Amjad Masad: evaluation

Transcript context

…Interestingly, if you want to be good at debugging, there's a lot you need to understand, which is basically what you're saying is that's the subset of things to understand is things that break. And to do that, you have to understand how it all works. What are servers? What are APIs? All these things. Okay, so we've been talking about how this is very good right now, building a prototype, building a v1, MVP, people can use it, you can deploy it. You deploy this app, people can start using it, and there's a scale it can reach. Do you see a future where you can build a Salesforce sized business fully Replit or other tools that can scale to hundreds of billions of dollars of value? Or is there just going to always be some limit of like, you need actual engineers and designers sitting on this thing, building it, thinking it? Awesome. If my law is directionally correct, even if the months are not, I'm not exactly right that the duration is correct, you're going to see a compounding effect of the power. It's actually quite hard to convince yourself. But if you really convince yourself that we are on a massive scale of improvement in AI, then the answer is yes. And it's absurd to my engineering mind that I'm saying it is, but know Ray Kurzweil, this futurist talks about how exponentials are really hard for humans to grasp. And so actually when we started building the agent, I told the team, it's easy and we fall in this trap before. It's easy to build and optimize for today. In '22, we built Copilot-like thing and autocomplete. We train our own models, we optimize the hell out of them. But at some point, that modality was kind of not the right modality, which is the autocomplete modality. And the right modality is actually this, I think for now, as being able to chat inside the programming environment and for the agent to create things for you. But in order for us to make that bat, a year ago the models were actually not there. The models could not do this, but we were like, okay, we're going to build for the models that are landing in six months. And truly six months later, the model started to land that are capable of this, of the reasoning that we needed and whatever. And so that was like saw it if you want, which is, oh, wow, we switched to it and the reasoning improved so much. And six months later, you have a son of you too. And so it's really almost like a six months cadence. And so if we're really on this trajectory, then I would say next year, you're able to just scale and maybe you get thousands of users paying you. The AI can do maintenance. We already showed the AI doing SQL queries and doing migrations, so I will be able to do maintenance, debugging, things like that. I think where it gets really tough is that when you're hitting scale and you want to architect a system that is resilient, and so that means you would start sharding databases and you would start using different queue systems and components and things like that. And I think the AI needs to have access to the entire suite of tools to be able to do this. And I think that's going to be the next bottleneck. And I think the AI needs to be a lot more reliable at doing that. But I could imagine whatever, five years from now, someone running a billion dollar company with zero employees where it's like the support is handled by AI, the development is handled by AI, and you're just building and creating this thing that people are finding valuable and are paying you for it. That being said, it's worth thinking about the economics of it. If the cost of software goes down a lot, then what is the price that you can charge on software? So can you actually build the next Salesforce if anyone can generate Salesforce? t the economics of it. If the cost of software goes down a lot, then what is the price that you can charge on software? So can you actually build the next Salesforce if anyone can generate Salesforce? And then the question is, and this is why I emphasize being generative, because I think then the thing that will make you better is by being able to iterate and improve the thing really quickly and generate new ideas.…

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