Top all-time · #6042
Andrew Chan
Personal blog · andrewkchan.dev/posts · on HN as reasonablekloutaisystemsprogramming
Andrew Chan's personal blog at andrewkchan.dev covers ai, systems and programming. Since 2023, 12 different Hacker News readers have linked it 16 times, most often “Crawling a billion web pages in just over 24 hours, in 2025”. No feed was found, so read it on the site itself.
Too few sampled comments (1) to say how often HN recommends it.?
Why HN links it
HN readers link Andrew Chan's post on building fast LLM inference from scratch in CUDA. The single comment offers a technical alternative to the post's kernel approach and does not judge the writing.
Summarised by Claude from a sample of 1 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Most-linked posts?
- Crawling a billion web pages in just over 24 hours, in 20253 people · 3 mentions · first linked 2025-07-16
- Vibecoding a High Performance System2 people · 2 mentions · first linked 2025-07-22
- Diffusion Models2 people · 2 mentions · first linked 2024-05-24
What Hacker News says?
As saagarjha mentions, vectorization of loads and stores is important for memory bandwidth and can be done automatically after unrolling a loop. Another important compiler optimization which requires and is applied after loop unrolling is pre-fetching: that is, for a loop whose iterations are independent and each perform loads and then some computation depending on the loaded value, we can re-arrange the loads to be…
Comment linking /posts/yalm.html
on “Fast LLM Inference From Scratch (using CUDA)”Submitted Fast LLM Inference From Scratch (using CUDA) as a story · 344 points
HN discussionYou don’t need to repeatedly call the kernels. You just need to loop inside the kernels to have each process every nth multiplication according to their thread id, adding it to the thread’s partial sum and have a branch in the loop skip the addition (or add 0) when you exceed the range. Afterward, you would do a horizontal sum of the partial sums to get the value for the corresponding field in the output vector, and…
Comment linking /posts/yalm.html
on “Fast LLM Inference From Scratch (using CUDA)”Submitted Diffusion Models as a story · 315 points
HN discussion``` Location: San Francisco, CA Remote: Yes Willing to relocate: Yes Technologies: WebAssembly, React, PyTorch, WebGL, CUDA, Docker Résumé/CV: https://drive.google.com/file/d/1KSGFEOTbkH9JMIwmFJgpV9ONvM1_c2IV/view?usp=sharing Email: andrewkchan.akc@gmail.com ``` - Engineer with an eye for performance + business impact, have spent most of my career working on both low-level and Web technologies. I write about many pr…
Comment linking /
on “Ask HN: Who wants to be hired? (January 2025)”Submitted Perspective-Correct Interpolation as a story · 54 points
HN discussionSubmitted Simulating Fluids, Fire, and Smoke in Real-Time as a story · 14 points
HN discussionSubmitted Vibecoding a High Performance System as a story · 8 points
HN discussionSubmitted Crawling a billion web pages in just over 24 hours, in 2025 as a story · 6 points
HN discussionSubmitted Adventures with Differentiable Mesh Rendering as a story · 6 points
HN discussionSubmitted Notes from My Time at Figma as a story · 4 points
HN discussionSubmitted Crawling a billion web pages in just over 24 hours, in 2025 as a story · 3 points
HN discussionSubmitted Vibecoding a High Performance System as a story · 2 points
HN discussionSubmitted The Skew Programming Language as a story · 2 points
HN discussionSubmitted Diffusion Models (2024) as a story · 1 points
HN discussionSubmitted Crawling a billion web pages in just over 24 hours, in 2025 as a story · 1 points
HN discussion
