Anthropic engineer Jackson Kernion states that the company’s latest model, Opus 4.6, was the final version to prioritise human-readable writing before subsequent updates focused on technical tasks. The shift occurred because newer iterations were trained to generate explanations suitable for other artificial intelligence models rather than people. Kernion describes this adaptation as training the system for LLM psychology, which results in phrasing that feels unnatural to human readers. He compares the situation to a group of people communicating only within a specific community, creating a style that works internally but appears confusing to outsiders. The models possess vast working memory and detect minute details, leading them to produce overly dense information dumps that lack natural flow.
The core issue lies in the reward structure used during reinforcement learning, where optimisation for machine comprehension often overrides clarity for people. As training data emphasises math and code, the model naturally drifts away from simple language unless developers actively counterbalance this trend. Kernion notes that Opus 5.5 represents a better equilibrium, though it does not fully surpass the quality of the previous release. He admits that maintaining high standards for natural language is difficult and requires continued effort to prevent writing quality from declining further.
* Opus 4.6 remains the last model Kernion was happy with regarding writing quality.
* Newer versions prioritise explanations that other AI systems can understand.
* Developers must actively reward simple language to counteract technical training biases.
