An Engineer Links Claude's Decline in Readability to Training

A Claude tuning specialist at Anthropic links the perceived degradation of style to recent training choices. He mentions a trade-off favoring mathematics and code, along with explanations tailored for AI models. Opus 5.5 aims for a better balance, though it is not guaranteed to surpass the writing performance of Opus 4.6.

Opus 5.5 Seeks Balance Without Dethroning Opus 4.6

Jackson Kernion claims that with Opus 5.5, Anthropic has found a better balance. He adds that he has not been as satisfied with the writing of a model since Opus 4.6. However, he clarifies that Opus 5.5 does not necessarily outperform the older model. Previously, he had presented Opus 4.6 as the last good writing model from Anthropic.

Explanations Tailored for LLMs, Not for Humans

Kernion specifies that the models have been trained to generate technical explanations intended for other AI systems, which leads to formulations described as "Claudeish." He explains that the model has been adjusted to align with the thinking of LLMs and has learned to write for AIs rather than for people. According to him, LLMs have a much larger working memory and grasp details at a finer level. In training, this results in a style that is effective for AI models but perceived by readers as overly dense information dumps. He compares this situation to a group of people who would only communicate among themselves, developing a style that is difficult for outsiders to follow.

Optimization for Code Requires Language Compensation

Kernion identifies the core of the problem in the reward structure of reinforcement learning, where certain rewards favor understanding by models and others by humans. The more training targets mathematics and code, the more there needs to be active compensation by rewarding simple explanations accessible to human readers. He links the degradation of written style to this optimization towards technical tasks. Meanwhile, performance is rapidly advancing in mathematics, coding, and reasoning, while the quality of writing has stagnated or worsened.