Nope, it isn't. There are thousands of "legacy" software products that cannot be converted into machine code while complying with runtime constraints, compile-time assertions or unit tests. Many of these programs are baked into a ROM and will continue running regardless of how optimal AI gets.

Fun theory, but it's more of a Rorschach test for how people perceive tech being deployed.

There are many other so-called models of computation that are useful for representing ideas such as actor models, abstract rewriting systems, decision trees, and so on. Without them, you might feel that something is missing, so relying on assembly alone would not be enough.

I see that we have progressed from just needing a "sufficiently advanced compiler" to needing a "sufficiently advanced LLM".

The sufficiently advanced compiler never showed up, and it was easier than a sufficiently advanced LLM. So don't hold your breath for this future you envision.

It's already true. Nothing runs but machine code. We will still have other abstractions in between, written by both LLMs and humans because they're easier to reason about and work with than assembly, both for text-modeling LLMs and for people who have to work with the software.

Is there enough assembly out there to train the LLMs?

Uh, assembly is a middleman.