My own guess is not that AI will by itself invent new coding languages, but
rather I believe that AI-aided programming (AAP) will engender new techniques
and higher levels—”programmers” will dialog with specially prepared AI about a
given solution or modification to a solution, and the AI will present check-
lists (or something to that effect) that the "programmer" can rapidly “go through”, and then the AI would use that detail as its input to do the actual coding that would be seen and consumed by the target computing machinery. This AAP-based coding would probably be done in new notations (languages), that need not value simplicity and
“expressiveness to humans” as part of their design desiderata.
For example, languages could be designed such that extremely complex / complicated expressions (amounting to entire subroutines) could be specified in a single line of code. Such hugely compact expressions can be thought of as like mathematical proofs which take other trained mathematicians days or even weeks to go through and validate— AI could be devised to code at such levels, using new, complex notations, and “rapidly-acquired-but-detailed” input (such as rapid-fire checklists, presented by AI's after dialog about a solution or modification in software), with human experts in working with AI for AAP. These "programmer SME's" would necessarily be familiar with the target computing machinery, and would work alongside the AI's to debug the AAP code.
For a conceptual "glimpse" illustrating what such compact AAP code might look like, the reader is invited to consider APL, a quirky programming language consisting of a high number of "graphical" characters whose central data type is the multi-dimensional array. APL code does not look like ordinary computing code/text-- its meaningful character set includes "exotic" single-character codes, not part of standard 7-bit ASCII, which allows only 127 characters.
Human beings would only in extremely rare instances have to “unpack” (analyze)
AAP code un-aided. When inevitable problems occurred, there would be “back-
translation” utilities to express the “AAP” code in a logically equivalent,
non-compact way, for humans to validate and even debug. However, with correctly trained software-expert AI's, the entire process could be (as I see it) enough more productive than traditional, un-aided human programming, to be a clear advantage to those companies that would invest in (or rent) software-AI "aides" for AAP, over traditional coding techniques.