Weaviate Podcast Cartesia AI with Karan Goel - Weaviate Podcast #113!
10 snips
Jan 28, 2025 
Karan Goel, leading force at Cartesia AI, discusses groundbreaking advancements in text-to-speech technology and neural network architecture. He shares insights into State Space Models, designed to overcome traditional model limitations. The conversation dives into the evolution of long context processing and the importance of emotional intelligence in AI communications. Karan also highlights the significance of many-shot in-context learning and its applications in education, as well as the development of user-friendly on-device models.
AI Snips
Chapters
Transcript
Episode notes
Cartesia's Founding Story
- Karan Goel shares how Cartesia AI spun out of Stanford research on state space models.
- Their team grew from five co-founders to 30+ people focused on building new architectures and products.
State Space Models Beat Attention Bottlenecks
- Attention mechanisms have a quadratic computation cost, limiting sequence length efficiency.
- State space models offer a scalable alternative allowing models to handle potentially limitless context.
Multimodal Data Enhances AI Memory
- Multimodal data like audio enriches machine memory by providing continuous, real-world contextual signals.
- This motivates building models that learn from complex sensory inputs over time.