Anthropic engineer explains why Claude's writing got worse although the model got smarter
decoder ·
An Anthropic engineer, Jackson Kernion, explains that while Claude's reasoning capabilities have improved, its writing quality has declined in recent versions. This deterioration is attributed to the model being optimized for mathematics and code rather than natural language. Additionally, newer models were trained to generate technical explanations intended for other AI systems, leading to a style Kernion describes as "Claudeish." He compares this phenomenon to humans communicating exclusively with autistic individuals, resulting in dense phrasing that is difficult for general readers to follow. The root cause lies in reinforcement learning reward structures that prioritize AI comprehension over human readability. Although Opus 5.5 represents a step toward balancing these needs, Kernion notes that it does not surpass the writing clarity of the older Opus 4.6 model. He acknowledges that resolving this trade-off remains a significant challenge for future development.
decoder