busfahrer

has anyone been able to get bonsai models to run at a proper speed in a x86 windows cpu-only setting, using llama.cpp?

For anybody else trying to grok backprop, I found it very helpful to look at karpathy's microgpt, a barebones LLM in 200 lines of python. The autograd itself is just 40 lines, and reading that code and then trying to do a clean room implementation of it is a great exercise:

https://gist.github.com/karpathy/8627fe009c40f57531cb1836010...

https://karpathy.github.io/2026/02/12/microgpt/

As a kid I played Ken's Labyrinth, a raycaster FPS from around the same time as Wolf3D. Only years later I found out that this was actually the same Ken Silverman that went on to create the Build engine for Duke Nukem 3D etc. Kind of funny considering the vast difference in tone between these two games.

edit:

https://en.wikipedia.org/wiki/Ken%27s_Labyrinth

I'm on the fence because I have a TSC-X controller and it's unclear if it's supported. Somebody on the forums posted a tool that converts generic joystick axes to keypresses, but not sure how well that works.

Or maybe Lisp is so pure that it embodies both Light Side and Dark Side, like a god that spawned the programming universe.

This isn't so far off, considering that some people consider the "Lisp in Lisp" bit from the 1.5 manual to be the "Maxwell's Equations in Software":

https://michaelnielsen.org/ddi/lisp-as-the-maxwells-equation...

The PS1 architecture is fascinating - it will also reveal why PS1 games have such a unique, recognizable style that games are trying to recreate nowadays.

To add some context, the author of this article is a science fiction writer, one of his stories was adapted into the movie "Arrival".

Not a one-liner, but I was impressed by this simple rogue-like written in 10 lines of BASIC:

https://bunsen.itch.io/the-snake-temple-by-rax

Having only dabbled in Lisp and knowing nothing about the plugins TFA mentions, there's also this FWIW:

https://github.com/jpalardy/vim-slime

Thank you, good sir!

Great, thanks! :-) and to mirror another poster: what kind of prompt parsing (prefill) speed do you get for that model? Also how is the speed for the 27B model?

I have been contemplating a M5 Pro MBP, but for the life for me I wasn't able to find benchmarks for real-world models, do you happen to know how many tokens per second roughly you get with MoE models like Qwen 3.6 35B/A3B or Gemma 4 26B?

Seeing as this is HN, I was expecting something on chemical properties of iron etc, but was pleasantly surprised

OTOH, with the IMSAI-8080 you'd have the satisfaction of using the same machine as David Lightman of Wargames fame

To add some context, this was made for the PICO-8 fantasy console, which (by design) is very limited, so it's immensely impressive to see something like this working on it. Honestly, even just the 3D logo from the intro is impressive, and the rest of the game even moreso.

https://www.lexaloffle.com/pico-8.php

So I'm guessing it's sort of like Fennel, which consciously doesn't completely hide its underlying Lua runtime?

https://fennel-lang.org/

I am considering a M5 Pro (18/20C) Macbook with 64GB of RAM, but I'm having a really hard time finding benchmarks of real world models:

Could somebody please provide some tokens-per-second numbers for example for Qwen 3.6 35B/A3B, specifically for Q4 and Q6 quants?

This is the whole point of Karpathy's nanochat which OP refers to, to train a GPT-2 level LLM for under $100, renting an 8xH100 VM.

This kind of is Postel's law, in a way:

https://en.wikipedia.org/wiki/Robustness_principle

Out of curiosity, what size Qwen did you use, at what quantization?

I think I remember a Carmack tweet where he mentioned in most cases he only considers it once he reaches three duplicates

This seems like a nice context to mention Sturgeon's law:

ninety percent of everything is crud

https://en.wikipedia.org/wiki/Sturgeon%27s_law

Low cost specialized models

Can you elaborate on this? Is this something that companies would train themselves?

I've noticed that subconsciously, I've started to use `bat` for interactive paging, even though that's not its main use case

Once you get your head around it a bit, doing a new in this circumstance will be second nature, since you will have realized that a `new XYZ` in jj leads to the same underlying git state as a `git checkout XYZ` in git

Don't think about git.

You can also make the reverse point, only think about git:

In theory it's perfectly possible if you have to jj binary lying around somewhere, to not even use it as your git client, but to temporarily do a `jj git init --colocate`, do a complex restructuring of your repo across 3 branches, and then delete the .jj folder again :-)

Just to highlight the other end of the spectrum, jj is flexible like that

I wonder, I wonder where one could find a good book about the software architecture of that game… oh, well

Even more impressively than fitting on a single floppy disk, Elite was built to fit into 56 kb of RAM

It's a pity that it's no longer available on iOS.

I do wonder how realistic it would be to maybe crowdfund the dev team with 3 weeks of Claude Code in order to achieve a working WASM port?

which is like a typical IKEA, except its maze of twisty passages run to infinity.

For the 4 people on HN who don't know, "maze of twisty passages" is a reference to the this (the?) text adventure game:

https://en.wikipedia.org/wiki/Colossal_Cave_Adventure