I've noticed anthropic models going overboard on code comments recently (e.g. perhaps the last week or two?). I've noticed it poignantly in my own use of opus 5.0 and opus 4.8. It seems to add 3-6 lines of comments above most method definitions, even where it's totally unnecessary (e.g. where the comment is stating the bleeding obvious).
I wonder if this is related to their announcement that they will start watermarking text? Their method adds a statistical biases to the generated text, which they state doesn’t work in certain situations (short text sequences being one of them). This could be their solution to ensure watermarks are added to more outputs.
You may think it's obvious, but what about users who cannot read code? Even better if you also have the LLM automatically write tests for this code, with comments of its own! Now whenever you want to change something, the LLM needs to update it in 4 separate places, reducing the risk of making mistakes because it has to quadruple-check its edits. Of course that means you spend a bit more on tokens, but you have a much higher chance of your PRs getting accepted because it looks like you went the extra mile.
I've noticed anthropic models going overboard on code comments recently (e.g. perhaps the last week or two?). I've noticed it poignantly in my own use of opus 5.0 and opus 4.8. It seems to add 3-6 lines of comments above most method definitions, even where it's totally unnecessary (e.g. where the comment is stating the bleeding obvious).
I wonder if this is related to their announcement that they will start watermarking text? Their method adds a statistical biases to the generated text, which they state doesn’t work in certain situations (short text sequences being one of them). This could be their solution to ensure watermarks are added to more outputs.
You may think it's obvious, but what about users who cannot read code? Even better if you also have the LLM automatically write tests for this code, with comments of its own! Now whenever you want to change something, the LLM needs to update it in 4 separate places, reducing the risk of making mistakes because it has to quadruple-check its edits. Of course that means you spend a bit more on tokens, but you have a much higher chance of your PRs getting accepted because it looks like you went the extra mile.