DeepMind CEO Challenges AI Industry's Scaling Strategy, Advocates World Models as Path to AGI

The Core Argument Against Pure Language Models

Hassabis, a Nobel Prize-winning AI researcher, made his case during a podcast appearance, explaining that while today's large language models excel at pattern recognition and statistical prediction, they don't truly grasp why events unfold as they do. These systems predict the next token based on correlations in training data rather than understanding the underlying causal mechanisms of the world.

This represents a direct challenge to the strategy that has dominated the AI industry for the past several years. Companies including OpenAI have invested billions in scaling language models larger and faster, driving rapid commercial adoption. However, Hassabis contends this approach has encountered what he describes as a fundamental wall when it comes to genuine scientific innovation and reasoning from first principles.

The World Model Alternative

Real scientific discovery, according to Hassabis, requires an internal simulation engine capable of running thought experiments, accurately simulating physics, and understanding how reality operates at a deeper level than word sequences. This is where world models come in.

DeepMind is actively building towards this vision with two key systems. Genie 3, released in August 2025, can generate interactive 3D environments from text prompts, creating dynamic worlds that users can navigate in real time at 24 frames per second with 720p resolution. The system has developed what researchers describe as a data-driven understanding of physical causality, knowing how snow should accumulate on surfaces or how objects behave under different gravitational conditions.

SIMA 2, the companion system, is an AI agent that can navigate and learn within these simulated worlds. Powered by Google's Gemini models, SIMA 2 has evolved from simply following instructions to becoming an interactive companion that can reason about goals, converse with users, and improve itself over time.

Early research indicates these hybrid approaches that combine language understanding with physical simulation outperform pure language models by 20 to 30 percent on complex reasoning tasks. They also show significant reductions in hallucinations about basic physics.

AGI Timeline and Competition

Despite his critique of current methods, Hassabis maintained his prediction that artificial general intelligence has a 50 percent chance of arriving by 2030. However, he emphasized that his definition sets a high bar, requiring capabilities like scientific creativity and continuous learning. He estimated that AGI remains 5 to 10 years away and will require what he termed two AlphaGo-scale breakthroughs, referring to major innovations comparable to DeepMind's historic achievement in mastering the ancient game of Go.

Speaking alongside Anthropic CEO Dario Amodei at a World Economic Forum session, Hassabis described the remaining missing ingredients for AGI, including better reasoning, planning, and robustness.

On the international competition front, Hassabis offered a sobering assessment of China's AI capabilities. He noted that Chinese companies like ByteDance are perhaps only 6 months behind the Western frontier, not the 1 to 2 years that many had assumed. However, he questioned whether Chinese labs can innovate beyond the current technological frontier rather than simply fast-following Western breakthroughs.

Industry Context and Implications

The debate arrives at a volatile moment for the AI industry. Following Google's Gemini 3.0 launch in late 2025, reports emerged of an internal code red at OpenAI amid concerns about diminishing returns from pure scaling strategies. Hassabis revealed that he is now in daily contact with Alphabet CEO Sundar Pichai, underscoring DeepMind's elevated role as the engine room of Google's AI initiatives.

The shift away from pure scaling towards more diverse approaches has broader implications. OpenAI's finance chief recently announced that 2026 would be the year of practical adoption, focusing on closing the gap between AI capabilities and real-world utilization. Industry analysts note that the pure scaling approach that dominated 2025 may have reached its limits, with data constraints potentially exhausting public text data by 2028 and compute capacity fully booked through 2026.

Hassabis described the current competitive environment as ferociously intense, potentially the most competitive period in the history of technology, with incredibly high stakes. The path forward, according to DeepMind's vision, lies not in simply making language models larger but in building AI systems that can truly understand and simulate the physical world, bridging the gap between pattern recognition and genuine causal reasoning.