I have no idea why Google continues to publish so much stuff that seems like it could help competitors, but I enjoyed reading this guide to scaling LLM training and inference.
jax-ml.github.io/scaling-book/
jax-ml.github.io/scaling-book/
How To Scale Your Model
Training LLMs often feels like alchemy, but understanding and optimizing the performance of your models doesn't have to. This book aims to demystify the science of scaling language models on TPUs: how...
jax-ml.github.io
April 10, 2025 at 12:23 AM
I have no idea why Google continues to publish so much stuff that seems like it could help competitors, but I enjoyed reading this guide to scaling LLM training and inference.
jax-ml.github.io/scaling-book/
jax-ml.github.io/scaling-book/
Disappointed in Claude in Cursor this week. It just sort of brute forces things to get rid of linter errors rather than stop and reconsider the approach, despite haveng "thinking" enabled. Still maybe having something written that was obviously wrong saved me time vs starting from scratch.
March 15, 2025 at 4:23 AM
Disappointed in Claude in Cursor this week. It just sort of brute forces things to get rid of linter errors rather than stop and reconsider the approach, despite haveng "thinking" enabled. Still maybe having something written that was obviously wrong saved me time vs starting from scratch.
I just chipped in to get a food that I eat almost daily tested for harmful chemicals. Pretty neat. laboratory.love
laboratory.love
Crowdfund chemical testing on everyday items
laboratory.love
March 9, 2025 at 4:56 AM
I just chipped in to get a food that I eat almost daily tested for harmful chemicals. Pretty neat. laboratory.love
My most significant open source contribution to date. Earl: a framework for scalable reinforcement learning research
Lets you do RL the way DeepMind does :-) Blog post about it:
www.garymm.org/blog/2025/03...
Lets you do RL the way DeepMind does :-) Blog post about it:
www.garymm.org/blog/2025/03...
Earl: a framework for scalable reinforcement learning research
www.garymm.org
March 4, 2025 at 12:42 AM
My most significant open source contribution to date. Earl: a framework for scalable reinforcement learning research
Lets you do RL the way DeepMind does :-) Blog post about it:
www.garymm.org/blog/2025/03...
Lets you do RL the way DeepMind does :-) Blog post about it:
www.garymm.org/blog/2025/03...
Nvidia Nsight Systems (performance profiler) takes like 3 minutes to start up on an M2 Pro. Oh the irony.
March 3, 2025 at 7:39 PM
Nvidia Nsight Systems (performance profiler) takes like 3 minutes to start up on an M2 Pro. Oh the irony.
This surprised me:
(computer science) seniors in the United States substantially outperform seniors in China, India, and Russia by 0.76–0.88 SDs and score comparably with seniors in elite institutions in these countries.
www.pnas.org/doi/10.1073/...
(computer science) seniors in the United States substantially outperform seniors in China, India, and Russia by 0.76–0.88 SDs and score comparably with seniors in elite institutions in these countries.
www.pnas.org/doi/10.1073/...
Computer science skills across China, India, Russia, and the United States | PNAS
We assess and compare computer science skills among final-year computer science undergraduates
(seniors) in four major economic and political power...
www.pnas.org
February 23, 2025 at 11:43 PM
This surprised me:
(computer science) seniors in the United States substantially outperform seniors in China, India, and Russia by 0.76–0.88 SDs and score comparably with seniors in elite institutions in these countries.
www.pnas.org/doi/10.1073/...
(computer science) seniors in the United States substantially outperform seniors in China, India, and Russia by 0.76–0.88 SDs and score comparably with seniors in elite institutions in these countries.
www.pnas.org/doi/10.1073/...
Tried o3 mini in cursor and it just stopped generating output without warning multiple times. I'm back to Claude for a few days at least. Anyone had a better experience?
February 4, 2025 at 6:39 AM
Tried o3 mini in cursor and it just stopped generating output without warning multiple times. I'm back to Claude for a few days at least. Anyone had a better experience?
Summary of my C++23 implementation of Deflate decompression (core of Gzip, Zip, PNG)
www.garymm.org/blog/2025/01...
www.garymm.org/blog/2025/01...
Starflate: Deflate decompression in C++23
www.garymm.org
February 1, 2025 at 5:45 AM
Summary of my C++23 implementation of Deflate decompression (core of Gzip, Zip, PNG)
www.garymm.org/blog/2025/01...
www.garymm.org/blog/2025/01...
While working on machine learning research at the Astera Institute1, I led a team that assembled a system that enabled researchers to quickly and easily run experiments that used up to a full datacenter’s worth of GPUs.
Blog post with details:
www.garymm.org/blog/2025/01...
Blog post with details:
www.garymm.org/blog/2025/01...
Assembling an infrastructure for machine learning research
www.garymm.org
January 28, 2025 at 6:04 PM
While working on machine learning research at the Astera Institute1, I led a team that assembled a system that enabled researchers to quickly and easily run experiments that used up to a full datacenter’s worth of GPUs.
Blog post with details:
www.garymm.org/blog/2025/01...
Blog post with details:
www.garymm.org/blog/2025/01...
"As strongly requested by the reviewers, here we cite some references [[35], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], [47]] although they are completely irrelevant to the present work."
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
Origin of the distinct site occupations of H atom in hcp Ti and Zr/Hf
The location of the H atoms in Ti, Zr, and Hf is crucial to the formation of the hydrides in these metals as it influences the crystal lattice transfo…
www.sciencedirect.com
January 21, 2025 at 4:09 AM
"As strongly requested by the reviewers, here we cite some references [[35], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], [47]] although they are completely irrelevant to the present work."
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
Another example of California dysfunction.
Container port efficiency rankings (out of 405 ports):
Long Beach: 373
Los Angeles: 375
Oakland: 397
I've had something I ordered stuck in the port of Los Angeles for 2 weeks.
Source: www.iaphworldports.org/news/iaphnew...
Container port efficiency rankings (out of 405 ports):
Long Beach: 373
Los Angeles: 375
Oakland: 397
I've had something I ordered stuck in the port of Los Angeles for 2 weeks.
Source: www.iaphworldports.org/news/iaphnew...
Latest container port efficiency ratings revealed | IAPH
The latest edition of the Container Port Performance Index (CPPI) 2023 – a comprehensive report that ranks 405 global container ports based on efficiency – was released earlier this month. A collabora...
www.iaphworldports.org
December 20, 2024 at 4:47 AM
Another example of California dysfunction.
Container port efficiency rankings (out of 405 ports):
Long Beach: 373
Los Angeles: 375
Oakland: 397
I've had something I ordered stuck in the port of Los Angeles for 2 weeks.
Source: www.iaphworldports.org/news/iaphnew...
Container port efficiency rankings (out of 405 ports):
Long Beach: 373
Los Angeles: 375
Oakland: 397
I've had something I ordered stuck in the port of Los Angeles for 2 weeks.
Source: www.iaphworldports.org/news/iaphnew...