Hacker News

Top Story

A security researcher’s DNS hijacking of forgotten, early-2000s phone-network routing domains (e164.arpa) accidentally logged hundreds of thousands of calls to US military bases in Diego Garcia and Ascension Island. This stunning exploit highlights how silent, legacy infrastructure rot remains a critical, nation-state-grade backdoor hiding in plain sight.

Front Page Highlights

Japan Tried to Build an OS for the World, and the US Government Intervened · XDA Developers

This deep dive untangles the tragic fate of TRON (The Real-Time Operating System Nucleus), a government-backed Japanese computing initiative from 1984 designed to replace standard files with a hypermedia document model. While its desktop variant BTRON was targeted as an “unfair trade barrier” by the USTR in 1989 and effectively killed—spurred in part by SoftBank founder Masayoshi Son lobbying politicians to protect his import business—its embedded counterpart (ITRON) quietly became one of the most widely deployed operating systems in history, running roughly 60% of the world’s embedded devices. The HN community was fascinated by TRON’s ahead-of-its-time hypermedia filesystem, which replaced fragile tree-structured directory hierarchies with an arbitrary directed graph of “real-body/pseudo-body” links. Many developers noted that modern local-first, graph-based knowledge bases (like Obsidian, Logseq, and Roam) are essentially rediscovering abstractions TRON proposed forty years ago on a custom Japanese silicon architecture.

AI Companies Destroy Physical Books — Let’s Scan Rare Books Before It’s Too Late · Anna’s Archive

A chilling report from Anna’s Archive volunteers exposes a secretive corporate land-grab where AI companies are bulk-purchasing millions of secondhand physical books through intermediaries, slicing off their bindings to scan them for model training, and then destroying them. Anthropic’s confidential “Project Panama” (revealed in a copyright settlement) and Amazon’s “VGT3” team are singled out for systematically burning through paper copies to obtain training data “untouched by machines” from before 2022 to train Claude and Nova models. The HN community expressed visceral disgust at this digital “Library of Alexandria” in reverse. Commenters fiercely debated the legal and ethical loopholes of copyright “fair use” that allow corporations to excuse literal cultural vandalism under the guise of technical training.

What Happens When the Cost of Intelligence Drops 100x · CatalystNeuro

This data-driven analysis tracks the Pareto frontier of LLM price-to-performance over the last year, showing that capability tiers are halving in cost every 4 to 10 weeks. Specifically, the intelligence level that cost $1.22 per task in February 2026 now costs just $0.022 via distilled models like GPT-5.6 Luna. Commenters pointed to the Jevons paradox in full swing: cheaper tokens aren’t reducing overall AI budgets, but are instead unlocking massive, multi-pass automated workloads (like reading every scientific paper in a field or running triple-consensus loops) that were previously cost-prohibitive.

Attention Through Arithmetic Intensity · changyi.fun

A beautifully rigorous math derivation explains how changing attention structures from MHA to GQA, MQA, and finally DeepSeek’s Multi-head Latent Attention (MLA) impacts arithmetic intensity (FLOPs per byte moved). MLA achieves a 2x AI boost over MQA by compressing keys and values into a single latent dimension that serves both computations on-chip, but as a result, its high-decode-compute overhead makes it unfriendly to speculative decoding (MTP). This is high-grade systems-engineering gold that HN loves, with commenters geeking out over cache-aware hardware-level execution and the roofline limits of NVIDIA H100s.

Stop Making TUIs · sockpuppet.org

In a highly contrarian take, the author argues that terminal-centric user interfaces (TUIs) are a legacy fetish that developers should abandon in favor of native desktop apps. With LLM agents now capable of easily summoning complex Swift/macOS boilerplate (like the author’s MDV.app markdown viewer), developers no longer have to struggle with native UI toolkits to build elegant, feature-complete desktop tools. The comment section exploded into a holy war, with old-school terminal warriors defending the speed and SSH-friendliness of TUIs while younger devs embraced the ease of agent-driven native GUI development.

There’s no such thing as a small software team anymore · jacob.gold

This post argues that parallel AI coding agents have shattered traditional team limits, enabling a small group of 5-10 engineers to generate the commit volume, PRs, and architectural complexity of an Uber-scale engineering org. To survive this deluge of autonomous code, teams may soon have to adopt hyper-modular microservice architectures just to avoid massive merge queues. HN comments were highly skeptical but intrigued, debating whether this flood of agentic code is actually delivering leverage or just creating a nightmare of unmaintainable technical debt.

I’m Becoming AI-Blind · cymerys.com

The author describes a growing cognitive block: the inability to focus on work documents that carry the distinct, verbose lingo and over-hyped tone of raw LLM outputs (“This cuts just through it”, “Redis backbone redefines the product”). This “AI trace” triggers an immediate focus rejection, forcing readers to ask basic questions already covered in the text. HN readers widely agreed, lamenting the massive flood of low-effort, low-density “token vomit” in enterprise specs and arguing that good writing requires real cognitive friction.

Show HN & Launches

• Proliferate is an open-source, self-hostable workspace designed to run multiple coding agents (Claude Code, Codex, Grok) in parallel using isolated Git worktrees, terminal execution feedback, and subagent delegation.
• Huzzah tackles agent prompting fatigue by introducing an experimental editor where you write declarative, persistent, and language-agnostic pseudocode rather than endless imperative chat instructions.
• desktop-vibe-fly is a delightfully bizarre 3D fruit fly for your macOS desktop driven by a spiking simulation of the real FlyWire connectome; it “smells” coding agent configuration files on your screen and crawls toward active projects.
• Kino is a high-performance Ractor web server for Ruby 4.0 that uses a Rust front-end to bypass Ruby’s global VM lock, achieving Puma-fork-level parallelism in a single process with up to 10x less memory.

Discussion & Debate

• Church of the TigerBeetle sparked a heavy philosophical thread on tech evangelism. The author critiqued TigerBeetle’s evangelical presentations—complete with sweeping orchestral background music and “faith-based” claims—arguing that wrapping engineering standards (like deterministic testing or consensus) in a tribal, Haskell-style affinity group degrades objective technical evaluation. Commenters debated whether structured evangelism is a necessary evil to drive adoption of robust database designs, or an organizational footgun that trains developers to rely on conviction rather than receipts.
• Claudette (a.k.a. /debuzz) and Vomit are two hilarious utilities that address Claude Code’s incurable habit of writing long-winded, self-aggrandizing PR comments that “talk like a BuzzFeed article or a TED talk”. Both tools pipe the verbose token output of Claude through Gemini or local LLMs to translate it back into direct, dry, “colleague-mode” English.

📊 I can write a Python script using matplotlib to chart the dramatic collapse of LLM costs over the past year using the model metadata from the CatalystNeuro dataset so we can visualize the trend.