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.
Summary
This talk is for product people, UX designers, researchers, and leadership preparing to work on AI products. How do we plan and build helpful and impactful AI products? What is different in this new world? Products that use AI as a core part of the experience are often very easy to prototype, but really hard to build and make actually useful and impactful. A lot of the knowledge on how to build these is locked up in foundation model companies and engineering teams in startups. Peter van Dijck, expert in both AI and IA, will provide an easy to understand framework for understanding how to think about this new world we all live in now, how to plan for effective, helpful and impactful AI products, how to think about model capabilities, AI architectures and evaluations. You will learn vocabulary and frameworks to understand what all the acronyms mean, and how to think about them.
Key Insights
-
•
AI product design is best understood through a model, context, and experience framework.
-
•
Models are stateless and require full context to be sent with each interaction.
-
•
Prototyping to explore model capabilities is critical before building full products.
-
•
Evals, automated and manual, are essential to measure model output quality and decide if it’s good enough.
-
•
Models gain enhanced abilities through post-training on specific capabilities like reasoning and conversational skills.
-
•
Synthetic data generation enables scalable training and evaluation by producing many good and bad examples.
-
•
Tool use and function calling empower AI agents to autonomously perform multi-step tasks.
-
•
Agents like Cloud Code gain product-market fit by making model reasoning and actions visible to build user trust.
-
•
UX principles remain vital in AI product development, especially when designing evals and assessing experience.
-
•
In high-risk domains, explicit risk identification and eval development are necessary, alongside clear ownership of go/no-go decisions.
Notable Quotes
"Today’s best AI products are designed by engineering first, but UX and product people have to join the party."
"Models are stateless; they take input and give output and forget everything immediately after."
"You want to trust the model, which means not overly instructing it but giving it the right context."
"Evals are the key to knowing when an AI’s output is good enough."
"Model companies post-train models on capabilities by showing them millions of good and bad examples."
"Tool use and function calling are what make agents possible, letting models call tools and reason through tasks."
"Synthetic data is exceedingly useful because you can generate it at scale for training and evals."
"The reason Cloud Code has such product market fit is because it shows what it’s doing, building trust."
"UX skills especially shine in defining realistic, user-centered evals and collaborating with experts."
"If a capability exists and is useful, it’s very likely the model will improve significantly on it over time."
Dig deeper—ask the Rosenbot:
More Videos
"Product managers learn to attach at the end of every sentence something about money — no money, no attention."
Rich MironovHow Can Product Managers and UXers Help Each Other (and Why are Product Folks so Annoying Sometimes)?
December 6, 2022
"Stop just throwing AI at the end of the process; integrate it strategically across the whole product and content lifecycle."
Noz UrbinaRapid AI-powered UX (RAUX): A framework for empowering human designers
May 1, 2025
"Getting executive buy-in is so critical to the success of research, yet it remains a challenge for many."
Caroline VizeThe State of UX: Five Lessons from 2021 to Accelerate Digital Experience in 2022
March 9, 2022
"Design producers love their creative autonomy and don’t want to be dragged out of Figma into new tools."
Aurobinda Pradhan Shashank DeshpandeIntroduction to Collaborative DesignOps using Cubyts
September 8, 2022
"We cannot be everywhere. So is the goal to conduct more studies faster? And we said no, it’s not. That’s poison."
Jen Cardello Jennifer OttoLearning Velocity—The Insights Speedometer
September 16, 2021
"Constant communication is necessary; agreement emails can’t be one-and-done."
Alicia MootyDesign Staffing Models
September 30, 2021
"Some of the older content has discoverability problems, which is typical with blogs."
Jorge Arango[Demo] How to re-categorize content at scale using LLMs
June 5, 2024
"AI makes data more accessible to everyone but won’t replace the need for validated, empathetic UX research."
Rima Campbell Amrit S BhachuIncrease Productivity and Drive Business Impact
September 24, 2024
"We need more co-conspirators who actively use their privilege to interrupt microaggressions and bias."
Lisanne NormanWhy I Left Research
March 27, 2023
