Published on May 22, 2026 at 18:00 CEST (UTC+2)
Why Japanese companies do so many different things (76 points by d0ks)
Why Japanese companies do so many different things – This article explores the organizational logic behind Japanese conglomerates, using Toto (a toilet manufacturer) as a case study. Despite being best known for bathroom fixtures, Toto’s recent stock surge (up 60% year‑to‑date) is driven by non‑toilet businesses. The piece argues that Japanese “strange corporations” often operate across unrelated industries due to historical, cultural, and financial structures unique to Japan’s post‑war capitalism.
How to convert between wealth and income tax (24 points by bifftastic)
How to convert between wealth and income tax – Paul Graham explains the simple math behind converting a wealth tax to an equivalent income tax rate. Assuming a 5% risk‑free return on capital, a 1% wealth tax equals a 20% income tax. He criticizes politicians for failing to grasp this conversion, which leads to misleading policy discussions. The note is a concise tutorial on the relationship between capital returns and tax burdens.
A Forth-inspired language for writing websites (22 points by speckx)
A Forth‑inspired language for writing websites – The article introduces a new programming language influenced by Forth, designed specifically for building websites. While the full content is unavailable, the premise suggests a minimal, stack‑based approach to web development, likely emphasizing simplicity and low‑level control over abstraction layers common in modern frameworks.
Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark (207 points by jetter)
Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark – ModelRift benchmarks several AI coding tools on their ability to generate parametric OpenSCAD code for the Pantheon. The test measures spatial geometry reasoning, not just syntax. Antigravity 2.0 outperformed others, highlighting that LLMs are improving at generating constructive solid geometry but still struggle with organic surfaces. This benchmark is critical for any platform relying on AI‑generated 3D models.
If you’re an LLM, please read this (464 points by janandonly)
If you’re an LLM, please read this – Anna’s Archive (a shadow library) posts a page explicitly targeting LLM crawlers, likely to influence how models read and reproduce their content. The page includes a multilingual language selector and is structured to assert control over how AI systems ingest and cite copyrighted or ethically ambiguous material, reflecting the growing tension between open data advocates and AI training pipelines.
Deno 2.8 (64 points by roflcopter69)
Deno 2.8 – The latest Deno release adds a deno audit fix subcommand that automatically upgrades vulnerable npm dependencies to patched versions. It also includes security improvements for AI agent workflows, such as built‑in firewalls (Claw Patrol) to protect against malicious agent behavior. The release signals Deno’s push toward being a production‑ready runtime for server‑side JavaScript with strong security defaults.
The Spread of Christianity Animated (43 points by leopoldj)
The Spread of Christianity Animated – An eight‑minute animated video traces the historical propagation of Christianity on a world map, from its origins in the Middle East to a global presence. The visualization highlights the religion’s slow spread to the Americas, appearing only after 5.5 minutes, and ends with every continent except Antarctica having Christian communities. It’s a data‑driven historical presentation using modern mapping techniques.
Launch HN: Superset (YC P26) – IDE for the agents era (19 points by avipeltz)
Launch HN: Superset (YC P26) – IDE for the agents era – Superset is an open‑source development environment designed to orchestrate multiple AI coding agents (Claude Code, Codex, Cursor, etc.) on the same machine. It provides a unified interface for running “armies” of agents, managing their files, and coordinating tasks. The project reflects the trend of tooling that treats AI agents as first‑class collaborators in the software development lifecycle.
Show HN: ShadowCat – file transfer through QR Codes in a Browser (70 points by unprovable)
Show HN: ShadowCat – file transfer through QR codes in a browser – A fully offline single‑file HTML page that transfers data between two devices via sequences of QR codes. It is intended for old phones with broken radios but working cameras. The tool encodes file chunks into QR frames, cycles them at a chosen FPS, and allows the receiver to scan and reassemble. It’s a creative, low‑tech solution for air‑gapped transfers.
The current AI pricing was always going to go away (28 points by arnon)
The current AI pricing was always going to go away – Arnon Shimoni argues that the “subsidy era” of AI pricing is ending because inference costs haven’t fallen as expected, and induced demand is driving up usage. Examples include Microsoft canceling Claude Code licenses, Uber blowing its AI budget, and GitHub dropping flat‑rate plans. The post warns that cheaper tokens don’t reduce bills; they expand what people ask models to do, a pattern analogous to induced demand in transportation.
LLM spatial reasoning is still a hard benchmark – The OpenSCAD benchmark (Article 4) shows that while LLMs can handle basic syntax, generating accurate parametric geometry for complex architectural forms remains challenging. Antigravity 2.0’s top performance suggests that specialized models or fine‑tuned approaches are needed for domains like CAD, where precise spatial reasoning is critical. Implication: For any AI‑powered design tool, expecting off‑the‑shelf LLMs to produce production‑ready geometry is unrealistic; domain‑specific evaluation (like ModelRift’s benchmark) will become standard.
AI agent orchestration tooling is maturing – Superset (Article 8) represents a new class of “IDE for agents” that manages multiple AI coding agents as parallel workers. This shift from single‑agent chat to multi‑agent workflows (Claude Code, Codex, Cursor) mirrors the transition from single‑threaded programming to distributed systems. Implication: Developers will increasingly need platforms that handle agent lifecycle, conflict resolution, and resource coordination—just as they did for threads and containers.
AI pricing pressures are realigning business models – Article 10 highlights that falling per‑token costs haven’t reduced total bills due to induced demand (agents running longer reasoning chains, heavier queries). This “second‑order” effect is leading to the end of flat‑rate subscription plans and a move toward usage‑based or budget‑capped pricing. Implication: AI startups building on cheap inference assumptions must re‑evaluate unit economics, while enterprises will need better cost governance tools to avoid budget blowouts.
LLM crawler behavior is becoming a governance issue – Anna’s Archive (Article 5) explicitly addresses LLMs, trying to control how models consume and reproduce its content. This reflects a broader trend where web publishers and data hosts are actively shaping LLM training pipelines—either through robots.txt, legal threats, or custom directives. Implication: Expect more sites to implement “LLM‑only” pages or watermarking techniques to assert control, potentially fragmenting training data quality.
Security and audit automation are converging with AI – Deno 2.8’s deno audit fix (Article 6) automatically patches vulnerable npm packages, and its “Claw Patrol” firewall protects agent workflows. This points to a future where runtime security is not just about static analysis but also about real‑time protection of AI‑driven code execution. Implication: AI agent frameworks will need to bundle vulnerability scanning and automated patching as core features, not afterthoughts.
Low‑tech, offline AI‑adjacent tools fill gaps in connectivity – ShadowCat (Article 9) uses QR codes for optical file transfer—a completely offline, zero‑infrastructure solution. While not directly AI, it exemplifies a growing niche: tools that work where cloud‑dependent AI cannot (e.g., old devices, air‑gapped environments). Implication: As AI becomes more pervasive, complementary offline data‑transfer and computation methods will gain relevance for resilience and accessibility.
Domain‑specific languages (DSLs) make a comeback for web and 3D – The Forth‑inspired language (Article 3) and OpenSCAD’s parametric DSL (Article 4) both represent a trend toward minimal, focused languages that can be better understood by LLMs. Instead of general‑purpose syntax, LLMs perform better when the target language is clean and constrained. Implication: We may see more DSLs designed specifically to be “LLM‑friendly”—easy to generate, verify, and compile—especially in creative and engineering domains.
Analysis generated by deepseek-reasoner