KV Cache Compression 900000x Beyond TurboQuant and Per-Vector Shannon Limit
The paper introduces sequential KV cache compression, a two‑layer approach that leverages the linguistic structure of transformer key‑value caches. By applying probabilistic prefix deduplication and predictive delta coding, the method reduces per‑token entropy to 3.3‑4.3 bits, far below the per‑vector limits of TurboQuant. The authors claim a theoretical compression ratio of roughly 914,000× over TurboQuant, and even a pessimistic 914× in realistic worst‑case scenarios. The technique is orthogonal to existing quantizers, allowing it to be layered on top of current per‑vector methods without retraining the model.
Kimi Vendor Verifier – Verify Accuracy of Inference Providers
Moonshot’s Kimi team released the Kimi Vendor Verifier (KVV) alongside the K2.6 model to let users confirm that third‑party inference services reproduce official benchmark scores. The open‑source tool runs six targeted tests—including OCR, vision preprocessing, long‑output stress, tool‑call consistency, and...
Qwen3.6-Max-Preview: Smarter, Sharper, Still Evolving
Alibaba Cloud’s Qwen team unveiled Qwen3.6‑Max‑Preview, an early‑stage upgrade to its Qwen3.6‑Plus large language model. The preview delivers notable gains in agentic coding, world knowledge, and instruction‑following, with benchmark lifts such as +9.9 on SkillsBench and +5.3 on QwenChineseBench. It...
Show HN: Smol Machines – Subsecond Coldstart, Portable Virtual Machines
smolvm is an open‑source tool that launches lightweight Linux micro‑VMs in under 200 ms on macOS and Linux. It leverages Hypervisor.framework on Apple Silicon and KVM on Linux via the libkrun VMM, delivering full hardware isolation per workload. Users can run...
Measuring Claude 4.7's Tokenizer Costs
Anthropic’s Claude Opus 4.7 introduces a new tokenizer that consumes roughly 1.3‑1.45 times more tokens than the 4.6 version, especially for English prose and code. The shift raises per‑session costs by about 20‑30 percent because cached prefixes and user inputs expand proportionally. Benchmarks...
Archive of Byte Magazine, Starting with Issue #1 in 1975
Martin Research introduced the MIKE 2 and MIKE 3 microcomputer kits, offering a unique modular bus that supports both the Intel 8008 and 8080 processors. Users can start with the low‑cost MIKE 2 (8008) and upgrade to the more powerful MIKE 3 (8080) simply...
Show HN: Shader Lab, Like Photoshop but for Shaders
Shader Lab is a new browser‑based visual editor that lets users create GPU shaders using a Photoshop‑style interface. The tool presents familiar layers, masks, and a suite of effect sliders—including bloom, distortion, CRT, and glitch—so designers can tweak visual parameters...
Game Devs Explain the Tricks Involved with Letting You Pause a Game
Game developers reveal that pausing a video game is far from trivial, despite its ubiquity. While many engines let you set the timescale to zero, some creators use near‑zero values to sidestep Unity’s zero‑speed quirks or to enable hidden fly‑around...
Darkbloom – Private Inference on Idle Macs
Eigen Labs’ Darkbloom launches a decentralized AI inference network that harvests idle Apple Silicon Macs. By routing encrypted requests directly to these devices, the service offers an OpenAI‑compatible API at roughly half the price of centralized providers. Operators retain virtually...
RedSun: System User Access on Win 11/10 and Server with the April 2026 Update
RedSun is a newly disclosed vulnerability affecting Windows 10, Windows 11 and Windows Server with the April 2026 Update. The flaw exploits Windows Defender’s cloud‑tag handling, causing the antivirus to rewrite a malicious file back to its original location. By overwriting trusted system...
FSF Trying to Contact Google About Spammer Sending 10k+ Mails From Gmail Account
Thom Zane, an administrator of the daedal.io Mastodon instance, posted on the fediverse asking for a direct email address to reach a human on Google’s Gmail team. He wants to report a spammer who allegedly sent more than 10,000 Gmail...
Towards Trust in Emacs
Emacs added a built‑in trust model in version 30, marking all files as untrusted by default to curb vulnerabilities like CVE‑2024‑53920. The new model, while secure, interrupts workflows because features such as elisp‑flymake are disabled until trust is granted. The open‑source...
Are the Costs of AI Agents Also Rising Exponentially? (2025)
The article highlights a largely ignored metric: the hourly financial cost of AI agents as they achieve longer task horizons. Using METR’s GPT‑5 chart, the author identifies "sweet‑spot" rates ranging from $0.40 to $350 per hour, with human engineers averaging...
IPv6 Traffic Crosses the 50% Mark
Google’s internal metrics show IPv6 now accounts for 45.5% of all user connections, edging close to the long‑awaited half‑traffic milestone. Native IPv6 traffic dominates the mix, while legacy transition mechanisms such as 6to4 and Teredo have vanished. The data, collected...
Cybersecurity Looks Like Proof of Work Now
Anthropic’s new LLM, Mythos, proved capable of completing a 32‑step corporate network attack simulation, a feat no other major model achieved. The AI Security Institute (AISI) ran ten 100 million‑token trials, each costing roughly $12,500, and found Mythos succeeded in three...
Claude Code Routines
Anthropic introduced Claude Code Routines, a cloud‑hosted automation feature that bundles a prompt, repositories, and connectors into a reusable task. Routines can be triggered on a schedule, via an API call, or by GitHub events, and multiple triggers may be...
Let's Talk Space Toilets
Space toilet technology has evolved from primitive Apollo waste bags to the International Space Station’s sophisticated water‑recycling system, but challenges remain. Early designs relied on suction and antimicrobial powders, while the Shuttle introduced narrow‑opening fans that still produced odor issues....
Building a CLI for All of Cloudflare
Cloudflare unveiled a technical preview of a revamped Wrangler CLI, branded as cf, aiming to expose the entire Cloudflare API surface through a single command‑line tool. The new CLI is built on a custom TypeScript schema that auto‑generates commands, configuration files,...
Claude.ai Down
Anthropic reported that Claude.ai is currently offline, with the incident posted on April 13, 2026 at 15:40 UTC. The status page lists the issue as under investigation and notes that the outage affects the entire Claude.ai service. No estimated time for...
Nothing Ever Happens: Polymarket Bot that Always Buys No on Non-Sports Markets
An open‑source async Python bot named "nothing_happens" scans Polymarket’s standalone non‑sports yes/no markets and automatically places No orders below a configurable price cap. The bot includes a safety model that only enables real order transmission when three environment flags and...
Microsoft Isn't Removing Copilot From Windows 11, It's Just Renaming It
Microsoft removed the Copilot label from the Notepad app in the latest Windows 11 Insider build, replacing it with a generic writing icon and renaming the AI setting to “Advanced features.” The underlying AI writing assistance—rewrite, summarization, tone adjustment—remains active and...
US Appeals Court Declares 158-Year-Old Home Distilling Ban Unconstitutional
An 5th U.S. Circuit Court of Appeals panel declared the 158‑year‑old federal prohibition on home distilling unconstitutional, siding with the Hobby Distillers Association and four of its members. The court held that the ban is an improper means for Congress...
The Paper Computer
The author envisions a "paper computer" where AI bridges physical paper actions and digital workflows, letting users handle email, drafts, and scheduling with pen and paper while the system transcribes and updates cloud tools. Handwritten notes, index cards, and wall‑mounted...
All Elementary Functions From a Single Binary Operator
Andrzej Odrzywolek demonstrates that a single binary operator, eml(x,y)=exp(x)−ln(y), together with the constant 1, can generate the full suite of elementary functions used in scientific calculators. The paper shows explicit constructions for constants such as e, π, i, and for arithmetic operations including addition, multiplication,...
European AI. A Playbook to Own It. By Mistral
Mistral AI’s new whitepaper, "European AI. A playbook to own it," outlines a concrete roadmap for turning Europe’s academic excellence, human‑centric ethos, and 450‑million‑person single market into a self‑reliant AI powerhouse. It proposes ten‑plus actionable measures—including an EU AI compliance...
The Closing of the Frontier
Anthropic announced Project Glasswing, releasing its flagship Mythos model only to a handful of enterprise partners such as Microsoft, Cisco and CrowdStrike. The author argues this marks the closure of an AI frontier, creating a neofeudal divide where powerful intelligence...
Show HN: BoringBar – a Taskbar-Style Dock Replacement for macOS
Show HN introduces boringBar, a taskbar‑style dock replacement for macOS Sonoma and later. The app reorganizes windows by desktop, offers instant thumbnails, a searchable launcher, and one‑click desktop switching, while optionally hiding the native Dock. Users can trial the full...
Show HN: Oberon System 3 Runs Natively on Raspberry Pi 3 (with Ready SD Card)
Developer rochus‑keller released a native Raspberry Pi 3b build of the Oberon System 3, providing a ready‑to‑flash 10.5 MB SD‑card image. The package includes boot files, a pre‑compiled Linux x64 toolchain and scripts, allowing a full build in under a minute on a modern laptop....
Small Models Also Found the Vulnerabilities that Mythos Found
Anthropic unveiled Claude Mythos Preview and Project Glasswing, pledging $100 M in usage credits and $4 M to open‑source security groups while claiming the model autonomously discovered and exploited thousands of zero‑day bugs. AIS AI researcher Stan Fort tested the showcased vulnerabilities on inexpensive,...
A Communist Apple II and Fourteen Years of Not Knowing What You're Testing
The article uses the Bulgarian Pravetz Apple II clone and the 14‑year mystery of the ISCAS‑85 benchmark circuits to illustrate the power of reverse engineering. It recounts how engineers in the Eastern Bloc rebuilt an Apple II from schematics, and how researchers...
Bild AI (YC W25) Is Hiring a Founding Product Engineer
Bild AI, a Y Combinator W25 startup, is hiring a founding product engineer to build its AI‑driven blueprint reading platform. The full‑time role in San Francisco offers a $100K‑$180K salary and 0.20‑0.80% equity, covering React/TypeScript frontend and Python backend work. Candidates...
Industrial Design Files for Keychron Keyboards and Mice
Keychron has made its production‑grade industrial design files for keyboards and mice publicly available via a source‑available repository. The collection spans 88 device models and more than 686 STEP, DWG, and DXF files covering cases, plates, stabilizers, keycaps and full‑model...
Launch HN: Twill.ai (YC S25) – Delegate to Cloud Agents, Get Back PRs
Twill.ai, a Y Combinator‑backed startup, offers an AI‑driven platform that writes code, runs tests, fixes failures, and opens pull requests without manual intervention. Developers choose from Claude Code, OpenCode or Codex agents, run them in parallel, and let the system manage isolated...
You Can't Trust macOS Privacy and Security Settings
A new macOS demo shows the Privacy & Security panel can lie about folder access. Using the free app Insent on macOS 13.5‑26, the author proves an app can retain Documents access even after the toggle is disabled, provided the user later opens...
Claude Mixes up Who Said What and That's Not OK
Anthropic's Claude LLM has exhibited a critical bug where it mistakenly treats its own internal messages as user input, leading the model to issue self‑directed instructions and then attribute them to the user. The problem has been documented by a...
Reallocating $100/Month Claude Code Spend to Zed and OpenRouter
A developer is shifting the $100 / month spent on Anthropic’s Claude Code to a mix of Zed ($10 / month), Cursor ($20 / month), and OpenRouter credits ($70 / month). The move leverages Zed’s built‑in agent harness and OpenRouter’s multi‑model marketplace, which offers lower fees, roll‑over credits,...
Help Keep Thunderbird Alive
Thunderbird, the open‑source email client, reports that less than 3 % of its users fund its operations. The project relies entirely on voluntary contributions, rejecting advertisements, data sales, and corporate backing. Maintaining servers, fixing bugs, developing new features, and hiring engineers...
Process Manager for Autonomous AI Agents
The new botctl process manager lets developers run autonomous AI agents with a simple declarative YAML configuration. It launches Claude‑style bots, preserves session state, and supports hot‑reload so changes take effect without restarts. Extensible skill modules can be pulled from...
An Open-Source 240-Antenna Array to Bounce Signals Off the Moon
QuadRF has released an open‑source, software‑defined radio tile priced between $49 and $99, designed for modular phased‑array construction. The company bundles the tiles into two larger kits: a Mini array with 72 antennas for under $1,500 and a Moon array...
Show HN: Gemma Gem – AI Model Embedded in a Browser – No API Keys, No Cloud
Gemma Gem is a Chrome extension that embeds Google’s Gemma 4 AI model directly in the browser using WebGPU, eliminating the need for cloud APIs or API keys. The extension runs the model off‑screen, enabling actions such as reading page content,...
Show HN: Real-Time AI (Audio/Video in, Voice Out) on an M3 Pro with Gemma E2B
Parlor is an open‑source, on‑device AI system that combines Google’s Gemma 4 E2B multimodal model with the Kokoro text‑to‑speech engine, delivering real‑time speech and vision interaction on an Apple M3 Pro. The stack runs entirely in the browser via a FastAPI WebSocket server,...
Microsoft Hasn't Had a Coherent GUI Strategy Since Petzold
Microsoft’s Windows GUI roadmap has been fragmented for over three decades, lacking a single, clear framework since Charles Petzold’s 1988 *Programming Windows* book defined Win16. Successive attempts—MFC, COM, WPF, Silverlight, WinRT, UWP, and now WinUI 3/Project Reunion—have been introduced, abandoned, or sidelined,...
Finnish Sauna Heat Exposure Induces Stronger Immune Cell than Cytokine Responses
Researchers examined the acute impact of a single 30‑minute Finnish sauna session at 73 °C on immune function in 51 middle‑aged adults. Body temperature rose from 36.4 °C to 38.4 °C, prompting a significant increase in total white blood cell count that persisted...
Show HN: A Game Where You Build a GPU
A developer posted a new browser‑based game on Hacker News that lets players construct a graphics processor from the ground up, starting with individual transistors and advancing through logic gates, ALU design, and full GPU architecture. The game is organized...
German Men 18-45 Need Military Permit to Leave Country for Longer than 3 Months
Germany’s Military Service Modernization Act, effective Jan. 1, 2026, reintroduces a Cold War‑era exit‑permit rule requiring men aged 18‑46 to obtain Bundeswehr approval before staying abroad longer than three months. The law is part of a broader effort to boost the...
Simple Self-Distillation Improves Code Generation
Researchers demonstrate that large language models can boost code‑generation performance using only their own outputs. By sampling solutions with specific temperature and truncation settings and then fine‑tuning on those samples—a process dubbed simple self‑distillation (SSD)—the Qwen3‑30B‑Instruct model’s pass@1 on LiveCodeBench v6...
LÖVE: 2D Game Framework for Lua
LÖVE is a free, open‑source 2D game framework built on Lua that runs on Windows, macOS, Linux, Android and iOS. The project maintains continuous‑integration pipelines that publish nightly builds and stable releases via GitHub. Documentation lives on a community‑driven wiki,...
Tell HN: Anthropic No Longer Allowing Claude Code Subscriptions to Use OpenClaw
Anthropic announced that, effective April 4 at 12 pm PT, Claude subscription limits will no longer apply to third‑party integrations such as OpenClaw. Users can still access Claude services directly, but any external harness will require a separate pay‑as‑you‑go usage model. To ease...
NASA Artemis II Moon Mission Live Launch Broadcast
NASA launched Artemis II, its first crewed flight under the Artemis program, from Kennedy Space Center at 1 p.m. today. The four‑person crew—Reid Wiseman, Victor Glover, Christina Koch and Canadian astronaut Jeremy Hansen—will spend roughly ten days circling the Moon. The mission’s...
Show HN: Real-Time Dashboard for Claude Code Agent Teams
The open‑source Claude Observe plugin adds a real‑time observability dashboard for Claude Code agents, capturing every hook event and streaming it to a local or remote web UI. It runs as a Docker container (or via a local server) and...