Normal Science

Prompt Injection as Role Confusion

Prompt Injection as Role Confusion First, I absolutely love this: This is a blog-style writeup of the paper. I wish every paper would come with one of these. Academic writing is pretty dry - the impact of a paper can be so much higher if you publish a readable version to accompany the formal one. Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell present some fascinating research into the challenge of having models distinguish their own privileged text (here wrapped in role tags like <system>, <...

Bipartite matching is in NC!

Since I’m a good mood today—at a beautiful science camp with my kids, high in the mountains near Big Bear Lake in California—I thought I’d blog about something positive. Last week, five authors posted a major paper to the Electronic Colloquium on Computational Complexity, which shows (or anyway, credibly claims to show) that the Bipartite Matching problem is in the complexity class NC. Assuming this stands, it resolves a central problem in parallel algorithms and derandomization that’s been open...

Seeing the world in radio waves with the QuadRF

Although the basic principle of radio direction finding is easy to understand (measure the phase difference between different antennas, then calculate the angle of arrival from this difference), th…

Scaling opencomputer from 1 VM to 1 million sandboxes

From one VM to a million sandboxes: the architecture redesign behind OpenComputer

GLM-5.2 is probably the most powerful text-only open weights LLM

Chinese AI lab Z.ai released GLM-5.2 to their coding plan subscribers on June 13th, and then yesterday (June 16th) released the full open weights under an MIT license. Similar in size to their previous GLM-5 and GLM-5.1 releases this is a 753B parameter, 1.51TB monster - with 40 active parameters (Mixture of Experts). GLM-5.2 is a text input only model - Z.ai have a separate vision family most recently represented by GLM-5V-Turbo, but that one isn't open weights. GLM-5.2 has a 1 million token co...

Should I run plain Docker Compose in production in 2026?

Yes, plain Docker Compose can still run production workloads in 2026—if you close the operational gaps it leaves: cleanup, healing, image pinning, socket security, and updates.

Computer Use is 45x more expensive than structured APIs

We benchmarked computer use against auto-generated API endpoints on the same admin panel. 53 steps and 551k tokens vs 8 calls and 12k tokens.

Bun is being ported from Zig to Rust

Incredibly fast JavaScript runtime, bundler, test runner, and package manager – all in one - docs: add Phase-A porting guide · oven-sh/bun@46d3bc2

Show HN: Tilde.run – Agent sandbox with a transactional, versioned filesystem

Tilde turns running AI agents and pipelines on real data into a transactional, auditable operation. Compose GitHub, S3, and Drive as one versioned filesystem. Roll back any run with one command. Every network call audited.

RaTeX: KaTeX-compatible LaTeX rendering engine in pure Rust

Rust TeX-style math layout with KaTeX-aligned golden tests. Ready-to-use packages for Web (WASM), iOS, Android, Flutter, and React Native—same display list everywhere.

Batteries Not Included, or Required, for These Smart Home Sensors

Engineering and computing researchers create simple metal tags with unique ultrasonic fingerprints to detect door openings and other movements.

Docker 29 has changed its default image store for new installs

Learn about the containerd image store

Reverse-engineering the 1998 Ultima Online demo server

Draxinar Reverse-engineering the 1998 Ultima Online demo server May 1, 2026 After 10 years of on-and-off work, I’m releasing a full reverse-engineering of the 1998 Ultima Online demo server: https://github.com/draxinar/ouo. About 5,000 functions disassembled from MSVC x86 and translated into portable C99, with each function compared instruction-by-instruction against the binary. UoDemo.exe For those who don’t know, Ultima Online is a 1997 MMORPG developed by Origin Systems Inc. It was one of the

Virtual violin produces realistic sounds

MIT researchers developed a “computational violin” — the first computer simulation that captures the detailed physics of the instrument and realistically produces the sound of a violin when its strings are plucked. Violin makers could use the model to test how a violin might sound when certain dimensions or properties are changed.

Comparing the Z80 and 6502 to Their Relatives

Last week’s adventures with the Exidy Sorcerer led me to write a Z80 version of the LZ4 decompressor I’d previously used on the SNES, the CoCo, and the Genesis. At this point, this has …

From CVS to Git, thirty years of source control

In April 2005, Linus Torvalds wrote Git in ten days because BitKeeper revoked its free licence to the Linux kernel. Twenty-one years later, no successor has emerged. A practitioner's history of source control from someone who used every major system since 1990, and lost code in most of them.

Testing macOS on the Apple Network Server 2.0 ROMs

It's time for another save point in the continuing saga of the various ROMs for the Apple Network Server , Apple's first through-and-through...

Collaborative Editing in CodeMirror (2020)

Bun Has Been Converted to Rust. Now What?

Bun was ported from Zig to Rust by an LLM and passed nearly all its tests - while shipping more than ten thousand unsafe blocks. Memory safety is the main reason you'd pick Rust. So what did the rewrite actually accomplish?

Discovering hard disk physical geometry through microbenchmarking (2019)

BlogRandom stuff…CommentsPosts Home Blog About Uncategorized Measuring Stuff Fixing Stuff Building Stuff Taking apart stuff &laquo; Elections 2019: Who is really getting a tax cut? A Cardboard Haswell Box &raquo; Discovering Hard Disk Physical Geometry through Microbenchmarking By Henry, on September 6th, 2019 Modern hard drives store an incredible amount of data in a small space, and are still the default choice for high-capacity (though not highest-performance) storag

Finding the differences in a series of power supplies

Show HN: nfsdiag – A NFS diagnostic application

A nfs doctor application. Contribute to lsferreira42/nfsdiag development by creating an account on GitHub.

Speedup in Lattice Boltzmann Cylinder Flow

Resolution robustness of vortex shedding in Lattice Boltzmann cylinder flow: a scaling study for reduced-cost simulation. - alikamp/Parks-KPBM-Scaling

Ripping a DVD, a federal crime in 1999, requires $22 and free software in 2026

I went to rip a movie from my own DVD collection. It surfaced dead anti-piracy infrastructure, a defunct Japanese DRM company, and a twenty five year old executable that was still trying to phone home

Rolling the Root Key

Have DNSSEC-validating recursive resolvers updated their Trust Anchor sets to include KSK-2024, and how can we measure whether this transition has been successfully adopted?

Homebridge 2.0 is here, and it speaks Matter

The open-source Apple HomeKit bridge is adding Matter, allowing it to support Alexa, Google Home, Home Assistant — and robot vacuums.

MIT’s virtual violin offers luthiers a new design tool

Computational model lets users tweak parameters to hear effect on the sound in early design process.

Show HN: Tight C, a systems language with 10 keywords

A minimalistic portable assembly lenguage. Contribute to alonsovm44/tc-lang development by creating an account on GitHub.

RuView – See through walls with WiFi

π RuView turns commodity WiFi signals into real-time spatial intelligence, vital sign monitoring, and presence detection — all without a single pixel of video. - ruvnet/RuView

America's Data Center Build-Out Is Falling Way Behind Schedule

Starlink shifts hardware from one-time purchase to $10/month rental

Starlink, SpaceX's top moneymaker, also raised service prices by $5 to $10.

Thunderbolt vs. USB-C: what the connector hides

USB-C is the connector. Thunderbolt is one of the protocols that runs through it. Here's what actually changes between TB3, TB4, USB4, and TB5.

Klaxon a livr earthquake map with no back end

Real-time global emergency event tracker. Live earthquakes, with tsunamis, wildfires, and conflicts coming soon. Data from USGS.

A Simpler Parametrization for Modern Optimizers

Bun's unreleased Rust port has 13,365 unsafe blocks

An audit of every unsafe block in Bun

Outlook on Windows silently scales your email by 1.5×

A Windows display setting most users never touch is multiplying pixel values inside Microsoft Word's HTML engine. Half your recipients see your 600px hero at 900px. The fix is an XML namespace from 1996.

Wasi: WebGPU – A Proposed WebAssembly System Interface API

Contribute to WebAssembly/wasi-webgpu development by creating an account on GitHub.

How to make firecracker faster to start Chromium in < 20ms

kernel got our start running chromium in a firecracker vm. we spent the past 14 months making it even faster. we're sharing how we get the most out of firecracker using snapshots, copy-on-write forks, uffd paging, and hot pools.

A 35B MoE on a 16 GB GPU, without the offload tax

A 33-35B MoE only fires ~8 of 256 experts per token, but holding it on the GPU costs you all of them. Luce Spark pins the experts your traffic uses, offloads the rest to CPU, and decodes the whole token in one fused graph, so a 33-35B MoE fits a 16 GB card and still runs near full-GPU speed (~100 tok/s vs ~119 all-GPU on a 3090). Self-tuning, one flag.

Freelang – a Libc-free, direct sys/kernel call language with weird concurrency

FreedomLang Features Runtime Genealogy Philosophy Code Fit Status GitHub No libc. No VM. No hidden jobs. Native x86-64. OS-process jobs. No libc. "Model the world as data. Treat bugs as fatal." FreedomLang is a small AOT systems language that lowers source through a compact IR to libc/CRT-free native x86-64. Linux emits ELF64 machine-code bytes directly; macOS and Windows emit platform assembly. Jobs are real OS processes with filesystem-visible state. View on GitHub See the Code ✓ Linux

Show HN: Rotunda - A browser built for agents with simulated typing

Hi HN! Pierce here.Rotunda is a firefox fork primarily intended for agent use, which I’ve been hacking on nights/weekends.There was a [lengthy](https://news.ycombinator.com/item?id=48024859) discussion last week on how expensive computer use models are. The cost is going to drop eventually, but I think on some level it's still usually the wrong primitive. The web gives us access to beautiful structured formats, plaintext, etc... why throw that away if we don't have to?I realized at some point that for 99% of automations I just want agents to be able to control my Chrome instance. But that’s easier said that done: CDP (the Chrome automation protocol) leaks a ton of state about being programmatically controlled, either by toggling window attributes or by running `page.evaluate()` commands right in the page context. Plus if you look at an automation running it's pretty obvious what happens: the mouse jumps around, fields are filled instantly, etc.Rotunda tries to fix this. Its standout features:- Realistic simulation of mouse movements and keyboard commands, powered by a trained RNN on my own timing patterns from the last week. (still feel weird about opting-in to a key logger but whatever)- Doesn’t lie about its host specs, only fibs about some client side details. Stealth browsers are too easy to flag statistically when you’re adding noise to canvas pixels or audio pipelines.- It runs on your local device with a CLI or Playwright API accessible to Claude, Codex, or whatever your harness-de-jure today looks like.- Patches modern Firefox (150) with an agentic harness to keep this updated over timeMPL-2.0 on GitHub: https://github.com/monkeysee-ai/rotundaLonger writeup on the design choices: https://pierce.dev/notes/a-browser-for-agentsAlso check out the demo on the site! https://www.rotunda.sh/Pretty excited by how this turned out but we’re still super early. Give it a try and please flag any issues!

Pyro Caml Continuous Profiler for OCaml

Semgrep open-sources Pyro Caml 1.0.0, the first continuous profiler for OCaml. Learn how it uses Memprof, OCaml Runtime Events, and Pyroscope to profile production workloads running under gVisor with under 5% overhead.

Tuning LLVM's SLP Vectorizer Cost Model

Similar to my last post, this writeup covers how I solved a performance regression on LLVM by analyzing a benchmark from a RISCV target.

We tested super-resolution pre-filter for LPR OCR. It did nothing

Does AI upscaling actually improve plate reads? We tested a 42K-parameter super-resolution model and a 1.21M-parameter pretrained model as pre-processing steps for ALPR OCR on thousands of labeled production crops. Both did nothing.

Automating Hermitage to see how transactions differ in MySQL and MariaDB

Submitting again after https://news.ycombinator.com/item?id=47990047 now that the paywall has expired.

Show HN: I built a sovereign OS, L1 blockchain, AI agent, and language

I built IONA over 10 years, working alone. It's a complete sovereign digital ecosystem — OS (PC + Phone), L1 blockchain, programming language, and on-device AI — all written from scratch in Rust, without Linux, BSD, or Windows code.What's in the GitHub org (github.com/Ionablokchain): - Iona-OS — bare-metal x86_64 kernel, boots in QEMU, 50+ tests - Iona-OS-Phone — ARM64 mobile OS - Iona-protocol — security-first L1 blockchain - Carpel — a systems language that eliminates lifetimes and Pin - Flux — a temporal-quantum language for a new computing paradigm - Nihilo-OS — an OS without files, where mistakes are parallel universes - Iona-AI — autonomous AI agent with self-healing kernel integrationAll written by one person over 10 years. I'd love honest technical feedback.

Evaluating Geekbench 6

Applications vary wildly in what they demand from a system, making it difficult for a single benchmark to provide a broadly representative score.

The adder at the heart of Intel's 8087 floating-point chip

In 1980, Intel released the Intel 8087 floating-point coprocessor, a chip that could make math up to 100 times faster. As well as arithmetic and square roots, the 8087 computed transcendental functions including tangent, exponentiation, and logarithms. But it all depended on a 69-bit adder: "The arithmetic heart of the floating-point execution unit is centered about a nanomachine comprised of the adder and its related registers, shifters and control circuitry," as the patent describes it. In thi...

DiffusionGemma

DiffusionGemma Last May Google briefly released an experimental Gemini Diffusion model. I tried the preview at the time and recorded it running at 857 tokens/second. It was an exciting model, but Google made no further announcements about it. That research has returned in the best possible way: as a new open weight (Apache 2 licensed) Gemma model, google/diffusiongemma-26B-A4B-it. NVIDIA are currently hosting the model for free on their NIM cloud API. I used that API to generate this pelican, wh...

Powering up a module from the IBM 604: an electronic calculator from 1948

1948 was an interesting time for computing. For decades, businesses had used punch card equipment that added and sorted electromechanically. Now these electromechanical relays and counting wheels were being used to build room-filling general-purpose computers such as Harvard Mark I (1944) and IBM's SSEC (1948). But slow electromechanical mechanisms were already becoming obsolete. World War II had fostered the development of electronics and vacuum tubes for radio, radar, and navigation. Electroni...