The significance of being able to run medium-sized models locally is underestimated.

It’s not that their capabilities have caught up with closed-source models. It’s that they change the premise of whether you can use them at all—when privacy is involved, when you’re offline, or when data can’t be uploaded. Before, these situations weren’t even an option.

My take is that the next wave of truly robust AI applications will emerge in situations where “the data can’t leave.”

That’s not a technology problem. It’s a trust problem.