LLMs are fundamentally inadequate to the tasks we desire to project them into. Not because they're dumb or unimpressive, but question-answering isn't collaboration. A model can remix everything humanity has ever written. It cannot act, observe, and adapt. That's where the utility we're chasing actually lives. No amount of scaling closes the gap.

LLMs won't design rockets, but it can make it so that many many more people can be a "rocket scientist" tomorrow.

With many extra backyard rocket scientists around, you're going to get new designs.