Ep 50: Fireworks CEO Lin Qiao on Why There Won’t be a Single Model, Will Hyperscalers Win Inference & AI Use-cases with PMF
Summary
In the 50th episode of the podcast, Lin Qiao, co-founder of Fireworks.ai, explores the multifaceted landscape of artificial intelligence (AI) and its implications for future development. Qiao emphasizes that data derived from mobile-first strategies is foundational to AI innovation, driving algorithmic capabilities. She discusses the flawed notion that a single AI model can address all challenges, promoting the coexistence of specialized models tailored for distinct use cases. The evolution of model capabilities has been significant, with advancements being noted from models like GPT-3 to the current iterations. Additionally, she contrasts the agile nature of startups with the more structured approaches of enterprises in adopting AI, underscoring a trend towards specialized solutions tailored to unique market demands. Qiao addresses the increasing complexity of AI systems and emphasizes the need for co-design in both hardware and software to optimize performance. She highlights the potential of hyperscalers to dominate the AI infrastructure landscape and discusses the importance of function calling and model orchestration, which enhance model performance. Finally, she points out the growing trend towards smaller expert AI models and a shift in investment focus towards post-training processes, indicating an evolving paradigm in how AI technologies are integrated and maintained in businesses.
Key Takeaways
- 1Data as the Foundation of AI Innovation
- 2The Importance of Co-Design in AI Systems
- 3Diversity in AI Modeling
- 4Trends Among Startups vs. Enterprises
- 5Complexity in AI Systems
- 6The Evolution Toward Smaller Expert Models
- 7Function Calling and Model Performance
- 8Quality Assurance in AI Development
- 9Post-Training Optimization Strategies
Notable Quotes
"One is again, anchor back to our vision is we believe the direction is specialization and customization."
"Uh, and that's how we operate at Meta is kind of the research team and kind of info teams sit very closely to discuss trade-offs."
"Uh, and we believe there's much better, um, solution if you can customize, if you can steer, uh, if you have control, uh, that's neat here."
"They're starting from their product is open-ended... they start to get more and more clear opinion."
"Building a system is very complex. It's not like just regular single model as a service inference."
"The future lies in hundreds of small expert models."
"We have to solve a lot of quality-related problems... let the model talk with itself."
"When I started Fireworks, there was actually an active debate, is AI here or not? Because it's pre-kind of awareness of foundation models."
"So the other side effect is all these models, all Gen AI models are PyTorch models."
"You just build on top of it or you tune, right?"
"I think it's kind of the nature of the training process. The training process is a very opinionated process that you have to pick which subset of problem across like thousands of problems in the world you care the most about."
"So we believe the future are the hundreds of small expert models. Because when you shrink the problem down to a narrow space, it's inevitably much easier for small model to thrive in pushing the quality."