Bret Taylor: belief

Transcript context

…I'm extremely bullish on the productivity gains from AI, but I do think the tools and products right now are somewhat immature and it's quite counterintuitive. So, for example, almost every software engineering firm I know uses something like Cursor to help their software engineers. Most people use Cursor right now as a coding autocomplete, though they have a lot of agentic solutions and there's a lot of... OpenAI has Codex and Cloud has... I can't remember the Anthropic products. So there's lots of agentic agents coming as well. One of the interesting things because the technology is immature, the code it produces often has problems. There's a lot of people approaching this to actually realize those productivity gains because as any engineer who's written a lot of code will tell you, it's pretty easy to look at and edit and fix code you wrote. Reviewing other people's code or particularly finding a subtle logical error in someone else's code is actually really hard. It's actually much harder than editing code that you wrote yourself. So if the code produced by a coding agent is often incorrect, it actually can take a lot of cognitive load and time to fix it. And in fact, if you end up producing lots of issues with your customers, you could end up producing a lot of features, but actually, is like mucking up the machine a little bit and having something that's not ideal. There's a couple of techniques that I think are interesting. First, I think there's a lot of AI starts now working on things like code reviews. I think this idea of self-reflection in agents is really important. Having AI supervise the AI is actually very effective. Just think about it this way, if you produce an AI agent that's right 90% of the time, that's not that great, but how hard would it be to make another AI agent to find the errors the other 10% of the time? That might be a tractable problem. And if that thing's right 90% of the time, just for argument's sake, you can wire those things together and have something that's right 99% of the time. So it's just a math problem and it turns out that you can make something to generate code, you can make something to review code and you're essentially using compute for cognitive capacity and you can layer on more layers of cognition and thinking and reasoning and produce things increasingly robust. So I'm very excited about that. The other thing though is root cause analysis. So we have an engineer at Sierra who exclusively focuses on the model context protocol server serving our cursor instance. And our whole philosophy is, if cursor generated something incorrect, rather than just fixing it, try to root cause it. Try to get it so the next time Cursor will produce the correct code and essentially, is context engineering. What context did Cursor not have that would've been necessary to produce the right outcome? Try to get it so the next time Cursor will produce the correct code and essentially, is context engineering. What context did Cursor not have that would've been necessary to produce the right outcome? So I think people who are trying to get productivity gains in departments like software engineering need to stop waiting for the models to magically work if they want to see the gains now. And you really have to create root cause analysis and systems and say, how do we go root cause every bad line of code and actually give the right context and produce the right system so the models can do it today? Over time that'll probably be less necessary and you'll have less context engineering necessary to do it, but you really have to think of this as a system and I think people are waiting for the models to just magically get better. And I'm like, well that will happen eventually, but if you want the gains now you got to put in the work. That's essentially why applied AI companies exist. And the work is non-trivial, but you can do it. And so, for customers using platforms like Sierra, yeah, AI agents aren't perfect, but we're creating a system that lets customers create a virtuous cycle of improvement. If you want to go from a 65% automated resolution rate to 75%, we have a billion tools to let AI help you do that, identify opportunities for improvement, figure out why people are frustrated, what new capabilities can we add to our agent to improve the resolution rate? And you let AI put the needles at the top of the haystack on your behalf and I think that's really the way to optimize these systems. I've never heard of this technique of improving Cursor by adding additional context. What's the actual way of doing that? You build an MCP server that everything runs through or is it like you add Cursor rules? What's the actual approach there?…

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