DeepMind veteran Thore Graepel leaves to launch AI reasoning

Thore Graepel, a research scientist with 16 years at Google's AI lab DeepMind, has departed to found a new company. His goal is to build artificial intelligence systems capable of human-like reasoning.

Graepel cites DeepMind's AlphaGo victory over Go champion Lee Sedol in 2016 as a key inspiration. He told Sifted that watching the AI win was a moment that demonstrated the potential of the technology. Now, he believes the field must move beyond current approaches.

The case for a fundamental redesign

Graepel argues that achieving true reasoning in AI requires a foundational shift. “To get to reasoning, we need to fundamentally redesign AI,” he stated. He contends that while large language models (LLMs) are powerful, they operate more on pattern recognition than genuine reasoning. The veteran researcher suggests the industry is at an inflection point similar to the transition to transformer models that powered the current AI boom.

He draws a distinction between two types of AI systems. The first is the current generation of LLMs, which he describes as operating on “associative memory.” The second, which he aims to build, would function on “causal reasoning.” Graepel believes integrating these two approaches is the key to creating more capable and reliable AI.

Challenges with current AI models

Graepel highlighted specific shortcomings of existing models. This, he argues, is a core limitation of the associative approach.

He also pointed to issues with reliability and cost. Running advanced AI models is expensive, and their outputs can be inconsistent. Graepel's vision involves creating systems that are not only more capable but also more efficient and trustworthy for complex tasks.

Building the new venture

Graepel is now focused on assembling a team and securing funding for his new startup. While he did not disclose the company's name or specific funding details to Sifted, he indicated the venture is in its early stages. His departure follows a trend of DeepMind alumni launching new AI companies.

Other notable figures have recently left to pursue similar ambitions. For instance, DeepMind cofounder Mustafa Suleyman left to launch Inflection AI and later joined Microsoft. Another cofounder, Demis Hassabis, remains at DeepMind's helm. Graepel's move places him among a growing cohort of experienced researchers aiming to advance the frontier of AI reasoning.

Graepel concluded by emphasizing the scale of the challenge. He said building reasoning AI is a profound technical problem. “It’s a deep technical problem,” he told Sifted. The researcher is now dedicating himself full-time to solving it.