Alessio Fanelli: recommendation

Transcript context

…I'll maybe throw AI under the bus a little bit here and just say that Claude wrote a lot of these example servers. No surprise at all. But I do think, sorry, I do think there's an interesting point in this that I do think people at the moment still to mostly still just apply their normal software engineering API approaches to this. And I think we're still need a little bit more relearning of how to build something for LLMs and trust them, particularly, you know, as they are getting significantly better year to year. Right. And I think, you know, two years ago, maybe that approach would have been very valid. But nowadays, just like just throw data at that thing that is really good at dealing with data is a good approach to this problem. And I think it's just like unlearning like 20, 30, 40 years of software engineering practices that go a little bit into this to some degree. If I could add to that real quickly, just one framing as well for MCP is thinking in terms of like how crazily fast AI is advancing. I mean, it's exciting. It's also scary. Like thinking, us thinking that like the biggest bottleneck to, you know, the next wave of capabilities for models might actually be their ability to like interact with the outside world to like, you know, read data from outside data sources or like take stateful actions. Working at Anthropic, we absolutely care about doing that. Safely and with the right control and alignment measures in place and everything. But also as AI gets better, people will want that. That'll be key to like becoming productive with AI is like being able to connect them up to all those things. So MCP is also sort of like a bet on the future and where this is all going and how important that will be. Yeah. Yeah, I would say any API attribute that says formatted underscore should kind of be gone and we should just get the raw data from all of them. Because why, you know, why are you formatting? For me, the, the model is definitely smart enough to format an address. So I think that should go to the end user. Yeah. I have, I think Alessio is about to move on to like server implementation. I wanted to, I think we were talking, we're still talking about sort of MCP design and goals and intentions. And we've, I think we've indirectly identified like some problems that MCP is really trying to address. But I wanted to give you the spot to directly take on MCP versus open API, because I think obviously there's a, this is a top question. I wanted to sort of recap everything we just talked about and give people a nice little segment that, that people can say, say, like, this is a definitive answer on MCP versus open API.…

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