Anthropic, Google, Microsoft and OpenAI launch the Frontier Model Forum

A body formed by four model developers

On 26 July 2023, Anthropic, Google, Microsoft and OpenAI announced the formation of the Frontier Model Forum ^[1]. The Forum was set up as an industry body focused on ensuring the safe and responsible development of frontier AI models ^[1]. The announcement presented it as a shared venue where the participating companies would work on safety questions connected to the most capable machine-learning systems ^[1].

How it defines a frontier model

The Forum defined frontier models as large-scale machine-learning models that exceed the capabilities of the most advanced existing models and can perform a wide variety of tasks ^[1]. The definition is comparative rather than fixed: a model belongs to the category by reference to what already exists, and by the breadth of tasks it can handle ^[1]. Nothing in the definition ties the term to a particular model size, training method or benchmark score ^[1].

Stated objectives

The Forum listed four objectives: advancing AI safety research, identifying best practices for frontier models, sharing knowledge with policymakers, academics and civil society, and supporting AI applications for societal challenges ^[1].

It said it would draw on member companies' technical and operational expertise to pursue this work, including by advancing technical evaluations and benchmarks and by developing a public library of solutions ^[1]. In its account of what it planned to focus on, the Forum named three areas: identifying best practices, advancing AI safety research, and facilitating information sharing among companies and governments ^[1].

The Forum also said it would establish an Advisory Board to help guide its strategy and priorities ^[1].

Who can join

Membership is open to organisations that develop and deploy frontier models, show a strong commitment to frontier model safety and are willing to contribute to joint initiatives ^[1]. The founding members were Anthropic, Google, Microsoft and OpenAI ^[1]; in May 2024, Amazon and Meta joined the Forum ^[2].

What the announcement set out

Read together, the stated objectives describe three kinds of work: research on AI safety, the writing down of practices that developers of frontier models should follow, and the movement of information outward to policymakers, academics and civil society and between companies and governments ^[1]. The concrete outputs named at the launch were technical evaluations and benchmarks, and a public library of solutions ^[1]. The Advisory Board was described as the mechanism for setting strategy and priorities ^[1].

The Forum therefore sits at the point where several companies that build large-scale models describe, in a common document, how such models should be defined, which safety questions they intend to work on, and who is allowed to join them in that work ^[1].