But that approach doesn't work with code, or with reasoning in general, because you would need to exponentially fine tune everything in the universe. The illusion that the AI "understands" what it is doing is lost.

Code generation progression in LLMs still carries higher objective risk of failure depending on the experience on the person using it because:

1. They still do not trust if the code works (even if it has tests) thus, needs thorough human supervision and still requires on-going maintainance.

2. Hence (2) it can cost you more money than the tokens you spent building it in the first place when it goes horribly wrong in production.

Image generation progression comes with close to no operational impact, and has far less human supervision and can be safely done with none.

LLMs aren't a "layer of abstraction."

99% of people writing in assembly don't have to drop down into manual cobbling of machine code. People who write in C rarely drop into assembly. Java developers typically treat the JVM as "the computer." In the OSI network stack, developers writing at level 7 (application layer) almost never drop to level 5 (session layer), and virtually no one even bothers to understand the magic at layers 1 & 2. These all represent successful, effective abstractions for developers.

In contrast, unless you believe 99% of "software development" is about to be replaced with "vibe coding", it's off the mark to describe LLMs as a new layer of abstraction.