World's Fair 2025
Five hard earned lessons about Evals — Ankur Goyal, Braintrust
Overview
This talk emphasizes the critical role of robust evaluation (evals) in the development and deployment of AI systems. It argues that evals should be actively engineered, not passively accepted, and that they are essential for both playing offense by identifying new use cases and defense by ensuring quality. The core thesis is that by treating evals as a continuous engineering process, organizations can better adapt to rapid model advancements and ship higher-quality AI products.
Who should watch
- AI Engineers
- Product Managers
- Builders working with LLMs
- Teams looking to improve AI product quality
- Those struggling with model updates and regressions
Key takeaways
- Effective evals provide tangible value, demonstrated by the ability to integrate new models within 24 hours and a clear path for incorporating user feedback into testing.
- Evals should be used proactively to explore and validate potential use cases before product launch, rather than solely for regression testing.
- High-quality evals require significant engineering effort; datasets must be continuously reconciled with real-world usage, and scoring functions should be custom-built to reflect specific project needs.
- Modern agentic systems rely heavily on context, with tool definitions and their outputs consuming a significant portion of the token budget, necessitating careful design for LLM success.
- Teams must engineer their products and mindsets to be agile, ready to capitalize on new model releases that can fundamentally change capabilities.
- Optimizing AI performance requires a holistic approach, considering the entire system including data, tasks (prompts, tools), and scoring functions, not just isolated prompt improvements.
- The output format of tools can significantly impact LLM performance; for instance, YAML can be more token-efficient for LLMs than verbose JSON.
Notable quotes
*The best data sets are those that you can continuously reconcile as you actually experience what happens in reality.*
*You have to think about tools in terms of what the LLM wants to see and how you can use exactly what you present to the LLM to make it work really well.*
*You want to create evals so that if there's something ambitious that you want to do in the future, you are very well prepared when a new model comes out to just push a button and find that out.*
Up next · Evals first
Watch next
Where the practice is headed once measurement is table stakes.
The Future of Evals - Ankur Goyal, BraintrustUnofficial community note. Prefer the recording for nuance.