3 months ago
- Accretive editing is a failure mode in AI tools where they add addendums instead of correcting text when updating.
- An example involves changing support from Amazon Bedrock to LiteLLM, where AI added an addendum rather than replacing the obsolete information.
- This issue cannot be fixed by simply telling the model to write less or adjust style, as it stems from how LLMs predict text based on multiple inputs.
- The cause may be that LLMs lack human perspective on document updates, focusing on predicting likely text rather than making it true for readers.
- To address accretive editing, focus on instructing AI to replace obsolete text with accurate text, aiming for a document that reads as if written correctly from the start.