Terence Tao Guest Post Calls for Public Frontier AI Infrastructure for Science

Summary

A guest post by Dimitris Koukoulopoulos, published on Terence Tao’s blog, argues that the scientific community needs a public institutional response to the rapid spread of AI-assisted research. Its central proposal is a “CERN for AI-assisted science”: publicly funded frontier AI systems with conversational and agentic interfaces, substantial compute and storage, and free baseline access for researchers. Projects needing exceptional resources would compete for larger allocations and coordinated multi-agent systems, using a mechanism comparable to peer-reviewed access to major scientific facilities. The author says dependence on a small number of private companies creates several risks. Sensitive or classified data and unpublished ideas may be exposed to systems whose retention, training and reuse policies researchers cannot independently control. Private-company incentives may also favor rapid demonstrations over verification, attribution, exposition and human understanding. The post points to the possibility that AI-assisted results could eclipse years of early-career work within days and create a two-tier research system in which company-affiliated researchers receive access to unreleased models. The proposed public facility would combine controlled frontier models, large-scale computing, technical staff and a simple user-facing research assistant that could help explore hypotheses, search literature and coordinate semi-autonomous agents under human direction. The article identifies Canada and the European Union as a possible starting partnership, citing Canada’s C$2 billion, five-year sovereign AI compute commitment, including up to C$1 billion for public supercomputing, and EU plans involving RAISE, AI Factories and up to €10 billion for AI Gigafactories. It says the two sides have agreed to explore cooperation on fundamental AI research, agentic scientific systems, advanced infrastructure and public-interest models. The author acknowledges unresolved questions about staying near the frontier, training or combining models, governance, funding, resource allocation, bureaucracy, stagnation and capture. The argument is presented as an opinion: the key choice is whether future scientific AI infrastructure becomes a public good or remains structurally dependent on private companies.