Aishwarya Naresh Reganti: evaluation

Transcript context

…Awesome. And you have a visual actually that we'll share of what this looks like. But just to kind of mirror back what you're describing, this idea of start with high control, low agency, the example you gave is the support agents just kind of giving suggestions, is not able to do anything, the user is in charge. And then as that becomes useful and you are confident it's doing the right sort of work, you give it a little more agency and you kind of pull back on the control the user has. And then if that's starting to go well, then you give it more agency and the user needs less control to control it. Awesome. I think the higher level idea here is with AI systems, it's all about behavior calibration. It's incredibly impossible to predict upfront how your system behaves. Now, what do you do about it? You make sure that you don't ruin your customer experience or your end user experience. You keep that as is, but then remove the amount of control that the human has. And there is no single right way of doing it. You can decide how to constrain that autonomy. I mean, a different example of how you could constrain autonomy is pre-authorization use cases. Insurance pre-authorization is a very ripe use case for AI because clinicians spend a lot of time pre-authorizing things like blood tests, MRIs and things like that. And there are some cases which are more of low hanging fruits. For instance, MRIs and block tests, because as soon as you know patient's information, it's easier to approve that and AI could do that versus something like an invasive surgery, et cetera, is more high risk. You don't want to be doing that autonomously. So you can kind of determine which of these use cases should go through that human and the loop layer versus which of the use cases AI can conveniently handle. And then all through this process, you're also logging what the human is doing because you want to build a flywheel that you could use in order to improve your system. So you're essentially not showing the user experience, not eroding trust, at the same time logging what humans would otherwise do so that you can continuously improve your system. So let me give you a few more examples of this kind of progression that you recommend. And the reason I'm spending so much time here is this is a really key part of your recommendation to help people build more successful AI products. This idea of start slow with high control and low agency and then build up over time once you've built confidence that it's doing the right sort of work. So a few more examples that you shared in your post that I'll just read. So say you're building a coding assistant, V1 would be just suggest inline completion and boilerplate snippets. V2 would be generate larger blocks like tests or refactors for humans to review. And then V3 is just apply the changes and open PRs autonomously. And then another example is a marketing assistant. So V1 would be draft emails or social copy, just like here's what I would do. V2 is build a multi-step campaign and run the campaign. And V3 is just launch it A/B tested auto-optimize campaigns across channels. Awesome. Yeah. And again, just to summarize where we're at, just to give people the advice we've shared so far. One is just important to understand AI products are different. They're non-deterministic. And you pointed out, and I forgot to actually mirror back this point, both on the input and the output. The user experience is non-deterministic.People will see different things, different outputs, different chat conversations, different maybe UI if it's designing the UI for you. And also the output obviously is going to be non-terministic. So that's a problem and a challenge. And then-…

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