论文22

Epoch 评测:模型撰写数据洞察表现

原标题:We also tasked models with creating an Epoch-style Data Insight, which is a concise article highlighting a single trend in AI. Fable 5.2 ...

Epoch AI@EpochAIResearch发布时间 (北京时间)

我们还让模型撰写一篇 Epoch 风格的数据洞察,即一篇聚焦单一 AI 趋势的简洁文章。Fable 5.2 和 Opus 5.5 提出了对 Epoch 读者来说有趣且有效的选题,但难以用简洁的方式呈现。 GPT-6.2 Astra 未能理解 Epoch 的受众,选择了一个过于小众的选题。 我们在之前对 GPT-6 Astra 的评测中也看到了同样的缺陷。这体现了我们在各模型中普遍看到的判断力不足。

中文译文尚未完成,先为你展示原文。

We also tasked models with creating an Epoch-style Data Insight, which is a concise article highlighting a single trend in AI. Fable 5.2 and Opus 5.5 came up with valid topics that would be interesting to Epoch’s audience, but struggled to present them in a concise way.

GPT-6.2 Astra failed to understand Epoch’s audience and chose an overly niche topic.

We saw the same flaw in our previous evaluation of GPT-6 Astra. It’s an example of the broader lack of judgment we see across models.