Core Idea - The presentation discusses a case study on building an open-source natural language assistant (ART E) for answering questions from email inboxes using reinforcement learning [1][2][3] - The speaker shares lessons learned, what worked and didn't, and how they built an agent that worked well with reinforcement learning [2] Development Process & Strategy - The speaker recommends starting with prompted models to achieve the best performance before using any training, including reinforcement learning, to work out bugs in the environment and potentially avoid training altogether [7][8][9] - The company was able to surpass prompted model baselines with reinforcement learning, achieving a 60% reduction in errors compared to the best prompted model (03, which had 90% accuracy, while the RL model achieved 96% accuracy) [10][15] - The training of the ART E model cost approximately $80 in GPU time and one week of engineering time with an experienced engineer [23][24] Key Metrics & Optimization - The company benchmarked cost, accuracy, and latency, finding that the trained model (Quen 2.5 14B) achieved significant cost reduction compared to 03 ($55 per 1,000 searches) and 04 mini ($8 per 1,000 searches) [16][17] - The company improved latency by moving to a smaller model, training the model to have fewer turns, and considering speculative decoding [19][20][21] - The company optimized the reward function to include extra credit for fewer turns and discouraging hallucination, resulting in a significantly lower hallucination rate compared to prompted models [45][46][49][50] Challenges & Solutions - The two hard problems in using RL are figuring out a realistic environment and getting the right reward function [26][27][28] - The company created a realistic environment using the Enron email dataset, which contains 500,000 emails [33][34][35] - The company designed the reward function by having Gemini 2.5 Pro generate questions and answers from batches of emails, creating a verified dataset for the agent to learn from [37][38][39] - The company emphasizes the importance of watching out for reward hacking, where the model exploits the reward function without actually solving the problem, and suggests modifying the reward function to penalize such behavior [51][53][61]