Log in or create a free Rosenverse account to watch this video.
Log in Create free account100s of community videos are available to free members. Conference talks are generally available to Gold members.
Hands-on AI #2: Understanding evals: LLM as a Judge
This video is featured in the Evals + Claude playlist.
Summary
If you’re a product manager, UX researcher, or any kind of designer involved in creating an AI product or feature, you need to understand evals. And a great way to learn is with a hands-on example. In this second talk in the series, Peter Van Dijck of the helpful intelligence company will show you how to create an eval for an AI product using an LLM as a judge (when we use a Large Language Model to evaluate the output of another Large Language Model). We’ll have a look at how that works, but also dig into why this even works. Are we creating problems for ourselves when we let an LLM judge itself? This talk is hands on; and there will be plenty of time for questions. You will go away understanding when and how to use LLM as a judge, and build some product sense around how the best AI products today are built, and how that can help you use them more effectively yourself.
Key Insights
-
•
Evals are a foundational feedback loop defining what 'good' means for AI products, helping to measure and improve systems continuously.
-
•
Evaluating fuzzy, subjective AI outputs requires innovative approaches such as using LLMs as judges to score results.
-
•
Binary (yes/no) scoring is more reliable than rating scales with ranges because LLMs lack internal memory and consistency.
-
•
Starting evals early (week one of a project) drastically improves AI product outcomes, but many teams delay due to perceived complexity.
-
•
High-risk or important tasks should be prioritized for evals instead of attempting broad coverage.
-
•
Assigning a dedicated owner or 'benevolent dictator' for evals who works closely with domain experts accelerates feedback and quality.
-
•
Creating a written constitution of principles helps concretize AI behavior goals and guides prompt and model training.
-
•
Most current eval tooling is too technical, slowing iteration cycles and making expert involvement inefficient.
-
•
Custom feedback interfaces tailored to expert users significantly speed up evaluating AI outputs in domains like healthcare and law.
-
•
Diverse perspectives from UX, product, strategy, and domain experts are critical in defining and refining what 'good' means in AI systems.
Notable Quotes
"Evals are everywhere, right? Everybody's talking about evals. It is like one of the key things in developing useful AI products."
"You want to ask an LLM to evaluate the fuzzy stuff because there’s no black and white output."
"LLMs don’t have memory, so rating on a scale from one to five is pretty random. Better to have yes or no answers."
"One of the biggest problems in AI building is evolving your prompts and having a fast feedback loop."
"By starting to categorize risk in detail, you naturally lead to better prompts and better evals."
"A constitution is a very good exercise: write down your system’s principles and values to help guide its behavior."
"Use custom systems for experts to quickly review and rate outputs, making feedback cycles much faster."
"Evals define a shared definition of good with tests to measure it, and that is the secret sauce for building great AI products."
"Model companies are students in a classroom wanting good points—they’re happy to run external expert evals to improve."
"The more I work with evals, the more I think UX and product people need to be involved because of the need for diverse perspectives."
Or choose a question:
More Videos
"Jazz soloists draw from a lifetime of patterns, sometimes quoting TV theme songs like the Muppets or Sanford and Son."
Jim KalbachJazz Improvisation as a Model for Team Collaboration
November 6, 2017
"We look to work with authors we like and enjoy being around because writing a book is a long, collaborative journey."
Louis RosenfeldCoffee with Lou: Should You Write a (UX) Book?
March 7, 2024
"Building influence has been such a challenging part of growing in my own career. It takes a lot of communication and understanding others’ goals."
Catt Small Micah Bennett Brian Carr Jessica HarlleeWhat's Next for ICs: Exploring Staff and Principal Designer Roles
February 22, 2024
"Sharing your research skills broadly within your company can create unexpected opportunities to improve experiences."
Marieke McCloskeyUser Science: Product Analytics & User Research
March 11, 2021
"Wonder Studio automatically segments actors, maps their movements to a 3D model, and renders a synthetic avatar video."
Llewyn Paine[Demo] Deploying AI doppelgangers to de-identify user research recordings
June 5, 2024
"Matching is very much an art, not a science; there are no hard and fast rules for what counts as a good match."
Joshua NobleCasual Inference
October 6, 2023
"We’re biologically wired to respond to intellectual challenges the same way as physical threats."
Sara LogelYour Colleagues are Your Users Too
March 29, 2023
"Learning is a marathon, not a sprint, especially in the context of online conferences."
Bria Alexander Louis RosenfeldWelcome
January 8, 2024
"If you have to learn a workaround, you want to learn it once and reuse it again and again."
Sam ProulxOnline Shopping: Designing an Accessible Experience
June 7, 2023
Latest Books All books
Dig deeper with the Rosenbot
What steps help build trust and healing in communities affected by historical traumas during engagement?
How can play help overcome the ‘messy middle’ between research insights and decision-making in organizations?
Why is interoperability and API access critical in modern user research tooling ecosystems?