This talk explores methods for optimizing expressive, multi-modal policies, such as those based on flow-matching, with off-policy reinforcement learning. The speaker presents two novel algorithms, FQ-RL and CAM, designed to overcome the instability of backpropagation through multi-step generative models, enabling effective online self-improvement and adaptation for robotic manipulation tasks.
Continuous flow models map naturally to a Schrödinger equation, the fundamental equation of quantum mechanics. This discovery proves that a trained generative model can be efficiently simulated on a future quantum computer, enabling a new, more powerful type of access to its learned distribution for tasks like Monte Carlo estimation and structure discovery.
Mistral's Pavan (Voxtral lead) and Guillaume (Chief Scientist) discuss the new Voxtral TTS model, its novel architecture using flow matching for efficient, high-quality speech generation. They elaborate on Mistral's strategy of delivering specialized, open-weight models and the Mistral Forge platform, which empowers enterprises to leverage their proprietary data through fine-tuning for privacy, cost-effectiveness, and superior performance. The conversation also covers Mistral Small, the future of AI for science, and the company's commitment to open-source and foundational research, including formal proving as a proxy for long-horizon reasoning.
YC Visiting Partner Francois Chaubard and YC General Partner Ankit Gupta break down diffusion, the machine learning framework behind generative AI models like Sora and Midjourney. They discuss its core principles, trace its evolution from complex KL-divergence methods to the elegant simplicity of flow matching, and explore its vast applications beyond images, from protein folding to robotics, arguing it's a key component for future AI systems.