I think this points to bigger issues with the AI field generally with respect to adoption and development. The output of these systems is fundamentally unverifiable. Building a system to verify these answers would be orders of magnitude more expensive and maybe computationally impossible. It looks really impressive because it's 95% of the way there, but 5% error rate is atrocious, in was expensive to get this far, and improving it much more will be more and more expensive. What we've essentially built is a search engine that is 5 times more expensive to operate and that alienates the content from its context, which makes it less valuable.

Maybe it's good enough for programming since you can immediately verify the output but I suspect we're pretty far away from the breakthroughs ai boosters insist we are mere months away from. We've had some form of self driving for a while now and the other company that seems close is waymo, and that seems to have taken over a decade of huge research and capital expenditures.