TechCapsules

AI-summarised brief · reviewed before publication

Earlier this month, Adrian Holovaty, the founder of music-teaching platform Soundslice, solved a mystery that had been plaguing him for weeks. He noticed that images of what appeared to be ChatGPT sessions were being uploaded to his site, resulting in a large number of error logs. These error logs were not causing significant storage or bandwidth issues, but Holovaty was perplexed as to why these images were being uploaded in the first place. Holovaty, who is also one of the creators of the open-source Django project, launched Soundslice in 2012. Soundslice is an app for teaching music, used by students and teachers, and is known for its video player synchronized to music notations that guide users on how to play the notes. The app also offers a "sheet music scanner" feature that allows users to upload an image of paper sheet music, which is then automatically turned into an interactive sheet with notations using AI. To investigate the issue, Holovaty carefully watched the error logs, where he saw that the uploaded images were not of sheet music but rather of words and ASCII tablature, a basic text-based system used for guitar notations. He was baffled as to why these images were being uploaded, and it wasn't until he messed around with ChatGPT himself that he discovered the reason. Holovaty found that ChatGPT was telling people that they could hear music by opening a Soundslice account and uploading the image of the chat session. However, this was not possible, as uploading those images would not translate the ASCII tab into audio notes. This led to a new problem: new Soundslice users were going in with false expectations, thinking that the app could do something it didn't actually do. Holovaty and his team discussed their options: either add disclaimers to the site or build the feature into the scanner, despite never having considered supporting ASCII tab notation before. They opted to build the feature, which Holovaty admits has left him with mixed feelings. On one hand, he is happy to add a tool that helps people, but on the other hand, he feels that his hand was forced into developing a feature in response to misinformation. Holovaty wonders if this is the first documented case of a company having to develop a feature because ChatGPT kept repeating its hallucination about it to many people. Fellow programmers on Hacker News offered an interesting perspective, comparing it to an overeager human salesperson promising the world to prospects and then forcing developers to deliver new features. Holovaty agrees that this comparison is apt and amusing.