kennethallen

no profile record

comments

vor 20 Tagendiscuss

Sorry, this kind of LLMglish is illegible.

  That rounding has teeth.

  Here is the answer, and it’s the most useful idea in this whole post.

  That last clause is the whole game.

The constantly referring back or forward to other Parts is very annoying. Please write your essay so that I can read it from beginning to end and it makes sense!

Also, this is far from "first principles". I have a general familiarity with GPUs but I had to look up a number of terms, like "LDS".

vor 23 Tagendiscuss

Looks like Blue Prince interiors. :D

vor 26 Tagendiscuss

I read this post and your post introducing zeroserve. One of the main parts of the original pitch was "no separate config file and scripts". Now you're adding a config file separate from the scripts. Is the pitch now that it's like Caddy but runs eBPF scripts in userspace?

vor 2 Monatendiscuss

These games all released with Denuvo on Steam and DRM-free on GOG. (Some of them have subsequently removed Denuvo on Steam.)

Mad Max Middle-earth: Shadow of War Deus Ex: Mankind Divided Yakuza: Like a Dragon

vor 4 Monatendiscuss

> This isn’t parser-style backtracking

How is this different from backtracking? You're doing a depth-first search over possible interpretations. The grammar is just expressed in the type system instead of usual spec formats.

Critiques in other comments are accurate. This is a tooling nightmare, but also probably a nightmare to read. Consider an expression like

  2026 March 10 to 13

What's the binding precedence? Does this mean March 10 through March 13, or midnight to 1 PM on March 10th? I think this breaks down outside of trivial examples that are better achieved in other ways.

vor 4 Monatendiscuss

What does this offer over feeding `-v` output into an LLM?

vor 4 Monatendiscuss

Extremely. Repo is littered with one-off Python scripts, among many other indicators.

vor 4 Monatendiscuss

And in every training corpus many times over.

vor 5 Monatendiscuss

Yeah, instantly switched to reader mode.

vor 6 Monatendiscuss

The fundamental reason JSON won over XML is that JSON maps exactly to universal data structures (lists and string-keyed maps) and XML does not.

vor 6 Monatendiscuss

There are many different supersampling patterns you can use: https://en.wikipedia.org/wiki/Supersampling#Supersampling_pa...

A grid can have unwanted aliasing effects. It all depends on the kinds of images you're working with.

vor 6 Monatendiscuss

mmap is better when:

  * You want your program to crash on any I/O error because you wouldn't handle them anyway
  * You value the programming convenience of being able to treat a file on disk as if the entire thing exists in memory
  * The performance is good enough for your use. As the article showed, sequential scan performance is as good as direct I/O until the page cache fills up *from a single SSD*, and random access performance is as good as direct I/O until the page cache fills up *if you use MADV_RANDOM*. If your data doesn't fit in memory, or is across multiple storage devices, or you don't correctly advise the OS about your access patterns, mmap will probably be much slower

To be clear, normal I/O still benefits from the OS's shared page cache, where files that other processes have loaded will probably still be in memory, avoiding waiting on the storage device. But each normal I/O process incurs the space and time cost of a copy into its private memory, unlike mmap.

vor 6 Monatendiscuss

If you fail to load an mmapped page because of an I/O error, Unix-like OSes interrupt your program with SIGBUS/SIGSEGV. It might be technically possible to write a program that would handle those signals and recover, but it seems like a lot more work and complexity than just checking errno after a read system call.

vor 6 Monatendiscuss

Two big advantages:

You avoid an unnecessary copy. Normal read system call gets the data from disk hardware into the kernel page cache and then copies it into the buffer you provide in your process memory. With mmap, the page cache is mapped directly into your process memory, no copy.

All running processes share the mapped copy of the file.

There are a lot of downsides to mmap: you lose explicit error handling and fine-grained control of when exactly I/O happens. Consult the classic article on why sophisticated systems like DBMSs do not use mmap: https://db.cs.cmu.edu/mmap-cidr2022/

vor 6 Monatendiscuss

It's 2025 and you're telling people to `pip install` into system packages instead of running via `uvx`.

vor 8 Monatendiscuss

I don't understand the use case here. Is this supposed to be for enterprise to control access to internal applications via network access policies?

vor 8 Monatendiscuss

The use case for these boxes is a local NVIDIA development platform before you do your actual training run on your A100 cluster.

vor 8 Monatendiscuss

Running LLMs will be slow and training them is basically out of the question. You can get a Framework Desktop with similar bandwidth for less than a third of the price of this thing (though that isn't NVIDIA).

vor 8 Monatendiscuss

Author on Twitter a few years ago: https://x.com/tqbf/status/851466178535055362

vor 9 Monatendiscuss

Oh, joy