We have a full slate of upcoming events: AI Engineer London, AWS Re:Invent in Las Vegas, and now Latent Space LIVE! at NeurIPS in Vancouver and online. Sign up to join and speak!

We are still taking questions for our next big recap episode! Submit questions and messages on Speakpipe here for a chance to appear on the show!

We try to stay close to the inference providers as part of our coverage, as our podcasts with Together AI and Replicate will attest:

However one of the most notable pull quotes from our very well received Braintrust episode was his opinion that open source model adoption has NOT gone very well and is actually declining in relative market share terms (it is of course increasing in absolute terms):

Today’s guest, Lin Qiao, would wholly disagree. Her team of Pytorch/GPU experts are wholly dedicated toward helping you serve and finetune the full stack of open source models from Meta and others, across all modalities (Text, Audio, Image, Embedding, Vision-understanding), helping customers like Cursor and Hubspot scale up open source model inference both rapidly and affordably.

Fireworks has emerged after its successive funding rounds with top tier VCs as one of the leaders of the Compound AI movement, a term first coined by the Databricks/Mosaic gang at Berkeley AI and adapted as “Composite AI” by Gartner:

Replicating o1

We are the first podcast to discuss Fireworks’ f1, their proprietary replication of OpenAI’s o1. This has become a surprisingly hot area of competition in the past week as both Nous Forge and Deepseek r1 have launched competitive models.

Full Video Podcast

Like and subscribe!

Timestamps

* 00:00:00 Introductions

* 00:02:08 Pre-history of Fireworks and PyTorch at Meta

* 00:09:49 Product Strategy: From Framework to Model Library

* 00:13:01 Compound AI Concept and Industry Dynamics

* 00:20:07 Fireworks' Distributed Inference Engine

* 00:22:58 OSS Model Support and Competitive Strategy

* 00:29:46 Declarative System Approach in AI

* 00:31:00 Can OSS replicate o1?

* 00:36:51 Fireworks f1

* 00:41:03 Collaboration with Cursor and Speculative Decoding

* 00:46:44 Fireworks quantization (and drama around it)

* 00:49:38 Pricing Strategy

* 00:51:51 Underrated Features of Fireworks Platform

* 00:55:17 Hiring

Transcript

Alessio [00:00:00 ]: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner at CTO at Danceable Partners, and I'm joined by my co-host, Swyx founder, Osmalayar.

Swyx [00:00:11 ]: Hey, and today we're in a very special studio inside the Fireworks office with Lin Qiang, CEO of Fireworks. Welcome. Yeah.

Lin [00:00:20 ]: Oh, you should welcome us.

Swyx [00:00:21 ]: Yeah, welcome. Yeah, thanks for having us. It's unusual to be in the home of a startup, but it's also, I think our relationship is a bit unusual compared to all our normal guests. Definitely.

Lin [00:00:34 ]: Yeah. I'm super excited to talk about very interesting topics in that space with both of you.

Swyx [00:00:41 ]: You just celebrated your two-year anniversary yesterday.

Lin [00:00:43 ]: Yeah, it's quite a crazy journey. We circle around and share all the crazy stories across these two years, and it has been super fun. All the way from we experienced Silicon Valley bank run to we delete some data that shouldn't be deleted operationally. We went through a massive scale where we actually are busy getting capacity to, yeah, we learned to kind of work with it as a team with a lot of brilliant people across different places to join a company. It has really been a fun journey.

Alessio [00:01:24 ]: When you started, did you think the technical stuff will be harder or the bank run and then the people side? I think there's a lot of amazing researchers that want to do companies and it's like the hardest thing is going to be building the product and then you have all these different other things. So, were you surprised by what has been your experience the most?

Lin [00:01:42 ]: Yeah, to be honest with you, my focus has always been on the product side and then after the product goes to market. And I didn't realize the rest has been so complicated, operating a company and so on. But because I don't think about it, I just kind of manage it. So it's done. I think I just somehow don't think about it too much and solve whatever problem coming our way and it worked.

Swyx [00:02:08 ]: So let's, I guess, let's start at the pre-history, the initial history of Fireworks. You ran the PyTorch team at Meta for a number of years and we previously had Sumit Chintal on and I think we were just all very interested in the history of GenEI. Maybe not that many people know how deeply involved Faire and Meta were prior to the current GenEI revolution.

Lin [00:02:35 ]: My background is deep in distributed system, database management system. And I joined Meta from the data side and I saw this tremendous amount of data growth, which cost a lot of money and we're analyzing what's going on. And it's clear that AI is driving all this data generation. So it's a very interesting time because when I joined Meta, Meta is going through ramping down mobile-first, finishing the mobile-first transition and then starting AI-first. And there's a fundamental reason about that sequence because mobile-first gave a full range of user engagement that has never existed before. And all this user engagement generated a lot of data and this data power AI. So then the whole entire industry is also going through, falling through this same transition. When I see, oh, okay, this AI is powering all this data generation and look at where's our AI stack. There's no software, there's no hardware, there's no people, there's no team. I want to dive up there and help this movement. So when I started, it's very interesting industry landscape. There are a lot of AI frameworks. It's a kind of proliferation of AI frameworks happening in the industry. But all the AI frameworks focus on production and they use a very certain way of defining the graph of neural network and then use that to drive the model iteration and productionization. And PyTorch is completely different. So they could also assume that he was the user of his product. And he basically says, researchers face so much pain using existing AI frameworks, this is really hard to use and I'm going to do something different for myself. And that's the origin story of PyTorch. PyTorch actually started as the framework for researchers. They don't care about production at all. And as they grow in terms of adoption, so the interesting part of AI is research is the top of our normal production. There are so many researchers across academic, across industry, they innovate and they put their results out there in open source and that power the downstream productionization. So it's brilliant for MATA to establish PyTorch as a strategy to drive massive adoption in open source because MATA internally is a PyTorch shop. So it creates a flying wheel effect. So that's kind of a strategy behind PyTorch. But when I took on PyTorch, it's kind of at Caspo, MATA established PyTorch as the framework for both research and production. So no one has done that before. And we have to kind of rethink how to architect PyTorch so we can really sustain production workload, the stability, reliability, low latency, all this production concern was never a concern before. Now it's a concern. And we actually have to adjust its design and make it work for both sides. And that took us five years because MATA has so many AI use cases, all the way from ranking recommendation as powering the business top line or as ranking newsfeed, video ranking to site integrity detect bad content automatically using AI to all kinds of effects, translation, image classification, object detection, all this. And also across AI running on the server side, on mobile phones, on AI VR devices, the