Julien Launay, Co-founder & CEO, Adaptive ML | Slush 2025
summary
Adaptive ML provides an on‑prem/cloud platform that uses reinforcement learning with synthetic data to specialize open LLMs for agentic tasks, then runs inference efficiently. A tuned JMA 4B model reaches GPT‑4.1‑level accuracy on bilingual (English/Korean) compliance checks while achieving a 4B‑parameter footprint, 7x latency reduction to hundreds of milliseconds, and roughly one‑tenth the cost (typical savings 50–90%) at trillions of monthly tokens.
AI-written summary · about
chapters (4)
- 0:00The rise of AI agents
- 2:51Reinforcement learning value
- 4:49Adaptive ML platform
- 5:15SK Telecom case study
related talks
30mAI Application Architecture: Building for Scale and Efficiency
38mUnraveling Multimodality with Large Language Models
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
36mProduction-Ready AI Agents: From LLMs to Small Language Models
32mLLM Future Is Multi-Model | Erik Davtyan #JSConfAM25- 12mKeynote: How LLMs Have Changed the Way Organizations Do AI
47mOperationalization of AI/ML models and use cases for ChatGPT Q&A
39mAccelerating Open-Source RL and Agentic Inference with vLLM