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 #1: Let’s write your first AI eval
This video is featured in the Evals + Claude playlist.
Summary
In this talk, Peter Van Dijck of the helpful intelligence company will walk you through writing your first eval. You will learn the basic concepts and the tools, and write an eval together. This talk is hands on; you can follow along, and there will be plenty of time for questions.
You will go away with an understanding of the basic building blocks of AI evals, and with the confidence that you know how to write one. And more importantly, you’ll build some intuition, 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 consist of a task, a golden dataset with known correct outputs, and an evaluator that measures correctness.
-
•
Manual AI prompt testing is slow and inconsistent; automated evals accelerate and scale evaluation.
-
•
UX and product teams can and should learn evals as a practical, non-technical skill.
-
•
Creating your own golden dataset is essential and cannot be outsourced or fully automated.
-
•
Models are fixed once trained; improvements happen by refining prompts and context design, not retraining the model.
-
•
Evaluations measure task performance, not the underlying model itself, allowing comparison across models.
-
•
Outputting a confidence score from models is unreliable due to lack of internal memory and inconsistent scale interpretation.
-
•
Biases are baked into models during training via evals used in post-training refinement.
-
•
LLMs can be used to judge other LLM outputs to evaluate tasks with non-binary answers.
-
•
Effective eval work requires collaboration across data analysts, engineers, subject matter experts, and UX/product teams.
Notable Quotes
"Evals are like a way to define what good looks like."
"The model was baked and once it’s baked, it does not learn again until they bake a new one."
"You need to be looking at the data. Nobody wants to, but that’s core work."
"Without a golden dataset, you have to build the golden dataset yourself."
"We’re not teaching the model anything; we’re improving our prompts and context."
"Confidence scores from the model are not a good idea because the model has no memory."
"Biases are baked in through the evals used during model training and post-training."
"LLMs judging other LLMs might sound crazy, but if you do it right, it works."
"Evals are a product and UX skill; learning them lets you make these systems do what you want."
"There is a large and growing capability overhang in these models we haven’t discovered yet."
Or choose a question:
More Videos
"What does a hysterectomy cost? $2,000, $13,000, or $40,000? It depends who you ask and who pays."
Meghan BausoneSystems Thinking and Design Innovation: Working with Leverage Points in Rural Maternal Health Systems
April 17, 2026
"We can't track actual benefit uptake after users click through to official government sites, only clicks and visits."
Alan Williams Rose DeebDesigning essential financial services for those in need
February 10, 2022
"Proving the value of design is a fool's errand. What matters is your value."
Shahrzad SamadzadehWhat Is My Value? Two Takes and Some Mistakes
January 8, 2024
"In remote teams, we get prototypes up in front of folks in a short time window for collaborative comments."
Adam Cutler Karen Pascoe Ian Swinson Susan WorthmanDiscussion
June 8, 2016
"The biggest gain is about getting value out sooner, with an 80 percent efficiency improvement from design to running app."
George Abraham Stefan IvanovDesign Systems To-Go: Reimagining Developer Handoff, and Introducing App Builder (Part 2)
October 1, 2021
"We talked all about resilience, curated by myself yesterday."
Bria AlexanderDay 3 Welcome
September 25, 2024
"Before AI, research teams spent over a quarter of their time consulting about research questions, often repeating themselves."
Mike OrenImproving Democratized Research with CustomGPTs and Gems
February 25, 2026
"I realized that dedicating my time and energy into something that wasn’t going to be used was not a good use of my time."
Megan BlockerA Selectively Scrappy Approach to ResearchOps
November 8, 2018
"Audacity means being willing to do things that seem dumb or risky because that’s how you innovate."
Deanna Washington Tim Allen Jeff Courcelle John Maeda Matt Raw Erica TjaderScaling Success: Paving the Path from DesignOps to VP
October 4, 2023
