Andrej Karpathy's implementation, including back-propagation/training, of a gpt2-like LLM in 100 lines of elegant python
This 2024 IOCC entry by Adrian Cable "implements an LLM inference engine in an impossibly minimal quantity of maximally incomprehensible C code". It can run llama2-7b in only 1750 bytes of C. That's basically a paragraph of C code to run an LLM đ€Ż
I was able to eventually convince ChatGPT that it actually works by asking it to de-obfuscate and explain the code.
I tried it on my 8-year-old Thinkpad (X1 Carbon 5th gen) with 16GB of RAM and got surprisingly good output at about 1 token/second!
:h quickref is worth the occasional review
some nice visualization of shuffling algorithms
css-doodle looks fun. It's a javascript library that extends CSS for shader-like effects
sensible takes on using LLMs
I first encountered Steve Litt's troubleshooters.com through his vim-outliner project as a new vim user (20 years ago). I never did get much use out of vim-outliner (I've since written my own outliner mode based on vim-quicktask), but I find something fascinating about his writing. I've never read his books (which are mostly self-published self-help books for nerds?) but his DIY ethic and self-aggrandizing voice reminds me of ESR and other nerds whose writing style I admire.
See also his Linux Productivity Magazine:
On the red "Melt the ICE" caps and other anti-government craft trends
If the Minneapolis Police Department didnât kill anyone in a year of active policing, and my combat unit didnât kill anyone in over a year of war, Minnesotans â and all Americans â are right to ask why ICE and the Border Patrol have killed two people in my state in two weeks.
The creator is fully aware that the site is a mere âweak toolâ that is âdoomed to dieâ, but the bus factor of one combined with its semi-legal nature means there can be no real continuity: there will never be a legally incorporated Archive.Today Foundation to carry on his work. Itâs a testament to their persistence that theyâre managed to keep this up for over 10 years, and I for one will be buying Denis/Masha/whoever a well deserved cup of coffee.
I've always thought it was impressive that Stackoverflow, one of the most important software dev sites on the web, racked their own servers, so this makes me a little sad.
Stack Overflow no longer has any physical datacenters or offices; we are fully in the cloud and remote!
uv installs packages faster than pip by an order of magnitude. The usual explanation is âitâs written in Rust.â Thatâs true, but it doesnât explain much. Plenty of tools are written in Rust without being notably fast. The interesting question is what design decisions made the difference.
From Douglas Hofstadter's new book on ambigrams.
In Marlin you can get a TSV of the mesh with G29 T to paste in here
A good starting point. Providing a max-width on the main content is the thing I wish all otherwise unstyled pages had.
I've been having mysterious problems with inconsistent z-offset with my Qidi Q1 printer, and this reddit post diagnosed it! (Screws holding the extruder together work themselves loose, and then the hotend is able to move a bit.)
"Attention sinks" fix this in recent models
"This project has everything: a web app, a physical controller, a custom CNC build, generated gcode, tons of fabrication, 3d modeling, 3d printing, material sourcing..."