以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Bridgewater

4 atoms · 跨 2 天 · 首见 2026-06-30 · 最近 2026-07-02

三色: 🟦 fact 4 · 🟥 take 0 · stance ▲1/▼0/◆2

来源: X 4

时态: fresh:4

标签: 好数字:2 · 好信源:1 · 好思考:1

2026-07-02 · 测试 Gemini Claude GPT 文档过滤任务准确率
Naive prompts scored around 50%. Expert-written prompts pushed accuracy to 78%.
tags: 好数字 · → daily
value: qty=50% (naive prompts) / 78% (expert-written prompts)
2026-07-02 · 微调 Qwen3-235B 达到准确率
they fine-tuned Qwen3-235B on Tinker instead. 84.7% accuracy. 29.8% fewer mistakes than the best frontier model. At 1/14th the inference cost.
tags: 好数字 · → daily
value: qty=84.7% accuracy, 29.8% fewer mistakes than best frontier model
2026-07-02 · 数据清洗方法
Their fix: train a model on the noisy dataset, then run it back over its own training data. Any example the model disagreed with got routed to senior investors, because either the example was genuinely hard or the label was wrong. The model's own confusion became a detector for bad labels.
tags: 好思考 · → daily
2026-06-30 · fine-tuned a model to classify financial docs
Bridgewater fine-tuned a model to do it reliably and cheaply
tags: 好信源 · → daily