Hacker News Reader

> it tells me that a substitution of Bm7 for Em produces a ii-V-I in C.

That's because it's using GPT-4. Which means it just makes stuff up, and you can't rely on it to tell the truth. It's a fun tool to see possibly novel output, but you can never rely on it to be accurate or correct.

View on HN

This is true. I have come across other instances where the explanations are incorrect. The app is not ready for prime time, but fine-tuned models can improve this.

You don't need to fine tune GPT4 for it. A tiny in comparison model trained specifically on chords would do much better for creativity and the explanations could be generated without any AI.

There are lots of ways to solve this problem. Procedural, tiny model, fine-tuned large model, expert systems, etc.

Reply on news.ycombinator.com