> Assembly experts still write code that runs faster than code produced by compilers.

They sometimes can, but this is no longer a guaranteed outcome. Supercompilation optimizers can often put manual assembly to shame.

> Impact of having 1.7x more bugs is difficult to assess and is not solved that easily.

Time will tell. Arguably the number of bugs produced by AI 2 years ago was much higher than 1.7x. In 2 more years it might only be 1.2x bugs. In 4 years time it might be barely measurable. The trend over the next couple of years will judge whether this is a viable way forward.