Writing an LLM from scratch, part 13 – attention heads are dumb
A pause to take stock: realising that attention heads are simpler than I thought explained why we do the calculations we do.
The latest blog posts to take off on Hacker News. 4,109 posts.
A pause to take stock: realising that attention heads are simpler than I thought explained why we do the calculations we do.
Among the most impressive manufacturing achievements of the US during WWII was the number of ships it produced.
I think I figured out what’s stressing me out about programming in dynamically typed languages: It’s about always being unsure if using a library in a particular way is going to work and if minor version upgrades aren’t going to break your code.
We've taught over 100 classes to over 1000 students 🤓
The one email I ever got from Steve Jobs, and it said …
You wouldn't download a site.
Choosing fairly when you don't know how many you're choosing from.
Why splitting your codebase too early can quietly destroy your team’s velocity — and what to do instead.
It was launched to measure moisture. Now it’s ghost-trailing jammers across an electronic battlefield.
Current reasoning models trained with Reinforcement Learning with Verifiable Rewards (RLVR) often rely on manually curated datasets, raising scalability concerns and potentially limiting future AI growth beyond human-defined tasks. To address this, we propose the Absolute Zero paradigm, where a single model…
A Texas firm recently charged with conspiring to distribute synthetic opioids in the United States is at the center of a vast network of companies in the U.S. and Pakistan whose employees are accused of using online ads to scam westerners seeking help with trademarks, book writing, mobile app development and logo…
A website for old Macintosh themes and some background.
And yet it does. Are we ready to deal?
It began with a spark—a dormant passion reignited after many years away from motorcycling. Life had steered me in different directions, away from the saddle, but something deep inside pulled me back to the thrill and freedom that only riding offers. Returning to motorcycling after such an extended hiatus was daunting…
but how much exactly?
Let's take a streetview roadtrip
Looking back on the last four years, what worked, what didn't.
💡Update May 18, 2025: I described TeleMessage's trivial vulnerability in an article for WIRED: How the knock-off Signal app used by Trump officials got hacked in 20 minutes Update May 19, 2025: Uh oh: DDoSecrets publishes 410 GB of heap dumps, hacked from TeleMessage's archive server Update
[Note that this article is a transcript of the video embedded above.] In December of 2024, a huge sinkhole opened up on I-80 near Wharton, New Jersey, creating massive traffic delays as crews worked to figure out what happened and get it fixed. Since then, it happened again in February 2025 and th
Planet 9 might be our Ultima Thule . Hell would certainly be freezing over at 700 AU. Whether this tentative observation is Planet 9 or not, it is still an interesting observation and an example, I think, of proof that AIs will not do all our science for us. (At least not LLMs).
What does it mean to get things done? In the abstract, you can complete a mathematical proof or a problem set, but the real world is much fuzzier. Suppose I…
Take a look at the following code.
When I was a little kid whenever we drove by mosques, I would be intrigued by the complex motifs they’re decorated by. I always tried to figure out the pattern; to me it was just a pattern, I never…
In this post I go through the journey of overusing Kubernetes and how systemd can actually do most of what I use it for.
How 'lurking weasel' phrases skewed Cold War intelligence