Karina Nguyen: evaluation

Transcript context

…Because everyone's in firing their managers. "Who needs managers anymore?" That's what I hear now. Just kidding. It's interesting that so much of your time was spent on teaching product teams how evals integrate and how important it is. And I've heard this a few times and I haven't personally experienced it yet, so I think it's an important thread to follow is just how writing these evaluations is going to become increasingly an important part of the job of product teams, especially when they're building AI features and working with LLMs. So can you just talk a bit more about what that looks like? Is it sitting there with an Excel spreadsheet basically showing, "Here's the input, here's the output, here's how good the result was"? Talk about what that actually looks like very practically. It certainly depends on what you're developing, but there are various types of evaluations. Sometimes I do ask product managers, or there's also new roles that we have, model designers, to go through some of the user feedback maybe or think of various user conversations that should have triggered... Under these circumstances, it should trigger Canvas. And then you have this ground truth label of, "Okay with this conversation it should look trigger Canvas, under this conversation it should not trigger Canvas." And you have this very deterministic kind of eval that for decision-making behaviors is like this. When we were launching tasks, for example, how do you make correct schedules is actually really hard for the model. But we built out some of the deterministic evaluations that is like, "Okay, if the user says 7:00 PM, the model should say 7:00 PM. " So if you can have deterministic evals whether it's pass or fail. And the way it works is all the... Sometimes I ask product managers to just go create a double sheet, have different tabs and what's the current behavior, what's the ideal behavior and why, and some notes. And sometimes they usually use it with evals, sometimes we use it for training. Because if you give the spreadsheet to o1 model, it can probably figure out how to teach itself a good behavior. And I think there are second type of evals that is more prevalent is human evaluations. And you can have specific trainers or you can have internal people to when you have a conversation of the prompt and then you have various completion of models, you choose the win rate. Which model is the best? Which model produce the highest quality comment or edit? And then you can have continuous win rates. And as you develop new models it should always win over the previous models. So it depends on what you want to measure. So interesting. Basically what I'm hearing, and there's something I'm learning about as I talk to people, is product development might move from this, "Here's a spec PRD, let's build it together and then cool, let's review it. Are we happy with this?" From that to, "Hey, AI, build this thing for me and here's what correct looks like," and I'm spending all my time on what does correct look like on evals essentially.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.