So to make this a git repo, I guess you'd make the snapshots be the law plus markup for all the changes, and the diffs are the metachanges. For each snapshot S, you could generate a git repo whose snapshots are the law and the whose diffs are the changes in the law (as described by S). However, this probably isn't the best way to query S.

  fun f(x1, x2, x3, ...):
    t1 = 3*x1 + 2*x2
    t2 = 4*t1 + 5*x3
    ...
    return (some subset of t vars)

The cost model is just the number of additions performed.

So from the browser's perspective, these are two distinct, unrelated URLs. There isn't a mechanism where it can know they are backed by the same file.

This browser behavior seems pretty reasonable. I'd try fixing it by doing something different on the server.

The way I have to resume downloads in Firefox is

1. Move the .part file to a temp file

2. Start the download again; a new .part file gets created

3. Pause the download

4. Move the temp file to the .part file, overwriting it

5. Unpause the download; it resumes from where the old .part file stopped

This is only necessary if a "one-time" download link gets interrupted part-way through, eg. by a flaky connection. Attempting to resume a one-time link like normal gets (I assume) a 400 response and FF kills the .part file.

I think the wider view is usually what people mean when they talk about nondeterminism in LLMs. Maybe because computers are traditionally so deterministic the wide view is almost taken for granted by programmers.

Why?