The Center for Data Innovation spoke with Lin Qiao, co-founder and CEO of Fireworks AI, a California-based startup that offers enterprises a platform to integrate generative AI into their business operations. Qiao spoke about the company’s mission, the importance of fine-tuning AI models, and what makes Fireworks’ application programming interface (API) unique.
Martin Makaryan: What does Fireworks AI do and what was the inspiration behind the startup?
Lin Qiao: Fireworks AI offers companies and organizations from various industries the generative AI platform they need to build their custom models and elevate their business to the next level. We do not train foundation models from scratch, but rather help our clients access and fine-tune existing AI models and adjust them to their needs through our API. Customers can interface with more than 100 models, be they “off-the-shelf” models, open-source options, or even customized models, including large language models (LLMs) and image generation models. For example, we provide access through our API to the Llama-3 LLM that Meta recently released, which a developer can easily fine-tune on our platform and deploy within minutes to assess the quality of fine-tuning.
The other co-founders and I come from backgrounds in the AI space where we worked for several years before deciding to launch Fireworks AI. I was working at Meta, where I not only helped build AI models but also the surrounding infrastructure to support AI systems, and I worked with big names like Walmart and Disney. During this time, I realized that many enterprises often do not know how to integrate AI into their business model and lack the kind of extensive human and technical resources that big companies have. Because of this lack of scale, their progress in harnessing AI is slow. We felt that we could have a positive impact by bringing our expertise to help a broader array of enterprises use AI to improve their products and services. Our mission is to reduce the time it would take such organizations to effectively use AI from years to weeks or even days through our generative AI platform.
Makaryan: Why is it important to fine tune AI models?
Qiao: Generative AI uses foundation models, which developers train on all kinds of word knowledge and data that is available to use. I predict that at some point, regardless of whether these models are open-source or closed-source, they will converge both in terms of quality and size. This is because there is a finite amount of data, and the model architecture that big developers have is more or less similar. This is why the gap between the big foundation models will start shrinking at some point, making it imperative to use proprietary data to fine-tune and customize models for specific cases. Every enterprise is different when it comes to its business model, goals, operations, interests, and needs. Fine-tuning becomes a critical aspect of harnessing AI, and this is where Fireworks AI comes in. We are not a professional service provider; our goal is to provide the automation and the tools that companies need to use their data for their own good.
Our approach makes it easier for companies of all sizes, especially those that do not have extensive AI infrastructure, to easily experiment with models while keeping their costs down, in part thanks to the fact our API can help limit the number of parameters on large models. For instance, rather than running the full version of GPT-4 with over 1 trillion parameters, they can use our solution to run a customized, smaller version with something closer to 7 billion parameters. Our platform can make developing their personalized AI models from 20 to 120 times cheaper compared to other services in the market. We have a variety of optimization techniques that allow us to pass on these cost savings to developers.
Makaryan: What makes Fireworks AI unique as a software as a service (SaaS) platform?
Qiao: I think our mission and our approach to building a better open AI for enterprise is what sets us apart. Essentially, the goal of both big names like OpenAI and smaller companies like us is to solve the thousands of problems in the world through AI, but we take radically different approaches. Instead of choosing a one-size-fits-all approach, we want to address different problems and different needs with a diverse range of models that we train on custom curated data to yield the best quality results.
Makaryan: How will new AI breakthroughs shape your company’s vision and future?
Qiao: I think that future advancements in AI will make us increasingly more relevant and increase demand for custom-built models. I believe that generative AI is going to empower new innovative businesses, disruptive ideas that will challenge the big players in each sector. For app developers, there are tectonic shifts that change the calculus on how to provide the best user experience to support their business. Of course, there are some aspects of app development that are constant, like the need for graphic processing units (GPUs) to power computing. Nonetheless, I think the overall impact of AI breakthroughs will be a significant transformation of market competition across industries, which will prompt enterprises to compete to stay relevant and succeed. I think this drive for innovation will make platforms like ours more valuable.
Makaryan: What has been the biggest challenge for you as an innovator?
Qiao: The challenges I have faced have evolved over time and were very different when I was working at Meta and now when I am the CEO of an AI startup. I think as a leader and innovator, the biggest challenge for me now is to balance between innovation, which requires flexibility, and production quality, which often requires low latency. This balancing act becomes harder as AI continues to transform the market. There are many decisions that we must make at Fireworks AI to ensure we are being aggressive with our mission and our product.