HAI Seminar with Sanmi Koyejo: Beyond Benchmarks – Building a Science of AI Measurement

The seminar provides a compelling and well-structured argument for reforming AI evaluation practices. Koyejo, a respected researcher, effectively diagnoses the limitations of current benchmarking approaches, particularly the gap between what benchmarks measure and the high-stakes claims they are used to support. He illustrates this with the GPQA example, contrasting a direct claim about answering questions with a broader claim about graduate-level reasoning, and rightly points out that the latter is less supported by the evidence. The historical context, from early punchcards to modern benchmarks, helps frame the evolution of the field. The proposal to integrate psychometric principles, such as Item Response Theory, is well-founded and offers a concrete path toward more rigorous measurement. The discussion of measurement validity, drawing on the history of psychological testing, is particularly insightful and relevant. The talk is not without its limitations: it is primarily an opinion piece, presenting ideas and frameworks rather than original experimental results. Some concepts, such as ‘amortized computation’ and ‘predictability analysis’, are mentioned but not deeply elaborated, leaving the audience wanting more detail. The speaker acknowledges this by pointing to future work and open problems. The sources cited, including the GPQA benchmark and the AI Index, are credible and relevant. The talk is aimed at an academic audience familiar with AI and evaluation, but the core ideas are accessible to a broader technical audience. Overall, the seminar is a valuable contribution to the ongoing discussion about AI measurement, offering both a critique of current practices and a constructive vision for the future.

The title accurately reflects the content: the speaker discusses moving beyond static benchmarks toward a more rigorous measurement science for AI.

8/10

The talk is given by a Stanford professor, well-grounded in psychometrics and AI evaluation, with references to published papers and benchmarks. The argumentation is coherent and supported by examples, though it remains an opinion piece without original experimental data.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction: The importance of measurement in AI and the historical context of benchmarking.
• 10:00Discussion of internal and external statistical validity of benchmarks, citing ImageNet interventions.
• 20:00Case study of GPQA benchmark: construction, validation, and the distinction between direct and ambitious claims.
• 30:00Introduction of the 'crisis of measurement' and the gap between benchmark design and high-stakes use.
• 40:00Proposal of three ideas: measurement validity, psychometric tooling, and predictability of interventions.
• 50:00Discussion of validity framing and lessons from psychometrics, including the history of intelligence testing.
• 1:00:00Case studies in safety assessment and capability measurement, and implications for governance and policy.

• GPQA: A Graduate-Level Google-Proof Q&A Benchmark — Referenced as an example of a modern benchmark with rigorous construction and validation.
• For Better or Worse, Benchmarks Shape a Field — Cited to support the argument that benchmarks shape the field's vision and progress.
• AI Index Report 2024 — Mentioned as a source of analysis on the state of responsible AI and benchmark limitations.

• AI Index Report 2024 — Supports the claim that there is a lack of standardized evaluation for AI models.