Lenny Rachitsky: evaluation

Transcript context

…Exactly. I feel there's a bunch of things that you actually have to get confidence in before you get to V3. And it's easy to get overwhelmed that, oh, my AI agent is doing these things wrong in a hundred different ways and you're not going to actually tabulate all of them and fix it. Even though you've learned how do you deal with the evaluation practices and stuff like that, if you're starting on the wrong spot, you are actually going to have a hard time correcting things from there. And when you start small and when you start with building a very minimalistic version with high human control and low agency, it also forces you to think about what is the problem that I'm going to solve. We use this term called problem first. And to me, it was obvious in the sense that that I do need to think about the problem, but it's incredible how well it resonates with the people that in all this advancements of the AI that we are seeing, one easy, slippery slope is to just keep thinking about complexities of the solution and forget the problem that you're trying to solve. So when you're trying to start at a smaller scale of autonomy, you start to really think about what is the problem that I'm trying to solve and how do I break it down into levels of autonomy that I can build later? So that is incredibly useful and we keep repeating this part and over and over with everyone we talk to. And there's so many other benefits to limiting autonomy because there's just danger also of the thing doing too much for you and just messing up your, I don't know, your database, sending out all these emails you never expected. And there's like so many reasons this is a good idea. Yep. I recently read this paper from a bunch of folks at UC Berkeley. Basically Matei Zaharia, [inaudible 00:21:54] and the folks at Databricks and it said about 74% or 75% of the enterprises that they had spoken to, their biggest problem was reliability. And that's also why they weren't comfortable deploying products to their end users or building customer facing products because they just weren't sure or they just weren't comfortable doing that and exposing their users to a bunch of these risks. And that's also why they think a lot of AI products today have to do with productivity because it's much low autonomy versus end-to-end agents that would replace workflows. And yeah, I love their work otherwise as well, but I think that's very in line with what at least we are seeing at my startup as well.…

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