Mira Murati团队和桥水基金一起搞了个AI筛选新闻的实验，先用GPT、Claude只有50%准确率，专家写提示词到75%还是不够，最后微调模型干到了84.7%，成本还低了13倍多。
Mira Murati的Thinking Machines Lab与全球最大对冲基金Bridgewater合作，测试AI用于投资新闻筛选。GPT、Claude、Gemini在六项过滤测试中平均准确率约50%。专家投资者编写提示词后准确率升至74-76%，但仍低于80%的信任阈值。通过TML的Tinker API微调开放权重模型，准确率达到84.7%，错误率比最佳前沿模型降低29.8%，且每任务成本仅为前者的1/13.8。
Mira Murati's Thinking Machines Lab and Bridgewater, the world's largest hedge fund, published joint...
Mira Murati's Thinking Machines Lab and Bridgewater, the world's largest hedge fund, published joint results on using AI for a basic but important task in investing: Deciding which news deserves an analyst's attention. First, Bridgewater tried the frontier models. GPT, Claude, and Gemini variants averaged around 50% across six filtering tests. Then, expert investors wrote the prompts themselves. Accuracy climbed into the mid-70s. Still shy of the 80% the investors said they'd need before trusting a system in daily work. Then came the fine-tune using TML's Tinker API, with Bridgewater training an open-weight model on its experts' real judgement calls. The results: 84.7%, with mistakes down 29.8% versus the top frontier model tested, and at a 13.8x lower per-task cost. Murati, on X: "Experts improving AI that empowers experts." 💬 4 🔄 0 ❤️ 4 👀 1880 📊 4 ⚡