This is a false analogy. LLMs do not "compile" natural language to high level code in the same way that a compiler or interpreter implements a high-level programming language in terms of machine instructions (or, for that matter, how a CPU implements machine instructions in hardware).

Programming and machine languages aim for a precise and unambiguous semantics, such that it's meaningful to talk about things like whether the semantics are actually precise or whether the compiler has a bug in failing to implement the spec.

Natural language is not just a higher level of abstraction on our existing stack. If a new model comes out, or you even run an existing model with a new seed, you can get different code out that behaves differently. This is not how compilers work.

If a program calls an API like

search_engine.get_search_results(query, length, order)

It doesn't "care" about the algorithm that produced that list of results, only that it fits the approximation of how the algorithm works as defined by the schema. There are thousands of ways the engine could have been implemented to produce the schema that returns relevance-based results from a web-crawler-sourced database.

In the same way, if I prompt an LLM "design a schema with [list of requirements] that works in [code context and API calls]", there are thousands of ways it could produce that code, but within a margin of error a high quality LLM should be able to produce the code that fits those requirements.

Of course the difference is that there is a stochastic element to LLM generated code. However it is useful to think of LLM's this way because it allows being able to leverage their probability of being correct, even if they aren't as precise as calling APIs but being explicit in how those abstractions are used.

This is a false interpretation, you've put "compile" in quotes when it doesn't appear in the parent comment and the actual phrasing used is more correct.