Earlier this month, Adrian Holovaty, the founder of the music learning platform Soundslice, solved a mystery that had been bothering him for several weeks. Strange images of what were clearly ChatGPT sessions kept being uploaded to the site.
After solving this problem, he realized that ChatGPT had become one of his company’s biggest hype men – but he was also lying to people about what his application could do.
Golovaty is best known as one of the creators of the open-source Django project, a popular Python web development framework (although he stepped down from project management in 2014). In 2012, he launched Soundslice, which remains “proudly bootstrapped,” as he told TechCrunch. He is now focused on his music career, both as an artist and as a founder.
Soundslice is a music learning app used by students and teachers. It is known for its video player synchronized with sheet music that prompts users on how to play the notes.
It also offers a feature called a “sheet music scanner” that allows users to upload an image of paper sheet music and, using artificial intelligence, automatically converts it into an interactive sheet music with notes.
Golovaty closely monitors the error logs of this feature to see what problems arise and where improvements need to be added, he says.
That’s where he started to see the ChatGPT sessions that were loaded.
They were creating a bunch of error logs. Instead of images of notes, they were images of words and a block of characters known as ASCII tablature. This is the basic text system used for guitar sheet music that uses a regular keyboard. (For example, your standard QWERTY keyboard doesn’t have a treble clef key.)
According to Golovaty, the volume of these ChatGPT session images was not so onerous that storing them cost his company money and reduced his app’s throughput. He was puzzled, he wrote in a blog post about the situation.
“Our scanning system was not designed to support this style of notation. So why were we being bombarded with so many screenshots of ASCII ChatGPT tabs? I was puzzled for a few weeks until I started working with ChatGPT myself.”
So he saw that ChatGPT was telling people that they could hear the music by opening a Soundslice account and uploading an image of the chat session. But they couldn’t do that. Loading these images would not turn the ASCII tab into audio notes.
He ran into a new problem. “The basic price was reputational: New Soundslice users had false expectations. They were confidently told that we were going to do things that we don’t do,” he told TechCrunch.
He and his team discussed their options: Putting a disclaimer on the site – “No, we can’t turn a ChatGPT session into audible music” – or building the feature into the scanner, even though he had never thought of supporting this unusual musical note system before.
He decided to create the feature.
“I have mixed feelings about it. I’m happy to add a tool that helps people. But I feel like we’ve been forced into this in a strange way. Do we really have to develop features in response to disinformation?” he wrote.
He also wonders if this is the first documented case where a company had to develop a feature because ChatGPT kept repeating its hallucination about it to many people.
Fellow programmers from Hacker News expressed an interesting opinion on this matter: Some of them said that this is no different from an overzealous salesman who promises the world and then forces developers to provide new features.
“I think it’s a very apt and funny comparison!” Golovaty agreed.
