DeepMind CEO Discusses Path Toward Artificial General Intelligence

Demis Hassabis, chief executive officer of Google’s DeepMind division, spoke to Alex Kantrowitz from Big Technology during a session in Davos on Thursday, September 10, 2026. In that interview he emphasized two concepts—“scaffolding” and “orchestration”—as central to the next wave of artificial intelligence research. He noted that while modern language models can call web‑searching tools, they do not retain what those searches reveal across sessions, hinting at an architectural shortfall that may keep them from achieving general intelligence. ¹

The discussion framed scaffolding as a common approach in LLM work where the model delegates sub‑tasks to specialized tools such as search engines or calculators. Orchestration refers to coordinating these tool calls within a single inference chain. However, Hassabis pointed out that even with sophisticated orchestration, the transformer backbone itself does not naturally preserve long‑term state; each new prompt starts fresh unless extra mechanisms are added. This limitation means that a model cannot “remember” past interactions without explicit design choices like recurrent networks or knowledge graphs. The lack of persistent memory was highlighted as one reason why current systems still fall short of true AGI. ¹

Expert commentary notes that this observation echoes earlier concerns about transformer architectures: backpropagation was first introduced by Geoffrey Hinton at Siemens Research in 1986, yet it has never been paired with built-in long‑term state management beyond simple recurrence. In practice, researchers have experimented with additional modules—recurrent layers, graph databases, and external memory banks—to give models a sense of continuity. Yet none of these solutions has achieved the seamless, lifelong learning seen in biological brains.

The interview also touched on what counts as AGI. Hassabis suggested that we may need one or two more major breakthroughs before reaching that milestone, citing the rapid progress observed since AlexNet won ImageNet in 2012. He argued that achieving true artificial general intelligence will require not only architectural changes but also reasoning capabilities that go beyond pattern recognition alone—a challenge that could demand new theoretical insights or novel hardware designs.

In summary, DeepMind’s CEO framed the present generation of large language models as powerful yet incomplete: they can orchestrate tools and scaffold tasks, but without long‑term memory and deeper reasoning they cannot be called truly intelligent. His remarks signal that the community must focus on integrating persistent state mechanisms—whether through recurrent networks, knowledge graphs, or external memory modules—and on developing efficient compute scaling strategies if it is to move from advanced LLMs toward genuine AGI.