IT之家 9 月 23 日消息,AI 模型在数学、编程和推理能力上进步迅速,写作水平却停滞不前,甚至有所退步,Anthropic 的 Claude 同样也存在这一难题。
为何 Opus 4.6 会成为 Anthropic 最后一款写作表现出色的模型?负责 Claude 微调的 Anthropic 工程师杰克逊 · 克尼恩今天(23 日)对此给出了解释:后续模型的优化重点更多放在数学和代码能力上。
这些模型还接受了大量面向其他 AI 模型撰写工艺解释的训练,而这正是 Claude 逐步形成奇特表达方式,也就是所谓“Claude 腔”(Claudeish)的主要原因。克尼恩称,模型被“调整得适应 LLM 心理”,最终学会了写给 AI 模型看,而不是写给人看。
克尼恩把这种现象类比为只与其他自闭症人士交流的人。此类群体会逐步形成一套在群体内部沟通顺畅、外人却难以理解的表达方式。LLM 具备远超人类的工作记忆,也能捕捉更细微的信息,因而在训练中会逐步形成一种格外合适 AI 模型理解的写法,但在人类读者眼中,这种表达就成了过度密集的信息堆砌。
IT之家获悉,克尼恩认为,难题的根源在于强化学习的奖励机制。有些奖励机制着重提高 AI 模型的理解成效,有些则着重让人类更容易理解。数学和代码维度的训练越多,就越需要主动抵消这种倾向,对简洁、容易让人读懂的解释给予更多奖励。
克尼恩称,Anthropic 在 Opus 5.5 上找到了更好的平衡,“自 Opus 4.6 以来,我还没有对一款模型的写作表现这么满意过。”
当然,这样也并不代表 Opus 5.5 已经超越 Opus 4.6。“这是一个很难应对的难题,笔者会继续改进。”
本文来源:https://www.ithome.com/1/006/510.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT之家 news on September 23: AI models are advancing rapidly in mathematics, programming, and reasoning, but their writing skills have stagnated or even regressed — Anthropic's Claude faces the same problem.
Why did Opus 4.6 become Anthropic's last model with outstanding writing performance? Jackson Kernion, the Anthropic engineer responsible for Claude fine-tuning, explained today (the 23rd): the optimization focus of subsequent models shifted more toward math and coding capabilities.
These models were also trained extensively to write craft explanations aimed at other AI models — which is the main reason Claude gradually developed its peculiar expressive style, the so-called "Claudeish." Kernion said the models were "tuned to fit LLM psychology" and ultimately learned to write for AI models rather than for humans.
Kernion compared this phenomenon to people who only communicate with other autistic individuals. Such groups gradually develop a communication style that works smoothly within the group but is difficult for outsiders to understand. LLMs have working memory far exceeding that of humans and can capture subtler information, so during training they gradually form a writing style especially suited to AI model comprehension — but in human readers' eyes, that expression becomes an excessively dense pile of information.
IT之家 learned that Kernion believes the root of the problem lies in the reward mechanisms of reinforcement learning. Some reward mechanisms emphasize improving AI models' comprehension effectiveness, while others focus on making content easier for humans to understand. The more training on the math and coding dimensions, the more necessary it becomes to actively counteract this tendency by giving more rewards to concise, readable explanations.
Kernion said Anthropic found a better balance with Opus 5.5: "Since Opus 4.6, I haven't been this satisfied with a model's writing performance."
Of course, that doesn't mean Opus 5.5 has already surpassed Opus 4.6. "This is a difficult problem to handle, and I will continue to improve."

