to tell if a maze is solvable, just hang it by its corners! The first maze stays in one piece, so there is no path from the entrance at the top to the exit at the bottom. The second maze splits apart along the solution.

Clearly LLMs must one day run in Space Step 1 we harden llm.c to pass the NASA code standards and style guides, certifying that the code is super safe, safe enough to run in Space. en.wikipedia.org/wiki/The_Power… (see the linked PDF) LLM training/inference in principle should be super

RELEASE DAY After almost 10 years of hard work, tireless research, and a dive deep into the kernels of computer science, I finally realized a dream: running a high-level language on GPUs. And I'm giving it to the world! Bend compiles modern programming features, including: -

I just wrote a change making GeLU 2x to 8x faster on CPU in llama.cpp and llamafile, with 8192x better accuracy. It's the world's most popular activation function. Now it has clean self-contained SSE2, AVX2, AVX512, ARM NEON impls you can copy and paste. github.com/ggerganov/llam…

System design is about trade-offs: you can achieve incredible performance, but at what cost? While building the research serverless system Apiary, I worked with H-Store/VoltDB, which is unbelievably fast (microsecond-level transactions, several orders of magnitude faster than

Video of my talk at LPC 2024: Crafting a Linux kernel scheduler that runs in user-space using Rust youtu.be/UK6XX27mK3c?si…

We needed a vector database that was inherently “Cloudflare”: deployed globally, distributed, and intrinsically multi-tenant. So we built one: blog.cloudflare.com/building-vecto…

AI Flame Graphs: Showing what's actually running on the HW and how we got there. Uses Intel EU stall profiling and eBPF. brendangregg.com/blog/2024-10-2…

NVIDIA's $7B Mellanox acquisition was actually one of tech's most strategic deals ever. The untold story of the most important company in AI that most people haven't heard of 1/12

I think programming GPUs is too hard. Part of the problem is sprawling, scattered documentation & best practices. Over the past few months, we’ve been working to solve that problem, putting together a “Rosetta Stone” GPU Glossary. And now it’s live! My take-aways in thread.

Distributed training is notoriously hard to learn - knowledge is scattered across papers and complex codebases. Enter picotron: implementing all 4D parallelism concepts in separate, readable files totaling just 1988 LoC!

I wrote up all my thoughts on whether NVDA stock is finally a decent short (or at least not a good thing to own at this point). I’m a huge bull when it comes to the power and potential of AI, but there are just too many forces arrayed against them to sustain supernormal profits.

The spell article reached Douglas. I saw him subscribe but thought it might be a prank by someone, until I saw this email from him. Very grateful that he took the time to read and respond to it. PS: I fixed the mistake he pointed out.

“Entropy is geometry. And geometry is entropy.” This is a new finding by Gabriele Carcassi, and I'll explain the reasoning below, along with the math. Don't worry, I'll hold your hand (metaphorically, of course, unless you're into that). (1/19)

Since Chain Replication with Apportioned Queries (CRAQ) is in the news because of DeepSeek 3FS, it's a good time to revisit the MIT 6.824 FAQ on the paper.

I taught an LLM to optimize proteins. It proposed a better carbon capture enzyme. Introducing Pro-1, an 8b param reasoning model trained using GRPO towards a physics based reward function for protein stability. It takes in a protein sequence + text description + previous

HOW TO DESIGN A FACTORY Lots of talk these days about bringing back manufacturing to America. In light of that trend I've decided to give you a how-to-guide on designing a factory

DESI has made the largest 3D map of our Universe to date Earth is at the center in this animation, each blue point is a galaxy: > 14 million galaxies in total up to 11 billion light years away The gaps are the regions where our own galaxy blocks the view

Say hello to one of our Milky Way’s neighbors! 👋 This new #HubbleFriday image peers into the Small Magellanic Cloud, a dwarf galaxy located about 200,000 light-years away. Read more: go.nasa.gov/4hv9RbX

Water can solve differential equations. Lukyanov, a Soviet engineer, was trying to calculate heat transfer in concrete structures. Hand calculation was cumbersome, so he developed an analog computer to physically model the math relationships. It worked *really* well.