World's Fair 2025
Evals Are Not Unit Tests — Ido Pesok, Vercel v0
Overview
This talk introduces the concept of evals at the application layer, distinguishing them from traditional unit tests. It emphasizes that Large Language Models (LLMs) can be unreliable, leading to unexpected failures in AI applications even when basic functionality appears to work. The core thesis is that robust evals are crucial for building reliable AI products by systematically testing and measuring performance across a spectrum of user-driven scenarios.
Who should watch
- AI Engineers building and deploying LLM-powered applications.
- Product Managers seeking to understand and improve the reliability of AI features.
- Developers encountering unexpected behavior or hallucinations in their AI products.
- Anyone interested in moving beyond basic LLM prototyping to production-ready applications.
Key takeaways
- LLMs are inherently unreliable, and this unreliability scales from simple apps to complex systems, posing a significant challenge for product development.
- Traditional unit tests are insufficient for AI applications because they cannot account for the probabilistic nature and potential for hallucinations of LLMs.
- Evals should be visualized as a basketball court, where the basket represents a successful outcome and the distance from the basket signifies difficulty. Data points represent user queries, and their position on the court indicates success or failure.
- Understanding your "court" is paramount; focus evals on data and queries that your users actually care about, avoiding "out-of-bounds" tests.
- Collect diverse data signals for evals, including user feedback (thumbs up/down), log analysis, and community forums, to gain a comprehensive understanding of application performance.
- Separate constants (data) from variables (tasks) in your eval setup to improve clarity, reusability, and scalability, especially when testing different system prompts or pre-processing logic.
- Prioritize deterministic, pass/fail scoring for evals to simplify debugging and ensure clarity across teams, though human review may be necessary in complex cases.
- Integrate evals into your CI/CD pipeline to catch regressions and measure the impact of changes before they reach production.
Notable quotes
*Improvement without measurement is limited and imprecise.*
*Evals give you the clarity you need to systematically improve your app.*
Up next · Evals first
Watch next
Field notes from running evals in production: what actually breaks.
Five hard earned lessons about Evals — Ankur Goyal, BraintrustUnofficial community note. Prefer the recording for nuance.