LLM Evaluation Needs a Real Science — Alex Shaw & Sanmi Koyejo

Frontier Data Summit 2026: Alex Shaw and Sanmi Koyejo on why AI evaluation needs a real science, and how open infrastructure like Harbor can keep benchmarks honest. Alex Shaw (Laude Institute, creator of Harbor and co-creator of Terminal-Bench) and Sanmi Koyejo (Stanford) discuss why agent performance is so hard to predict, why aggregate benchmark scores mislead, how benchmarks get saturated or exploited, and what it takes to maintain evals over time. They cover broken RL environments, reward hacking, reusing rollouts, applying measurement science from other fields, and how open versus private benchmarks should coexist. IN THIS VIDEO - Why agent evaluation needs a measurement science, not just leaderboards - The predictability gap: what a benchmark number really tells you - Why aggregate scores mislead, and what to look at instead - Shortcuts, reward hacking, and benchmarks that models learn to exploit - Saturation, staleness, and keeping benchmarks maintained - Broken RL environments and how bad tasks can hurt training - Open, private, and layered approaches to benchmark access SPEAKERS Alex Shaw, Founding Member of Technical Staff, Laude Institute; creator of Harbor; co-creator of Terminal-Bench Sanmi Koyejo, Stanford University; Director, Stanford Trustworthy AI Research lab Francesca Vera, AI Community Manager, Snorkel AI (moderator) CHAPTERS 00:00 Introductions 00:47 Building a science around AI measurement 02:05 Why agents are hard to evaluate 04:18 Fast-moving models and what a score means 06:04 The gap between automated and ideal evaluation 07:07 The problem with aggregate scores 07:45 Benchmarks as measurement and coordination 10:16 Shortcuts and benchmark exploitation 12:01 Saturation and long-lived benchmarks 13:44 Continuous maintenance of benchmarks 16:27 Broken RL environments 18:11 Where eval infrastructure goes next 20:20 Measurement science from other disciplines 23:50 Evaluating the evals 25:37 How people use Terminal-Bench 27:52 Open vs. private benchmarks RESOURCES Harbor: https://harborframework.com Terminal-Bench paper: https://arxiv.org/abs/2601.11868 Terminal-Bench repo: https://github.com/harbor-framework/terminal-bench Snorkel AI Open Benchmarks Grants: https://benchmarks.snorkel.ai/