Claude 变聪明，写作却变差
论证要点
- AI 模型在数学、代码和推理上进步很快，但写作质量停滞甚至下降，Anthropic 的 Opus 4.6 是最后一个被 Kernion 认可的好写作模型。
- 原因之一是模型被优化用于数学和代码，但更关键的是模型被训练成面向其他 AI 模型输出技术解释，这造成了 Claude 特有的怪异措辞。
- Kernion 称模型已被「适应 LLM 心理」，学会为 AI 模型而非人类写作，他把这比作只与其他自闭症者交流的人，形成群体内有效但外人难懂的风格。
- LLM 的工作记忆远超人类，能捕捉更细粒度的细节，训练中因此涌现出一种对 AI 模型有效、但人类读起来像「过度密集信息倾倒」的写作风格。
- 问题根源在强化学习的奖励结构：部分奖励优化 AI 模型理解，部分优化人类理解；越训练数学和代码，就越需要主动奖励人类能读懂的简单解释来抵消。
- Kernion 称 Opus 5.5 找到了更好的平衡，但他自 Opus 4.6 以来没对模型写作这么满意过，并强调这不代表 Opus 5.5 超越旧模型，问题仍难解决。
对实践的启示
做 AI 应用的人要注意：模型能力提升不等于输出更适合人读，面向代码和数学的优化会牺牲自然语言质量。若产品面向人类读者，需在提示或微调中主动奖励简洁易懂的表达，不能默认新模型写作更好。
作者承认的反方与局限
Kernion 承认 Opus 5.5 找到了更好的平衡，但表示这不意味着它超越 Opus 4.6，并称这是个难解的问题，团队会继续改进。
原文摘引
The model has been "adapted to LLM psychology," Kernion says, and learned to write for AI models, not for people.
The more you train on math and code, the more you have to actively push back by rewarding simple explanations that human readers can follow.
It's a hard problem to solve, and we'll continue to make improvements.
▸ 展开原文全文（10 段，抓取自 The Decoder）
Anthropic engineer explains why Claude's writing got worse although the model got smarter
AI models are improving fast at math, code, and reasoning. Their writing quality, though, has stalled or even gotten worse.
Anthropic employee Jackson Kernion, who works on Claude fine-tuning, recently explained why Opus 4.6 was the last good "writing model" from Anthropic. The answer is partly predictable: newer models have been optimized for math and code.
But they've also been trained to produce technical explanations aimed at other AI models, and that's what really drives Claude's odd phrasing ("Claudeish"). The model has been "adapted to LLM psychology," Kernion says, and learned to write for AI models, not for people.
Kernion compares this to humans who only communicate with other autistic people. They develop a style that works within that group but feels hard to follow for outsiders, he says. LLMs have far more working memory than humans and pick up on details at a much finer level, and so, during training, a writing style emerges that works well for AI models but produces what human readers experience as "overly-dense info dumps."
Optimizing for code comes at a cost to natural language
The issue, according to Kernion, comes down to the reward structure in reinforcement learning. Some rewards optimize for AI model comprehension, others for human comprehension. The more you train on math and code, the more you have to actively push back by rewarding simple explanations that human readers can follow.
With Opus 5.5, Anthropic found a better balance, Kernion says, adding that he hasn't been "as happy about a model's writing since Opus 4.6." That doesn't mean Opus 5.5 surpasses the older model, though. "It's a hard problem to solve, and we'll continue to make improvements," Kernion writes.
AI News Without the Hype – Curated by Humans
Subscribe now