What are your thoughts on a future where code is represented as a structured model, rather than text? Do you think that AI-powered coding assistants benefit from that?

It’s important to realize that “modern “AI” doesn’t understand human level meanings any better today (in many cases: worse!). So it is not going to be able to serve as much of a helper in a general coding assistant.

i.e. “operator like things” at the nodes …

a b i 1 x []

Chicago Bears offense has 94-yard TD drive in 2 minute drill at minicamp - Windy City Gridiron

NOT DATA … BUT MEANING!

A slogan that might help you get past the current fads is:

Most coding assistants — with or without “modern “AI” — also do reasoning and manipulation of structures.

2025 Stanley Cup Finals Preview: Can Connor McDavid Finally Do It? - The Ringer

Long ago in the 50s this was even thought of as a kind of “AI” and this association persisted into the 60s. Several Turing Awards were given for progress on this kind of “machine reasoning”.

Another canonical form could be Lisp S-expressions, etc.

in structures, such as:

2025 NASCAR Xfinity Series entry list for Autódromo Hermanos Rodríguez - NASCAR.com

These structures are made precisely to allow programs to “reason” about some parts of lower level meaning, and in many cases to rearrange the structure to preserve meaning but to make the eventual code that is generated more efficient.

First, it’s worth noting that the “syntax recognition” phase of most compilers already does build a “structured model”, often in what used to be called a “canonical form” (an example of this might be a “pseudo-function tree” where every elementary process description is put into the same form — so both “a + b” and “for i := 1 to x do […]” are rendered as

plus(a, b) for(i, 1, x, […])

NHL Draft: Flyers take strong winger Porter Martone at sixth overall - PhillyVoice

/ \ and ⁄ / | \

+ for