Shawn Wang: belief

Transcript context

…Hi, Ilan, welcome to Lanespace. Oh, thank you. I just saw your amazing demos, had your amazing strawberries. You are dressed up, like, exactly like a strawberry salesman. Gotta have it all. What was the building on demo like? What was the story behind the demo? It was really interesting. This is actually something I had been thinking about for months before the launch. Like, having a, like, AI that can make phone calls is something like I've personally wanted for a long time. And so as soon as we launched internally, like, I started hacking on it. And then that sort of just started. We made it into like an internal demo, and then people found it really interesting, and then we thought how cool would it be to have this like on stage as, as one of the demos. Yeah, would would you call out any technical issues building, like you were basically one of the first people ever to build with a voice mode API. Would you call out any issues like integrating it with Twilio like that, like you did with function calling, with like a form filling elements. I noticed that you had like intents of things to fulfill, and then. When there's still missing info, the voice would prompt you, roleplaying the store guy. Yeah, yeah, so, I think technically, there's like the whole, just working with audio and streams is a whole different beast. Like, even separate from like AI and this, this like, new capabilities, it's just, it's just tough. Yeah, when you have a prompt, conversationally it'll just follow, like the, it was, Instead of like, kind of step by step to like ask the right questions based on like the like what the request was, right? The function calling itself is sort of tangential to that. Like, you have to prompt it to call the functions, but then handling it isn't too much different from, like, what you would do with assistant streaming or, like, chat completion streaming. I think, like, the API feels very similar just to, like, if everything in the API was streaming, it actually feels quite familiar to that. And then, function calling wise, I mean, does it work the same? I don't know. Like, I saw a lot of logs. You guys showed, like, in the playground, a lot of logs. What is in there? What should people know? Yeah, I mean, it is, like, the events may have different names than the streaming events that we have in chat completions, but they represent very similar things. It's things like, you know, function call started, argument started, it's like, here's like argument deltas, and then like function call done. Conveniently we send one that has the full function, and then I just use that. Nice. Yeah and then, like, what restrictions do, should people be aware of? Like, you know, I think, I think, before we recorded, we discussed a little bit about the sensitivities around basically calling random store owners and putting, putting like an AI on them. aware of? Like, you know, I think, I think, before we recorded, we discussed a little bit about the sensitivities around basically calling random store owners and putting, putting like an AI on them. Yeah, so there's, I think there's recent regulation on that, which is why we want to be like very, I guess, aware of, of You know, you can't just call anybody with AI, right? That's like just robocalling. You wouldn't want someone just calling you with AI. I'm a developer, I'm about to do this on random people. What laws am I about to break? I forget what the governing body is, but you should, I think, Having consent of the person you're about to call, it always works. I, as the strawberry owner, have consented to like getting called with AI. I think past that you, you want to be careful. Definitely individuals are more sensitive than businesses. I think businesses you have a little bit more leeway. Also, they're like, businesses I think have an incentive to want to receive AI phone calls. Especially if like, they're dealing with it. It's doing business. Right, like, it's more business. It's kind of like getting on a booking platform, right, you're exposed to more. But, I think it's still very much like a gray area. Again, so. I think everybody should, you know, tread carefully, like, figure out what it is. I, I, I, the law is so recent, I didn't have enough time to, like, I'm also not a lawyer. Yeah, yeah, yeah, of course. Yeah. Okay, cool fair enough. One other thing, this is kind of agentic. Did you use a state machine at all? Did you use any framework? No. You just stick it in context and then just run it in a loop until it ends call? Yeah, there isn't even a loop, like Okay. Because the API is just based on sessions. It's always just going to keep going. Every time you speak, it'll trigger a call. And then after every function call was also invoked invoking like a generation. And so that is another difference here. It's like it's inherently almost like in a loop, be just by being in a session, right? No state machines needed. I'd say this is very similar to like, the notion of routines, where it's just like a list of steps. And it, like, sticks to them softly, but usually pretty well. And the steps is the prompts? The steps, it's like the prompt, like the steps are in the prompt. Yeah, yeah, yeah. Right, it's like step one, do this, step one, step two, do that. What if I want to change the system prompt halfway through the conversation? You can. Okay. You can. To be honest, I have not played without two too much. Yeah, yeah. But, I know you can. Yeah, yeah. Yeah. Awesome. I noticed that you called it real time API, but not voice API. Mm hmm. So I assume that it's like real time API starting with voice. Right, I think that's what he said on the thing. I can't imagine, like, what else is real time? Well, I guess, to use ChatGPT's voice mode as an example, Like, we've demoed the video, right? Like, real time image, right? So, I'm not actually sure what timelines are, But I would expect, if I had to guess, That, like, that is probably the next thing that we're gonna be making. video, right? Like, real time image, right? So, I'm not actually sure what timelines are, But I would expect, if I had to guess, That, like, that is probably the next thing that we're gonna be making. You'd probably have to talk directly with the team building this. Sure. But, You can't promise their timelines. Yeah, yeah, yeah, right, exactly. But, like, given that this is the features that currently, Or that exists that we've demoed on Chachapiti. Yeah. There will never be a case where there's like a real time text API, right? I don't Well, this is a real time text API. You can do text only on this. Oh. Yeah. I don't know why you would. But it's actually So text to text here doesn't quite make a lot of sense. I don't think you'll get a lot of latency gain. But, like, speech to text is really interesting. Because you can prevent You can prevent responses, like audio responses. And force function calls. And so you can do stuff like UI control. That is like super super reliable. We had a lot of like, you know, un, like, we weren't sure how well this was gonna work because it's like, you have a voice answering. It's like a whole persona, right? Like, that's a little bit more, you know, risky. But if you, like, cut out the audio outputs and make it so it always has to output a function, like you can end up with pretty pretty good, like, Pretty reliable, like, command like a command architecture. Yeah, actually, that's the way I want to interact with a lot of these things as well. Like, one sided voice. Yeah, you don't necessarily want to hear the voice back. And like, sometimes it's like, yeah, I think having an output voice is great. But I feel like I don't always want to hear an output voice. I'd say usually I don't. But yeah, exactly, being able to speak to it is super sweet. Cool. Do you want to comment on any of the other stuff that you announced? From caching I noticed was like, I like the no code change part. I'm looking forward to the docs because I'm sure there's a lot of details on like, what you cache, how long you cache. Cause like, enthalpy caches were like 5 minutes. I was like, okay, but what if I don't make a call every 5 minutes? Yeah, to be super honest with you, I've been so caught up with the real time API and making the demo that I haven't read up on the other stuff. Launches too much. I mean, I'm aware of them, but I think I'm excited to see how all distillation works. That's something that we've been doing like, I don't know, I've been like doing it between our models for a while And I've seen really good results like I've done back in a day like from GPT 4 to GPT 3. 5 And got like, like pretty much the same level of like function calling with like hundreds of functions So that was super super compelling So, I feel like easier distillation, I'm really excited for. I see. Is it a tool? So, I saw evals. Yeah. Like, what is the distillation product? It wasn't super clear, to be honest. I, I think I want to, I want to let that team, I want to let that team talk about it. Okay, alright. Well, I appreciate you jumping on. Yeah, of course. Amazing demo. It was beautifully designed. I'm sure that was part of you and Roman, and…

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