Discover new drugs in months, not years, with AI: DeepMind CEO

The use of artificial intelligence (AI) can unlock research with significant impact on industry and society, such as in the pharmaceutical sector and in public health. 

But DeepMind CEO Demis Hassabis, speaking at Mobile World Congress 2024 last week, also warned that questions around the principles and ethics around AI need to be discussed. 

“We’ve worried about [bad actors utilising AI] from the beginning of DeepMind,” he said. “We had an ethics charter since our founding and that eventually became the Google AI Principles.

“It’s a social-technical issue that… should involve [governments, Big Tech and civil society] to discuss what values we want these systems to have, what do they represent and how do you prevent bad actors from accessing and repurposing these technologies for harm.” 

AI can solve intractable problems 

Thanks to advances in AI, the discovery of new drugs could soon shrink from an average of 10 years to a matter of months, said Hassabis. 

This follows the success of AlphaFold, an AI programme developed by DeepMind, a subsidiary of Google’s parent company Alphabet, to predict the 3D structure of proteins.  

This was a feat that had eluded researchers for the better part of five decades until AI came along – what the MIT Technology Review termed “one of biology’s grand challenges”. 

“Being able to predict a protein just from the genetic sequence would be enormously useful in drug discovery, understanding disease, understanding life,” said Hassabis. 

An AI-fuelled breakthrough means that the 200 proteins known to science have all been “folded” recently, to predict their 3D structures, by researchers in basically one year – an effort that before AI might have taken a billion years, Hassabis estimated – using AlphaFold. 

“Over a million biologists have used AlphaFold. If you know the structure of a protein, you can target a drug compound to bind to the surface and block or change the protein’s behaviour.” 

Subscribe to the GovInsider Bulletin for the latest public sector and innovation updates.

Do humans need to remain in the loop? 

Whether it is the discovery of new materials, creating more accurate weather-prediction systems, or solving big problems in mathematics that could change the game for other domains, AI will be behind the breakthroughs of the not-so-distant future. 

“What I see over the next few years is human experts using these amazing AI to further their knowledge. For now, it's the human experts that still have to come up with the hypotheses and the problems they want the systems to solve, and to find the connections between whatever insights these tools reveal. 

“But maybe in a decade-plus time, [the AI will] be able to start formulating their own conjectures.  

“If you combine one of those systems with language models, maybe we can just ask them for explanations about what they're doing just like you would with a top human scientist,” said Hassabis. 

In 2022, Harvard Business Review wrote that the current opaque nature of AI means that the technology has a long way to go before being ready to make unsupervised decisions.  

Finetuning still needed for generative AI models 

On 26 February, Alphabet reportedly lost US$90 billion in market value after Google’s Gemini AI was accused of producing racially inaccurate images of historical figures , and refused to answer certain moral questions.  

Google CEO Sundar Pichai called the issues that saw the AI feature taken offline temporarily “completely unacceptable”. 

On a case that should get AI model developers to sit up and listen, Hassabis noted that Gemini was given a new multi-modal function that could deal with not just text data sources, but also audio, music, image, video and code. 

He called the errors “one of the nuances that comes with advanced AI that as a field, we’re all grappling with.  

“As Google, we serve 200-plus countries, every country around the world, so you don't know where the users coming from and what their background is going to be or what their context they're in.  

“So you want to show a very universal range of possibilities there. But of course, that well-intended feature was applied too bluntly,” he said, adding that he hoped the issues would be resolved shortly, within weeks. 

Will the phone still be a phone? 

As more governments develop their digital services to be mobile-friendly, Hassabis predicted that the phone of the near future might not look like the phones of today. 

The next five years is, to him, “an incredible opportunity for mobile and telcos” with the advent of AI.  

“I think we’re really going to have next-generation smart digital assistants that are useful in our everyday lives… and I think there will be questions about what the right device type is.” 

As AI features such as Gemini or Circle to Search come to market, Hassabis asked, “Is the phone even going to be that perfect form factor? 

“Maybe we need glasses or some other devices so that the AI system can actually see a bit of the context that you’re in, and so be even more helpful.”