Excessive confusion and stupidity #50
Description
Backend impacted
The PyTorch implementation
Operating system
Linux
Hardware
GPU with CUDA
Description
Thank you so much for this work. I think you have improved over the original open source Moshi demo and this is a big step towards something useful without specific fine-tuning.
However, and I'm not sure if it's because I am doing an outgoing rather than incoming call scenario, but using the demo client/server: the model seems very easily confused, bad at following instructions, and generally stupid.
It is getting the important names confused with other people who might be involved, such as the person that answered, and cannot adjust to the idea that that it made an outgoing call. When I had an example name in a call script in the instructions, it mixed that up with the real name that I clearly indicated.
I am hoping someone will build something similar but a bit larger, maybe 24B or so.
Extra information
See above.
Environment
Fill in the following information on your system.
- Operating system version: Ubuntu 24