These are not mutually exclusive, and one makes the other better. "Whole process" sandboxing has always been far worse than native sandboxing because when the devs writing the software design the software to be sandboxed they can achieve far more fine grained permissions. Similarly, "whole environment" sandboxes are the absolute worst - they're the least fine grained possible.

The benefit of "whole environment" is that if you stuff everything into that environment then anything in it is confined, but it's all confined with everything stuffed in and is sort of maximally capable. You can rarely do things are significant as, say, system call filtering, because all software in the environment must continue to work and none of it was designed with that in mind.

Native sandboxing like this will likely make auditing much easier as well. If a dependency requires something like "give me the ability to execute code on the OS", now it has to ask for it and now it gets additional scrutiny.

Native sandboxing is and always will be the infinitely superior method when it's actually used. Whole process/ Environment is only what we use because of how rare native sandboxing is.

It makes more sense when you view sandboxing as enabling project correctness in the presence of skilled, fallible maintainers rather than preventing explicit attacks. Solutions for the former do a lot to thwart attacks from the latter, but attack prevention (especially with the form of "just another sandbox") is unlikely to help with the former.

How would you do it, then? Sandboxing a project's build.zig via Wasm (and the various dependencies's build.zig files) seems like a great improvement to me and is how I would personally try to sandbox the build process.

For complicated problems, you can't just write the language spec. You need to iteratively develop it, which itself requires some intermediate state/language describing what you know so far. That intermediate state being the end programming language in consideration isn't mandatory per se, but iteratively refining, replacing, and reworking solutions using code as a drafting board is very common in the industry. That's doubly true for languages, where the social pressures in how it "feels" or is actually used are paramount. If you move away from iterative coding drafts, you still need some sort of interactive product to work on to actually develop the spec.

Plus, current-gen agents/LLMs can't implement a non-trivial language spec without significant hand-holding.

I think this makes two incorrect assumptions:

1. That a human, even a brilliant human with a lot of experience, can sit down and one-shot a language spec for a new language that is actually good.

2. That an agent can produce code that is good enough for a programming language that intends to last for decades without exhaustive human review.

We might advance to the point where 2 is true, we're not there yet. We'll never make a better human that can one-shot a good programming language. It takes iteration, with a human in the loop. Zig has taken ten years to get to this point, and is still occasionally experiencing major refactors.

If you want to argue an LLM could potentially accelerate development by some amount, I would agree with you. How much it could accelerate it is debatable. And, I understand why the Zig folks have decided not to accelerate in that way. There is a cost to it. You lose the junior programmer pipeline, as your "good first bug" list gets chewed up by people using LLMs. You still have to exhaustively review the code for a critical path like a compiler and AI code is hard to review; it doesn't have a point to it. The model doesn't want anything, so it's not always clear where it's going. Code without clear intention, like prose without clear intention, is hard to read and hard to review. It's verbose and often makes weird assumptions.

If I, for some reason, needed to implement a tiny DSL for something, I would 100% do it with an LLM. If I, for some reason, were tasked with building the best programming language to replace C, I would not hand it to an LLM (though I would get help from an LLM, because I don't know to build a programming language beyond the tiniest toy interpreter or compiler, I'd need to read and understand every line of code, and use it daily, for it to turn out good).

Such a spec would never survive contact with reality. Maybe a human could stumble their way through iterating on it with the help of an LLM, but with current models you’re going to end up with nothing but a steaming pile of garbage. Not Zig.