UCL lab will try to break AI’s dependence on tech giants
The Science of Fundamental AI Research Lab, known as SOFAIR, launched on 23 June at the Royal Academy of Engineering on Carlton Terrace, with UCL leading a partnership that includes Cambridge, Oxford, and Edinburgh. It is one of two new labs sharing government funding through the UKRI’s Engineering and Physical Sciences Research Council, with the aim of rethinking how AI systems are built rather than simply making the existing ones bigger.
Almost every AI tool a Londoner might use this year, from a chatbot drafting an email to a hospital algorithm flagging a scan, runs on some version of the same handful of underlying designs, built by a small number of companies with the computing power to train them. SOFAIR, based at UCL, has been given public money specifically to break that pattern.
The problem SOFAIR exists to address is one of concentration rather than capability. The SOFAIR Lab will focus on developing next generation open-source AI technologies that are more widely accessible.
The reasoning behind that goal is simple; modern AI relies heavily on a small number of popular architectures trained on vast data, requiring immense computing infrastructure, meaning that only a handful of organisations, largely outside the UK, currently have the resources to build systems at the frontier. SOFAIR’s job is to find designs that do not need that scale of infrastructure to work well.
The point is not that today’s systems are unreliable by accident. It is that most of them are variations on the same small set of designs, so any structural weakness in that design tends to show up everywhere at once, and open-source alternatives built differently are one way to test whether it has to be that way.
SOFAIR draws its researchers from computer science, mathematics, statistics, and neuroscience, rather than computer science alone, on the premise that the current bottlenecks in AI design might not be solved by more computer scientists doing more of the same work. The lab’s brief is to explore architectures capable of running on ordinary, widely available hardware rather than the specialised data centres that only a few companies can afford, which is the practical mechanism behind the accessibility goal, not just its stated ambition.
SOFAIR is one half of a two-lab programme. Its counterpart, the British Open-ended Learning and Discovery Lab, is led by Oxford with UCL and Imperial College London also involved, and tackles a related but separate problem: rather than what AI is built from, BOLD is working on how AI systems learn, aiming for models that can work alongside people and operate in the physical world without relying on the same centralised computing power SOFAIR is also trying to move away from. Each lab has received an initial £8 million, with a further funding decision due after an assessment in autumn 2026.
Whether that bet pays off will not be visible in a Londoner’s phone or GP surgery for some time. What UCL is promising, for now, is that the AI shaping those services in future will not have to depend on the same small handful of companies it does today.
For more London arts and culture, visit EyeOnLondon at eye-on-london.com
Follow us on:
Subscribe to our YouTube channel for the latest videos and updates!
We value your thoughts! Share your feedback and help us make EyeOnLondon even better!
