Logan Kilpatrick

Like I feel this way sometimes in some of Google's code bases, because they're large and there's lots of teams contributing to them.

And as you sort of try to form an intuition of how these different pieces of software are connected together, it's actually not easy to do that.

And that goes to this like emergent,

sort of framing of the problem around context engineering, where people are trying to build systems.

It's not based on the user prompt.

It's not like trying to do a bunch of optimization based on the user prompt.

It's trying to do a bunch of optimization on like getting the relevant data to actually help answer the user prompt.

And I think that's the large code base stuff like some of the most, you know, insert again your favorite AI coding tool.

This is some of the sweet spot and the magic of what they figured out is like how to actually go and get the right context to answer questions and use all these like more traditional code understanding tools like graph techniques and things like that to understand the relationship between different parts of a code base.

And I think that's actually working pretty well.

Like I could imagine that continuing to scale up and sort of be good enough for even the most large code bases to work really well with these tools.

Yeah, you're spot on.

I think this is part of the challenge for us from a product perspective is we do have this huge breadth of what you can do now from actually writing code and building a bespoke app to we have...

You mentioned the live API stuff, the stream real-time mode, and being able to build interactive sort of co-present AI experiences that don't just see text, but they can understand the audio that you're saying.

You can actually share your screen.

And back to this context engineering problem, what makes using AI products so hard in many ways is that you have to do, in most products, you have to do the context engineering yourself.

And this is actually one of the breakthrough projects

product experience things that a lot of these IDs have done that like has made this helpful is that actually all the context most or if not all the context, most of the context you need is in your ID somewhere.

So if you can just traverse through what's in your ID, you will eventually get all the right context to theoretically solve the problem that you have at hand.

That's actually not true on in most products like the the scope of if I'm trying to solve some like