives to define AGI in ways that align with their progress, making independent assessment difficult. The conversation between Shane Legg and Hannah Fry, featured in the video accompanying this article, highlights how the definition has evolved and why it remains contested.

Legg has consistently argued that AGI should be measured by performance across a broad range of cognitive tasks, not by a single benchmark or capability. This positions AGI as a threshold of generality rather than a specific score on a specific test. The practical implication is that AGI arrival is likely to be gradual rather than sudden, with systems progressively mastering more domains until the breadth of capability crosses a widely accepted threshold.

02 Shane Legg's Timeline Prediction and Methodology

Shane Legg has been making AGI timeline predictions for over a decade, and his estimates have been notably more consistent than those of many of his peers. In the conversation with Hannah Fry, Legg discusses his current estimate that AGI could arrive within the next several years, potentially by 2028 or shortly thereafter. This estimate is based on a combination of scaling laws, observed capability improvements, and his assessment of remaining technical challenges.

Legg's methodology is not based on a single metric but on tracking progress across multiple dimensions simultaneously. He monitors benchmark performance, the rate of improvement on those benchmarks, the emergence of new capabilities that were not present in earlier models, and the gap between current capabilities and those required for general intelligence. The consistency of improvement across these dimensions gives him confidence that the trajectory is real rather than an artifact of benchmark overfitting.

It is worth noting that Legg's predictions have been criticized by some researchers as overly optimistic. Yann LeCun, for example, has argued that current approaches to AI, particularly autoregressive language models, have fundamental limitations that will prevent them from reaching general intelligence without significant architectural innovations. The debate between optimists and skeptics is not merely academic: it influences investment decisions, regulatory priorities, and public expectations.

03 Current Capabilities vs AGI Benchmarks

The gap between current AI capabilities and what would constitute AGI is narrowing in some dimensions and remains wide in others. Language models have achieved remarkable performance on knowledge-intensive tasks like MMLU, a benchmark of academic and professional knowledge across 57 subjects. The best models now score above 85 percent on MMLU, approaching or exceeding human expert performance. Mathematical reasoning, measured by benchmarks like GSM8K, has similarly improved dramatically, with frontier models now scoring above 90 percent.

However, benchmarks capture only a slice of what general intelligence requires. Current AI systems struggle with tasks that humans find trivial: adapting to entirely new situations without training data, maintaining coherent long-term plans, understanding physical causality, and learning efficiently from small numbers of examples. A child who sees a single cat can recognize cats for life; a language model requires thousands of examples to achieve similar reliability. This gap in sample efficiency is one of the clearest indicators that current architectures, however impressive, may not be sufficient for general intelligence.

The frontier of capability is also moving in ways that benchmarks do not capture well. Recent models have demonstrated abilities that were not anticipated by their developers, including the ability to reason about novel situations, write working code from natural language descriptions, and engage in multi-turn planning. These emergent capabilities suggest that the relationship between scale and capability is not fully understood, and that progress may continue to surprise in both directions.

04 Scaling Laws and the Diminishing Returns Debate

The scaling hypothesis, which holds that increasing model size, training data, and compute will continue to yield proportional improvements in capability, has been the dominant paradigm in AI development for the past several years. The empirical scaling laws derived by researchers at OpenAI, DeepMind, and Anthropic have guided investment decisions worth billions of dollars. The question of whether these laws will continue to hold, or whether returns will diminish as models scale, is one of the most consequential uncertainties in the field.

Evidence is mixed. Some recent model releases have shown improvements that appear consistent with continued scaling, while others have suggested that the rate of improvement on certain benchmarks may be slowing. The interpretation of this evidence is complicated by the fact that benchmarks become saturated as models approach ceiling performance, making it difficult to distinguish between diminishing returns and benchmark limitations. New, more challenging benchmarks are continually being developed to address this issue.

Legg acknowledges the uncertainty but argues that even if scaling laws eventually plateau, the current trajectory has enough momentum to carry AI systems to near-AGI capabilities. The combination of architectural improvements, better training methods, and more efficient inference may compensate for any slowdown in raw scaling. The critical question, in his view, is not whether scaling alone will produce AGI, but whether the combination of scaling and algorithmic improvement will do so within the expected timeframe.

05 The Role of Reinforcement Learning and World Models

While language models have dominated recent AI progress, many researchers believe that general intelligence will require capabilities that go beyond language. Reinforcement learning, which trains systems to make decisions through trial and error in simulated or real environments, is seen as a key ingredient for developing world models: internal representations of how the physical and social world works, which can be used for planning and prediction.

DeepMind has been a leader in reinforcement learning since its founding, with landmark achievements including AlphaGo, AlphaZero, and MuZero. These systems demonstrated that reinforcement learning could produce superhuman performance in narrow domains, but the question of whether these techniques can be combined with language models to produce general intelligence remains open. Legg suggests that the integration of language understanding with world modeling and planning is one of the key remaining challenges.

Recent work on reinforcement learning from human feedback, used to align language models with human preferences, has shown that RL techniques can be effectively combined with large-scale pretraining. The extension of this approach to more complex environments, where systems must plan over longer time horizons and reason about physical consequences, is an active area of research. Whether this line of work will produce the missing capabilities for AGI, or whether entirely new approaches will be needed, is one of the central open questions in the field.

06 Safety and Alignment Challenges

The prospect of AGI raises safety concerns that are qualitatively different from those associated with current AI systems. A system that can perform any intellectual task that a human can perform could also perform tasks that humans would prefer it not to perform, including tasks that are harmful. The alignment problem, ensuring that AI systems pursue goals that are consistent with human values, becomes more urgent as capabilities approach general intelligence.

Legg has been a consistent voice for taking safety seriously, arguing that the technical challenges of alignment are comparable in difficulty to the challenges of building AGI itself. Current approaches to alignment, including reinforcement learning from human feedback and constitutional AI, have shown promise but also have known limitations. They depend on human raters whose preferences may be inconsistent, biased, or incomplete, and they may not generalize to the more capable systems of the future.

The safety challenge is compounded by the competitive dynamics of AI development. Companies and countries racing to build AGI have incentives to prioritize speed over safety, particularly if they believe that being first confers a decisive advantage. Legg acknowledges this tension but argues that the major AI companies have shown genuine commitment to safety, investing in research that does not directly contribute to capability improvement. Whether this commitment will hold as the finish line approaches is one of the most important unanswered questions.

07 Societal Preparation and Economic Implications

If AGI arrives within the timeframe that Legg and others predict, the societal implications are profound. A system capable of performing any intellectual task that a human can perform could automate large portions of cognitive work, from programming and legal analysis to scientific research and creative writing. The economic effects could be comparable to or larger than those of previous general-purpose technologies like electricity and computing.

The pace of adoption will depend on many factors beyond raw capability. Regulatory frameworks, economic incentives, social acceptance, and the development of complementary infrastructure will all influence how quickly AGI capabilities are deployed. Some economists argue that the bottleneck will not be AI capability but the speed at which organizations can restructure to take advantage of it, a process that typically takes years or decades rather than months.

Legg suggests that the most important form of societal preparation is not specific policy interventions but building the institutional capacity to respond adaptively as capabilities emerge. This means investing in AI safety research, developing regulatory frameworks that can evolve with the technology, and fostering public understanding of AI capabilities and limitations. The goal is not to predict exactly what will happen but to build the resilience needed to respond well to whatever does happen.