Dr. Percy Liang (Stanford) discusses his team’s approach to language model evaluation under the AIQ program. They use AI models to automatically cover the test input space and find benchmarks or other metrics that transfer smoothly across scales. Both of these aims inform the third aim of safety evaluation.
• The views, opinions and/or findings expressed are those of the presenter and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.
Per Executive Order 14434: Inaugurating the Era of Super Intelligence, DARPA is working to update its communications to incorporate the term "super intelligence" as directed.
Source: DARPAtv
