I think of it like speech synthesizers. First they were their own machines, then cards you plug into a computer, then once people figured out how to mash human speech together, they were, in some cases, a good 1.5 GB. Now, Siri voices, which are tons better than the concatinative models, used with the VoiceOver screen reader are a good 70 MB, Google TTS, even though it's awful and laggy with TalkBack, offline voices are a good 30 MB for a language pack, and in iOS 18, we can use our own voices as VoiceOver voices. So I think eventually we'll figure out how to run amazing AI stuff, even better than today, on our devices. And I think tons more people are working on LLM's than were ever working on TTS systems.