UKtech50 2025 winner: Demis Hassabis, co-founder and CEO, DeepMind

The development of AlphaZero

It was while at UCL that Hassabis met Suleyman and Legg, and the three came up with the idea that neuroscience, AI and machine learning together could create powerful algorithms, thus creating DeepMind.

One of the company’s first early achievements, before it was bought by Google, was using deep reinforcement learning to create an unbeatable AI model that could win Atari games such as Space Invaders and Pong. DeepMind then moved onto more complex games, gaining the attention of tech giant Google.  

Following the Google acquisition, Hassabis stayed on as DeepMind’s CEO, and in 2016, the company launched its AI system AlphaGo, which famously defeated world champion Lee Sedol in the complex, ancient board game of Go.

This showed, for the first time, the true potential of general AI to discover solutions humans may not have even considered, not just in the field of video games, but in areas including medicine and science.

AlphaGo itself, as well as the concept of artificial general intelligence (AGI), led to the development of AlphaZero, a game-playing system which taught itself to master chess as well as other games, without any human guidance. However, the journey of DeepMind’s AI systems has since moved on.

Scoring a Nobel Prize

As well as advancing research on AI safety and the development of a partnership with London’s Moorfields Eye Hospital for the use of artificial intelligence to identify and treat degenerative eye conditions, DeepMind developed AlphaFold in 2018, which is a system to solve the protein folding problem. 

The system accurately predicted the 3D shape a protein will fold up to when it’s in the body, a concept which was first articulated by another Nobel Prize winner, Christian Anfinsen, in 1972. DeepMind used 150,000 proteins whose structure had already been identified to train AlphaFold to predict their shape. This then led to the development of Alphafold2, for which Hassabis won a Nobel Prize.

It was 2020 when Hassabis, together with DeepMind colleague, American chemist and computer scientist John Jumper, created the second version of the AI model for structure prediction. The model can accurately predict the 3D model of protein structures by taking the protein’s amino acid sequence, not only on single protein chains. There are currently more than 200 million predicted protein structures available in the AlphaFold database, a far cry from the 150,000 that were known before Hassabis created the AI model.

The AlphaFold database is currently being used by several sectors all over the world, including pharmaceutical research, healthcare and environmental technology. In Singapore, researchers are using it to come up with ways to diagnose and treat Parkinson’s disease quicker, while the US is using it to combat antibiotic resistance. In Norway, researchers have used it to figure out how to increase honeybees’ chance of survival through looking into vitellogenin, a protein fundamental to bees’ immune system. Meanwhile, researchers at the University of Portsmouth in the UK have screened 100 candidate enzymes to engineer faster and cheaper plastic-recycling enzymes.

The future of AI

So, what’s next for Hassabis and AI? Despite progress, he believes there is still a while to go before many of the capabilities of AI will have evolved to live up to its current hype.

In March 2025, speaking at an event to mark the availability of audio generation model Chirp 3 on the Google Vertex AI platform, Hassabis said that AI will have evolved to AGI, whereby the AI system exhibits “the cognitive capabilities” of humans, within that time. “That’ll be a moment when we have finally arrived with a kind of general intelligence, which is the original aim of the whole field of AI,” he added.

While AI models have evolved drastically in the past few years, there are still challenges when it comes to combining them with planning algorithms. “If your AI model has a 1% error rate and you plan over 5,000 steps, that 1% compounds like compound interest,” Hassabis said at the event.

By the time 5,000 steps have been worked through, the compounded error, according to Hassabis, means the possibility of the answer being correct is random. “For a games model, you have the rules of Chess or Go,” he said, which aids the planning algorithm in making the correct decision. “In the real world, you don’t have perfect information. There’s hidden information that we don’t know about, so we need AI models that are able to understand the world around us.”

For Hassabis, one of the interesting developments expected to appear over the next few years is the deployment of multiple AI agents that work together to solve a problem, and AI agents themselves can be used in a general AI system to solve problems.

Only a few weeks ago, DeepMind launched its latest venture, the AlphaGenome tool, which accurately predicts how single variants or mutations in DNA sequences affect the biological processes regulating genes.

However, this too has limitations when it comes to accurately capturing the “distant regulatory elements, like those over 100,000 DNA letters away, [which is] still an ongoing challenge”, according to DeepMind. “Another priority for future work is further increasing the model’s ability to capture cell- and tissue-specific patterns.”