Fireworks AI [Dev Day] Fireside Chats - with Malte Ubl (CTO, Vercel)

um how would you describe Vzero in one sentence v V V V V V V V V V V V V V V V V V V V V Zero is this is such a good question because that's changed over time we first thought we were building a tool for making um for helping front end developers um build web pages and then we figured out no it's actually um the perfect tool to turn a back-end engineer into a fullstack engineer me personally I you couldn't chase me across the block

and I wouldn't get that Figma file into a web page right but now I can um and then finally we were discovered no what it actually is is a tool that turns PMS into people who build prototypes and which turns designers into people who build web pages right and so it's it's all these things in one thing yeah I saw Ray nodding his head we need that so I still remember about a year ago we have being in a room and discussing uh how we

can help like like visualize vis and make that happen at that time it's literally a moonshot project right uh it is really like amazing vision if we can but we don't know and we think it's very far away from realizing it right so what makes you the determination of starting a moonshot project while at that time the model capability was not there um and yeah so what drives that yeah I think we

um Vzero is now in a quite competitive market but um that space actually only started I think last October maybe right and and Vzero is actually more than a year older than that and so we actually started when um I'm not sure that GPD4 had come out it was definitely way too expensive to use and so we still the first version of of VZ shipped on on GPD 3.5 as a as a base model and I still remember actually being in the room because I think at the time everyone who

was interested in coding and AI was thinking okay code web pages they're just text and these models are good at generating text let's give them the task of of making of designing web pages and whenever you tried that and maybe some of you have on GPD 3.5 it would do something but it would be horrible um and so but we didn't give up and at some point someone literally like and it was so because we were a remote company but we happened to be in the same room so we

raised our hand like guys I figured it out i changed the prompt to ask for Tailwind and so Tailwind is this new way of writing HTML and CSS in one bundle and it was such a coincidence because Tailwind is a relatively new technology was just old enough to get popular just before the knowledge cut off date of GPD 3.5 and so because it intertwined HTML and CSS in one kind of inline notion these older models were able to reason

about it way more than I'm going to write a CSS file here I'm going to write an HTML file here and so it actually worked right and so that kind of that insight led us to go from something that just wasn't um a good thing to something that's really good and then later obviously as models got better this whole the product category got even better but it was it was an interesting journey to live yeah it sounds like serendipity but I think there there's there's some kind of essence here also like what's your opinion of using offtheshelf API versus you know create

this data fine where we discuss a lot today with other guests and kind of um you know have your own model yeah absolutely so the the I I already mentioned that the way we think about Vzero's model is a is as a composite model so the harsh reality is that if you want to be competitive in this space today for large tasks you have to use Sonnet they're just the best um but the thing is they're not going to be the best anymore in two months maybe it's