Extract from Katherine B. Forrest’s article “The Knowns and Unknowns of Crossing the AI Frontier”
One of the ways we can tell that technological developments in AI are moving fast— really fast—is the current dialogue relating to AI “Frontier” models. A Frontier model is a “highly capable model” that “could possess capabilities sufficient to pose severe risks to public safety.” (Anderljung, et al., “Frontier AI Regulation: Managing Emerging Risks to Public Safety,” November, 2023).
The White House Executive Order (“E.O.”) on AI, issued on Oct. 30, 2023, refers to these models as “dual-use” foundation models. The U.K. Government Office for Science has published a special report on “Future Risks of Frontier AI,” and the AI Seoul Summit, held in May 2024, was followed by an “International Scientific Report on the Safety of Advanced AI.”
There is broad agreement that the capabilities of large scale neural networks such as those that enable the generative AI foundation models that the world woke up to with the issuance of ChatGPT in the late fall of 2022, are rapidly increasing.
Models trained on huge data sets that have billions of “parameters,” and can perform a broad range of tasks unsupervised, can approach or cross the threshold into a frontier model. Parameters are the values assigned to the data the neural network is trained on and establish relationships between pieces of data. As additional data is used to train a model, the parameters can adjust.
