A $500 RL fine-tune of a 9B open model beat frontier models on catalog review (fermisense.com)

In a groundbreaking development, companies like Bridgewater Associates and Harvey have successfully fine-tuned a 9 billion parameter open-source model using reinforcement learning (RL), significantly outperforming leading frontier models like GPT-5.5 and Claude Opus 4.8 on specific tasks. Bridgewater's model, trained on labels from expert investors, reduced mistakes by approximately 30% during document relevance assessment for investment purposes. Similarly, Harvey's legal agent, which navigates complex tasks such as transaction due diligence, has demonstrated superior performance in handling long-horizon workloads that traditional models struggled with. This trend highlights a pivotal shift in the AI/ML community towards leveraging open-source models combined with proprietary task-specific data. Notably, Intercom's AI agent, Fin Apex, trained on billions of customer interactions, has also achieved higher resolution rates than leading models, all while reducing operational costs. These examples underscore a broader movement towards tailored AI solutions that are more efficient and effective in real-world applications, suggesting that proprietary data and reinforcement learning stages may constitute a new best practice for developing AI systems in specialized domains.