[sampton]

Multimodal LLM is the true solution but Apple is probably looking for something they can run on-device, at least current generation of devices.

[azeirah]

True solution to the problem stated in the article? They're talking about fine-grained control over model outputs, how would multimodality help there?