22 karma · joined January 5, 2025
I don't think there is any concrete steps in this direction. However, those ideas still live in the transcendental syntax since it is the successor of his geometry of interaction programme.
> or is there anything holding back Girard's theories from practical application
In theory, I don't think so. In practice, it's very difficult to approach his ideas. Not only because he's hard to read but also because you have to read and understand a lot of his previous works but also of other people's work to put everything into context. Once you do it, you then need enough practical knowledge to have an idea of what applications can come out of it.
> I know that Girard's Geometry of Interaction was supposed to have problems with additives, of which exact implications I do not quite comprehend, which may or may not be relevant
It's not very clear. From what I understand, he now considers he was "doing it wrong" and then the problem disappeared because he now distinguishes between "local" (asystemic/particular) and "global" (systemic/generic) mechanisms which are apparently and mistakenly mixed in logic. Full additives are global and live in a "system" (which defines contraints over a "free" computational space -- think of "complex systems") although we can define weaker "local" additives. This difference between local (he calls it first-order but it has nothing to do with FOL/predicate calculus) and global (he calls it "second-order") is mentionned in his "Logique 2.0" paper.
There might be links or transcendental syntax may subsume both geometry of interaction and ludics but nobody studied the relationship with ludics as far as I know.
My current understanding is that we should distinguish "proof-objects" (a sort of certificate) and what I call "proof-process" (a recipe, a method for building proof-objects). GoI/TS mostly speak about proof-objects. If you know about proof-nets, I don't see them as the essence of proofs behind sequent calculus proofs but as formal certificate constructed by sequent calculus proofs (seen as recipes). The link between ludics and GoI may be there but I don't know. There's nothing written about this.
Girard's ideas are indeed cryptic and may sound like a caricature of (a particularly grumpy) continental philosopher. One of my goals is to make these ideas more understandable and more down-to-earth, to show that they can be illustrated in a toy programming language. But there's a still a lot of work to be done (way more than what you could imagine just from reading Girard's papers). Although I studied Girard's last works for 4 years, I'm still far from understanding all its consequences. That's because it is related to pretty much everything in logic and computation (what is a type/formula/specification, how do you know whether an object is of some type, what is a proof, what is meaning, what is a system, what are logical rules, what all models of computation have in common, what all logical systems have in common, what is a program, what is an algorithm).
The idea is to go beyond the current proof-as-program correspondence (on which proof assistants like Coq are based). After its analysis of the notion of proof, Girard wanted a computational space in which elements would be able to "test themselves", without relying on some "hard-coded" semantics. In programming terms, it would correspond to the ability to build types from programs. There is no primitive types, not even the arrow type of functional programming. We just have Prolog-like bricks of terms which can interact with each other. They can express programs or tests for programs. You build everything with that like how you do chemical experiments. To give a probably familiar illustration, I'm often asked whether we could "do transcendental syntax" with the untyped lambda-calculus. It would be possible if for any type T you could find a finite set of lambda-terms acting as "tests" such that interaction with it ensures being of type T. The space of lambda-terms is not "large enough / flexible" for that. You need to introduce exotic objects like how you extend the set of rationals Q to the real numbers R to solve more equations. You can also think of how to tell whether a finite automaton recognizes some (infinite) rational language without relying on an external semantics. You only have input words and infinite testing by feeding it all words of the language. But yet, we are able to do it with an external semantics. So the question is how you internalize semantics in the object language.
If there's anyone who has any inspirations or insights, I'm open and interested. The right path to follow is still undefined. I just have some tools coming from my understanding of Girard's works.