Lack of infrastructure for public interest AI

Public interest AI requires public support

Addressing the gaps in infrastructure and institutional support for public interest AI is a priority for The Collaborative. This problem definition outlines our understanding of the problem and how we see it evolving. We expect this statement to evolve as we learn and in response to changes in the field.

More so than most transformative technologies throughout history, the AI industry is almost entirely privatised. The state seeks a standard-setting and accountability role but is not present either at the frontier or in the adoption of the technology as a builder. Therefore, public interests tend to be displaced in favour of private interests.

Frontier AI needs physical, natural, and financial capital at a huge scale, at times with adverse environmental, economic, and social consequences. Public interest AI systems can be built with a fraction of the resources hyperscalers demand. Smaller, frugal, open models, powered by high-quality datasets show that an alternative exists.

Towards a different contest

An alternative where money alone is not the key is built on trust. It is a question of earning the trust of the people, institutions, and communities who are the stewards of high-quality, sensitive data, like doctors and patients, who have legitimate concerns about their data serving private interests.

An alternative that is less about resources and competition, and more about resourcefulness and partnerships, is built on collaboration. Innovations and transformational ideas exist in the public interest AI field, but they do not yet make up a whole that is greater than the sum of its parts. The field must go from silos to a movement that enables public interest actors to build AI alternatives that serve people and communities at scale.

An alternative that moves past the winner-take-all race at the frontier is built on different rules. Vast compute, unfettered access to data, and the best talent money can buy are rules for a contest that only a few dominant players can participate in, let alone win.

The public interest demands a different contest, with different rules and a different notion of what winning means.

Public interest actors need connective tissue, within and across groups, to become active builders and owners of this technology. They must not be made to feel their only options are feeding proprietary data flywheels or missing out on the benefits. With infrastructure and support, doctors and patients, educators and learners, public administrators and citizens can build and govern the technology they need.

The public interest demands a different contest, with different rules and a different notion of what winning means. Not vast compute, but frugal, more sustainable compute. Not generalized models, but smaller ones, built on high-quality, specific data. Not a sole focus on the frontier, but focusing on adoption and spread, understanding the value of the specific expertise sitting in industries and sectors in Germany, Nigeria, Brazil, and elsewhere. And not only private capital, but public resources — public finance and budgets, alongside public data, compute, and institutions.

Learn more about the problems we are focused on: