In Part 1, I described how duplicating a block of seven middle layers in Qwen2-72B — no weight changes, no training — produced the #1 model on the HuggingFace Open LLM Leaderboard. The method, which I called RYS (Repeat Your Self), was discovered using nothing but hard math probes and EQ-Bench on a pair of RTX 4090s.
Run, compile, and link C programs instantly with CJIT.
You can go far without writing TLA+ syntax now, but you still need to understand temporal logic.
There are many interesting, useful, and fun C libraries in the software ecosystem. While one could go and rewrite these libraries in Swift, usually there is no need, because Swift provides direct interoperability with C. With a little setup, you can directly use existing C libraries from your Swift code.
A draft design document describing lightweight userland cooperative threads for SBCL.
How wanting more is not always better
A well-written, *FAST* data science database for your laptop.
NVL72 puts all 72 GPUs in a rack on one NVLink fabric, but only rack-local jobs get the speedup. How SkyPilot keeps multi-node training and inference on NVLink with topology-aware placement, in-place recovery from GPU failures, and automatic defragmentation.
If you have ever tried to minimize a function with gradient descent, you probably noticed that some functions are a joy to optimize and others are a nightmare. The difference often boils down to two properties: strong convexity and L-smoothness. These two concepts define a “sandwich” of quadratic bounds around your…
But a Redfin analysis found that AI data centers have benefits: In northern Virginia, taxes on data center equipment are likely contributing to a surge in education spending.
“The dreamers look skyward with longing. The Internet hums with its usual promise — you will find your people, your myth, your wonder, and maybe your transformation.”
I've been an Apple user for over a decade and generally very happy with that life choice, but since iOS / iPadOS 26 and MacOS Tahoe were released a few month...
Microsoft is losing Builders fast. They're switching to MacOS and Linux. The biggest pull keeping people on Windows, outside of shear inertia, is content creation and gaming. However, even these are falling to Linux.
When I started using computers, we had a Sinclair ZX Spectrum at home and a nano-reseau of Thomson MO5 s at elementary school. I distinctly remember how unpleasant it was to type with them. These must have been the worst keyboards I ever used [1] . Ever since, I have paid close attention to the keyboards I use. Here…
Rust struct serialization problem caused by many nullable SQL columns.
A question like this recently came up again. I once tried to find an answer, but only posted it in some "forum" a decade ago, and I think it's worth saving this little bit of trivia from digital nirvana.
This summary of Brad’s HIPS 2025 keynote looks at how HPC programming has (and has not) changed over the past 30 years
a personal blog
I was reading Jason Gorman’s article “The Wall Confronting Reliable Coding Agent Autonomy” and he says: the journey from 90% to 99% reliability is just as hard…
The IPv4 "exhaustion" is not real. The addresses got hoarded and the market devolved into a wild sub-leasing economy where anyone can choose where their IPs appear from, who owns them on paper, and scrub them clean of any history.
Working on an international tech team comes with communication challenges that go beyond just language. Here’s some practical advice I’ve gathered in my ten years of working in Japan.
Abusing named pipes in proprietary software to get LocalSystem privileges in seconds.
A parallax effect built entirely in CSS with scroll-driven animations, no scroll listeners required.
The Picard team is happy to announce the new major version 3.0 of MusicBrainz Picard, now available to download . MusicBrainz Picard is the official tag editor for the MusicBrainz database and helps you get your music collection sorted and cleaned up with data from MusicBrainz.
Meet North Mini Code, Cohere's first open-source agentic coding model: a 30B MoE that codes strongly on minimal hardware.
Pruning agent context down to what the answer actually needs, while keeping 96% of recall
There was a joke at the time that although the National Security Agency had opened up in recent years and started taking their defensive role more seriously, the one aspect of cryptography which they would never discuss publicly was side channel attacks — and this itself constituted a side channel, revealing the…
The FreeBSD Release Engineering Team is pleased to announce the availability of FreeBSD 14.5-RELEASE. This is the sixth release of the stable/14 branch.
Virtual reality for a 286 processor
Sometimes an article starts with a simple question or a single message on one of the available social platforms. This is the case today with this article where @ooberober asked: How small can the r…
