Lenny Rachitsky: belief

Transcript context

…Which shows that agency is what matters. If you have a product manager that has an idea, there's no reason for why that idea cannot be more well fleshed out. How many times do you have a product manager that just continualize ideas, but it just feels like they are extremely unsure on how to execute on it? They just want to say things for sake of saying things? But for the people that have ideas and a lot of, I guess, agency, they can go out and prove out what they want without any sort of external resources. I think even more acutely for product folks listening to this, it's the salesperson coming to you being like, "Hey, I want this thing. It's going to help me with my sales team." And you're like, "I don't have a million things to build. I don't have time for this." And so that problem goes away, which I think will make a lot of product leaders really happy. The model that this is sitting on, is it Sonnet? Yeah. So just to break down how it ultimately works, we have a model that does planning. And I would say right now Sonnet is a really, really good planning model. I think OpenAI's GPT-4o is also good. But the crazy thing is what we try to do is we try to make the Anthropic based model or Sonnet model try to do as much of the high level planning as possible. And then what we try to do internally is run all the models necessary to do high quality retrieval for the agent. As you could see, the agent needed to understand what the rest of the code base ultimately did. We actually make sure we run models to actually chunk up the entire code base and understand the code base so that obviously it would not be a good idea if we had a 100 million line code base to send that entire code base to Anthropic. First of all, you couldn't do that. That's over 1.5 billion tokens of code. So obviously that would be three or four orders of my actually larger than the largest context lens right now. But you also wouldn't want to do that from a cost and latency standpoint too. So that's one. And the second piece that you saw was the model is able to very quickly make edits to the software as well. We have custom models that we built that are post trained on top of popular open source models that can make these edits really, really quickly to the code base. And the reason why you would want to do that is it's A, faster, and B, also that model can actually have more of the code base in context too. So it can be better at applying changes than even Anthropics model too. So I think the way we like to think about it is, our only goal is how do we build the best product possible? How do we build the best product possible and how do we make the ceiling as high as possible? And we will go out and build models and train models wherever necessary. But if we're not going to be good at a task and we think the open source is better or Anthropic's better, we'll go and just use the open source or Anthropic.…

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