They don't only require "good enough to kill industries" (which is doubtful but certainly feasible), that's just step one. I think about it in terms of potential failure modes:
- if models don't reach worker-substitution levels, they fail
- if models reach that level, but it's too expensive to run and a worker's still cheaper, they fail
- if models reach that level, but the resulting tech is cheap enough to use, they fail (since open models can compete)
- If the models work but there's social rejection leading to regulation (due to mass unemployment for example), they fail
- if the models work but there are significant deal breakers (like a fundamental inability to keep them safeish to use) they fail.
So it's not really a single AI killer reason, it's more that the success case requires things to land in a very specific future where models work, and they're cheap enough, and expensive enough, and valuable enough, and exclusive enough, and safe enough, and...
Each "and" is a multiplier reducing their chances, and there's a ton or factors. Not imposible, but not where I'd put my money.