A 100% Eval Pass Rate Is A Bad Sign

20 Oct 2026 · 1 min
0:53

In this brief dialogue, Hamel Husain explains that consistently achieving a 100% pass rate across multiple evaluations is a warning sign rather than a success. He argues that such perfect scores indicate that the evaluations are too easy or saturated, providing no useful feedback for improvement. Husain advises designing more challenging evaluation scenarios that avoid saturation, enabling meaningful "hill climbing"—gradual performance improvements through harder problems. The conversation also prompts a deeper discussion on the concepts of saturation and hill climbing in evaluation design.

More from Hamel

0:53
youtube.com

Why You Shouldn't Write 500 Evals First

17h ago
0:46
youtube.com

How Should AI Evals Be Used In Performance Reviews?

18h ago
0:52
youtube.com

This Support Bot Recommended A Competitor In The Exec Demo

1d ago
0:29
youtube.com

Overnight Agent Demos Are Vanity Metrics

1d ago
0:39
youtube.com

Your Golden Dataset Isn't Golden

2d ago
0:37
youtube.com

Why You Shouldn't Simulate Your Users

3d ago
24:25
youtube.com

How to Evaluate Coding Agents

3d ago
4:19
youtube.com

How To Get Your Team To Invest In Evals

5d ago
0:47
youtube.com

Stop Splitting One Task Across Multiple Agents

6d ago
3:23
youtube.com

How To Decide Who Owns AI Evals

6d ago
0:41
youtube.com

The Two Biggest Mistakes Teams Make With Evals

1w ago
See all content