UK urged to embrace cheaper AI models

The UK should embrace cheaper AI models for many public services instead of automatically reaching for “extraordinarily expensive” frontier products, the head of the Alan Turing Institute has urged.

George Williamson, whose organisation plays a leading role in developing the UK’s national policy on artificial intelligence, also pressed British authorities to widen debate over the technology’s rollout. Many people feel “excluded” from the conversation over AI safety, even as fears about its risks grow, he said.

Williamson, who took over in May after a turbulent period at the Turing, said his priority was for the 11-year-old institute to focus on developing national resilience in the fields of AI, misinformation and tech sovereignty.

“As we try to think about how we might deploy AI at scale in our public services and for national good, some of these technologies have the potential to be extraordinarily expensive given the capability locked up in them,” Williamson said in an interview.

“Work we have done has shown that you can get very near frontier performance on much more application-specific models,” he told the FT, referring to models that are geared to a particular function rather than a wide spectrum of capabilities. “And if you work through the logic of it, that is much more computationally efficient — and at the end is much more financially effective.”

Recent hacks by AI agents, such as OpenAI’s incursion into Australian government websites, have underscored the technology’s potential effects on big national IT systems and forced politicians to recognise the rising stakes, Williamson said.

The incidents showed the risks in how the technology “plays out in society”, even as it promised to deliver a “huge opportunity for good” in areas such as public services, national security and drug discovery, he said.

“I speak to a lot of people who are not at the Turing, just pretty ordinary people [and] they feel very excluded from this discussion about AI,” Williamson said. “Because really lots of these issues are moral, they’re ethical, they are about policy or regulation.”

The hacking incidents highlighted how AI technologies that “behave in a certain way in a lab” sometimes work in “quite different and unexpected ways when deployed in real external settings”, Williamson said.

The Turing — along with other UK institutions such as the AI Security Institute — had a role in developing “mitigations” and a “science of verification and assurance” for these dangers, he added.

Williamson was previously head of His Majesty’s Government Communications Centre, where he worked to create national security tools and technologies. Before that, he was in the diplomatic service for more than 20 years, including a spell as the Foreign, Commonwealth and Development Office’s director-general in technology.

“For me the Turing sits at just the right place in the national system,” Williamson said. “It’s close to government but it’s independent of government. It works with the national security community, but with a whole set of other parts of society [too].”

Williamson’s predecessor Jean Innes stepped down last year, after then technology secretary Peter Kyle branded the institute a “failing organisation”. The government had been pressing the Turing to sharpen its focus on defence, national security and sovereign AI capabilities.

The institute was roiled by staff unrest during Innes’ tenure, with some protesting about the way the organisation was being run. Innes disputed the characterisation of the Turing as suffering from low morale and widespread disillusionment with senior management.

Williamson said institute teams that had worked in silos were now planning “much bolder, more ambitious” cross-disciplinary work. He said he was holding employee sessions termed “tea with George”, which sounded “a bit quaint” but were an effort to create an “open setting where people tell me what’s on their minds”.

“Most management problems in organisations are really down to communication and whether you feel confident enough to communicate early with staff,” Williamson said. “We have communicated very early, so we’ve been doing lots of work over the last couple of months on strategic direction, talking about national resilience, talking around how we might achieve that.”