Aishwarya Naresh Reganti: prediction

Transcript context

…So let me follow that thread. We worked on a guest post together that came out a few months ago. And the thing that stood out to me most that stuck with me most after working on that post is this really key insight that building AI products is very different from building non-AI products. And the thing that you're big on getting across is there's two very big differences. Talk about those two differences. Yes. And again, I want to make sure that we drive home the right point. There are tons of similarities of building AI systems and software systems as well, but then there are some things that kind of fundamentally change the way you build software systems versus AI systems. And one of them that most people tend to ignore is the non-determinism. You're pretty much working with a non-deterministic API as compared to traditional software. What does that mean and why does that have to affect us is in traditional software, you pretty much have a very well-mapped decision engine or workflow. Think of something like Booking.com. You have an intention that you want to make a booking in San Francisco for two nights, et cetera. The product has kind of been built so that your intention can be converted into a particular action and you kind of are clicking through a bunch of buttons, options, forms, and all of that, and you finally achieve your intention. But now that layer in AI products has completely been replaced by a very fluid interface, which is mostly natural language, which means the user can literally come up with a ton of ways of saying or communicating their intentions. And that kind of changes a lot of things because now you don't know how your user's going to be here. That's on the input side. And the output is also that you're working with a non-deterministic probabilistic API, which is your LLM. And LLMs are pretty sensitive to prompt phrasings and they're pretty much black boxes. So you don't even know how the output surface will look like. So you don't know how the user might behave with your product, and you also don't know how the LLM might respond to that. So you're now working with an input, output, and a process. You don't understand all the three very well. You're trying to anticipate behavior and build for it. And with agentic systems, this kind of gets even harder. And that's where we talk about the second difference, which is the agency control trade-off. What we mean by that, and I'm kind of shocked so many people don't talk about this. They're extremely obsessed with building autonomous systems, agents that can do work for you. But every time you hand over decision-making capabilities or autonomy to agentic systems, you're kind of relinquishing some amount of control on your end. And when you do that, you want to make sure that your agent has gained your trust or it is reliable enough that you can allow it to make decisions. And that's where we talk about this agency controlled trade-off, which is if you give your AI agent or your AI system, whatever it is, more agency, which is the ability to make decisions, you're also losing some control and you want to make sure that the agent or the AI system has earned that ability or has built up trust over time. So just to summarize what you're sharing here, essentially, people have been building product, software products for a long time. We're now in a world where the software you're building is one, non-deterministic, can just do things differently. As you said, you go to booking.com, you find a hotel, it's going to be the same experience every time. You'll see different hotels, but it's a predictable experience. With AI, you can't predict that it's going to be the exact same thing, the thing that you plan it to be every time. And then the other is there's this trade-off between agency and control. How much will the AI do for you versus how much should the person still be in charge? And what I'm hearing is the big point here is this significantly changes the way you should be building product. And we're going to talk about the impact on how the product development lifecycle should change as a result. Is there anything else you want to add there before we get into that?…

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