dotAI 2024 - Pierre Stock - The future of Edge Agents

Filmed at dotAI on October 17, 2024 in Paris. More about the conference on https://www.dotai.io Edge devices such as smartphones offer unique opportunities to harness the power of LLMs in scenarios where privacy and latency are key. In this talk, we’ll investigate techniques to obtain the best performance given latency/memory constraints for models in the 1-3B parameters range. We’ll go through Mistral’s edge offering and detail our view on how to make the most of customization and function calling as this scale. Who is Pierre Stock? Pierre Stock is a Research Scientist and first employee at Mistral AI, mainly focusing on Code and Edge models and their efficient deployment to production. Before that, he was a Research Scientist at Meta and developed privacy-related machine learning techniques such as Federated Learning and Differential Privacy specifically for LLMs. Pierre Stock was a resident PhD student at Facebook AI Research (FAIR, CIFRE program) for 3 years.