Wendy Wang

speaker

80 appearances 1 recordings 1 series first heard Jul 2026 last heard 20 Jul

Wendy Wang’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

Recordings per month over the last 12 months — 1 in all, peaking in Jul 2026 with 1.

Appearances

newest first · ▶ plays the moment

Product Impact Podcast | Secrets to unlocking the value of AI · 17. Your AI Product is Failing — Microsoft's UXR Team Knows Why · 20 Jul 2026 podcast

So we developed a set of loss patterns.

This loss pattern taxonomy became the shared language.

across engineering teams, across product teams, data science teams, so that we could all rally around these loss patterns or gaps in our response quality that then the prompt engineering team can use to build evals around and also prompts that they can uh measure and then be able to run some flights to see how these different prompts that were

specifically drafted around these loss patterns.

Could move the key user metrics such as retention and session sessions per user.

So we really developed this flywheel from the user UXR evals all the way to moving user metrics in production.

So when we were able to s show the success of moving using the loss pattern taxonomy and being able to drive up our user metrics.

That was when I believe the the or you know, the organization was able to rally around this methodology and the the flywheel that we developed.

Sure, sure.

For UX researchers, I think we can see it like a an insight.

For example, we learned that a lot of times when users are coming to AI, they write their own prompt, but there's actually an underlying goal that the user is not explicitly stating in their prompt.

So when the AI is able to detect the underlying goal and go above and beyond and give them something in addition to what they asked for.

the user felt like the y this AI was able to anticipate what they needed and provide that without them explicitly asking for that.

So when we see that, you know, maybe our current model behavior wasn't actually doing that, there's a gap there, right?

So what we

Then we would develop a a loss pattern.

around that specific behavior.

And then use that terminology that every then all the teams can win every time they heard that terminology, they know, oh, okay, this is the behavior that we need to improve for the model.

Um so basically, and and then the taxonomy is basically a set of those loss patterns with specific examples.

And then also we made sure that the loss patterns were represented in prod.

Showing 21–40 of 80 · page 2 of 4