It's interesting that OpenAI models suddenly started increasing how often they add arrows to links shortly after this behavior peaked at a lower level in Anthropic models. It would be amusing if this were because they trained a reward model on human ratings where UIs made by the latest Opus were preferred and coincidentally arrows were slightly more common, leading the reward model to treat them as an implicit indicator of quality, and henceforth rewarding that for every later model in a self-reinforcing loop. Similar to the em-dash overuse cascade.
1 comments