In Taiwan, we fund more than 150 different vertical model trainings. So, instead of training one single supermodel that sees everything, solves everything in a kind of adequate but not very good way, we focus on those narrower domain-specific models precisely because that gives the users, the people, easier steering, easier control over individual model alignment and tuning. And also it’s much more energy efficient because it does not need to memorize all the Studio Ghibli clips or something like that. And so for each task, like summarization, translation, facilitation, and so on, it is both faster, more energy efficient, and much more tunable. So, I think the future is public, pluralistic and portable AI models.

Keyboard shortcuts

j previous speech k next speech