UCL will lead a new national research lab, the Science of Fundamental AI Research (SOFAIR) Lab, to develop open-source artificial intelligence systems that run on widely available computing hardware. Formally launched in London on June 23, the lab is part of a £60 million program funded by UK Research and Innovation's Engineering and Physical Sciences Research Council (EPSRC). Each of the two labs in the program will initially receive £8 million, with further funding contingent on a review in autumn 2026. SOFAIR brings together researchers from UCL, the University of Cambridge, the University of Oxford, and the University of Edinburgh to examine alternatives to the small number of architectures that underpin many current AI systems.
The lab's researchers will draw on computer science, mathematics, statistics, and neuroscience to build open-source models designed for commonly available hardware. The goal is to reduce dependency on a limited group of model providers and make advanced AI systems accessible to researchers, universities, and public bodies. This approach aligns with the broader push for AI for Science & Research, where reducing infrastructure barriers can accelerate academic and institutional adoption.
SOFAIR will test alternatives to current AI architectures
Professor David Barber of UCL Computer Science, who leads the SOFAIR Lab, said many current AI systems still suffer from basic issues such as inaccurate responses. "These systems often use similar underlying architectures, so SOFAIR will bring together the broader sciences and fresh ideas to create a new generation of open-source models," Barber said. "This will reduce dependency on the small number of model providers, boosting UK sovereignty and its position as a global player in AI."
The lab will build on UCL's existing AI work in healthcare, education, public services, science, and industry. UCL has linked the initiative with its AI for People and Planet framework, which covers development, safeguards, and social impact. The research focuses on structural limits that may not yet attract commercial investment, according to Professor Geraint Rees, Vice-Provost for Research, Innovation and Global Engagement at UCL.
"Commercial AI labs are, understandably, focused on near-term applications. SOFAIR exists to do the work they can't: fundamental research into the structural limits of current AI systems, work that benefits UK citizens but doesn't yet have a commercial market," Rees said.
Oxford-led BOLD Lab to examine how AI learns
The second lab, the British Open-ended Learning and Discovery (BOLD) Lab, will be led by the University of Oxford with involvement from UCL and Imperial College London. BOLD will investigate how AI systems can move beyond current training methods, working alongside people and responding to real-world complexity in physical environments. The research will also examine whether these tasks can be performed without relying on large centralized computing resources.
Both labs are expected to build partnerships across universities, industry, and the public sector. The program includes support for turning research into commercial products and university spinouts, though no commercialization targets or delivery timetables have been published.
Further funding depends on autumn assessment
EPSRC will assess the progress of both labs in autumn 2026 before deciding how the remaining £60 million program funding will be distributed. Professor Charlotte Deane, Senior Responsible Owner for the UKRI AI Programme and Executive Chair of EPSRC, said the labs will "back the bold, high-reward ideas that can shape the future of AI."
The launch of SOFAIR took place on what would have been Alan Turing's 114th birthday, at the Royal Academy of Engineering in London. Kanishka Narayan, the UK's AI and Online Safety Minister, formally opened the lab.
Why this matters for science and research professionals
The SOFAIR Lab's emphasis on open-source models that run on commodity hardware could lower the entry barrier for scientific computing. Researchers in fields like genomics, climate modeling, or materials science often rely on specialized AI tools; more accessible architectures may reduce cost and increase reproducibility. The initiative's focus on fundamental limits-rather than near-term product development-creates space for long-horizon work that complements existing academic AI research. With funding tied to a progress review in 2026, the lab's early outputs will be a signal of how seriously the UK invests in publicly oriented AI research.
Your membership also unlocks:
