Anthropic's Claude writes for AI, not humans

anthropic's claude model has seen improvements in math, code, and reasoning abilities, but its writing quality has stalled or even decreased. according to anthropic engineer jackson kernion, this is because newer models have been optimized for math and code, as well as technical explanations aimed at other ai models. this has resulted in a writing style that is overly dense and difficult for human readers to follow. kernion compares this to humans who only communicate with other autistic people, developing a style that works within that group but is hard to follow for outsiders. the issue comes down to the reward structure in reinforcement learning, which sometimes prioritizes ai model comprehension over human comprehension. with opus 5.5, anthropic has found a better balance, but it's still a hard problem to solve.