Speculative KV coding: losslessly compressing KV cache by up to ~4×
Lossless compression of a target model's KV cache by up to 4×, using a cheaper predictor model to drive an arithmetic coder.
The latest blog posts to take off on Hacker News. 3,707 posts.
Lossless compression of a target model's KV cache by up to 4×, using a cheaper predictor model to drive an arithmetic coder.
Tutorial for tying the Ian's Secure Shoelace Knot, in which the laces are formed into two loops, both of which are passed through the “hole” in the middle.
Joris Rijsdijk, Christoph Peters, Michael Weinmann, Ricardo Marroquim. 2026–07 in ACM Transactions on Graphics (Proc. SIGGRAPH) 45, 4. Official version
An S-1 full of fantasies, insiders who will pocket millions, index companies that have changed the rules: it's all a recipe for regular people to have their pockets picked.
As a part of my work I do security research for various apps and websites. I wanted to see if LLMs could reproduce a common class of exploits I've found in multiple apps. So I built a deliberately vulnerable book review app and spent $1,500 finding out…
with apologies to Terry Bisson
Loris Cro's Blog
A from-the-ground-up walkthrough of how modern LLMs work, from tokens to transformer blocks to the next-token loop
I was recently diagnosed with anti-NMDA receptor encephalitis . It is an autoimmune disorder where your body’s normally helpful antibodies start acting strangely . This leads to inflammation in the brain. This blog briefly discusses some of my experience and prognosis.
I wrote the other day about Uber blowing its 2026 AI budget in four months, and how that wasn't particularly surprising given they would have set that budget in 2025, …
I have spent a large portion of my career working in Java. In that time, you get used to huge classes. New functionality? Just add a new method and field to the class. The cost of each new field is rarely considered. Performance is often considered from a classic computer science perspective by considering asymptotic…
Abusing an unauthenticated Bluetooth protocol to turn a PC speaker into a Rubber Ducky.
Including a little dive into the physics behind it
An in-depth analysis that explains how this console works internally
A Ligo research note on redundancy in natural protein fold space and what kind of data still carries signal.
I gave up blogging a long time ago (apart from an occasional thing about ArduPilot), I tend to just write code and hope people find it…
A brief explanation of my pessimism about efforts to radically improve education.
Performance results naturally depend on the underlying hardware. The following benchmarks show the execution times for sorting 50 million doubles using different sorting implementations. The measurements were taken on an Apple M1 system using Clang and on an AMD Ryzen 3 Linux system using GCC , both compiled with the…
I am now generating this website with Clojure (yes, right after rewriting it in GNU Make and shell ; I have a problem, ok?). In a personal tradition, I used writing a static site generator as the project to play with and learn Clojure as a new language. While I’ve long scoffed at Clojure for its copious syntax ( three…
The three companies (along with Mozilla, which is on one of their “ad features in the browser” kicks again) are drawing up a built-in advertising measurement system, called Attribution Level 1 , as a standard feature of web browsers. The system is intended to measure the effectiveness of advertising by enabling…
Gmail's rapid descent into belittling their users, powered by AI.
A story of sharp edges, segfaults, and standards
QBE 1.3 took a while to cook, but it is the most significant release since 1.0 with around 7k new lines of code and 1.5k deleted ones. In addition to the usual bug fixes, QBE gained a new and original IL matching algorithm, new optimizations from Roland Paterson-Jones, Scott Graham added support for the Windows ABI…
What my father taught me.
My blog, mostly about programming
When we first announced the transition to Plasma Wayland, one of Martin's slides from stated, "It's done when it's done!"
How do you ask questions that people don’t just hippopotamus?
A parallax effect built entirely in CSS with scroll-driven animations, no scroll listeners required.
Paul McGowan makes stereos. To paraphrase his insight: The musicality isn’t a feature you add to an amplifier. It’s what’s left when you stop ruining it. To expand: Customer delig…