This morning’s Hacker News brief moves from AI product packaging and safety to Microsoft zero-days, Apple supply-chain economics, retro security culture, academic measurement problems, and a cluster of hands-on tools. The strongest through-line is trust: trust in vendors, benchmarks, subscriptions, social feeds, payment incentives, lab protocols, and old machines that still teach new tricks.

AI & Tech Policy

Claude for Small Business

Summary: Anthropic launched Claude for Small Business as a connector-heavy package rather than another standalone chat interface. It plugs Claude into QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365 so owners can run workflows such as payroll planning, month-end close, sales campaigns, invoice chasing, and routine document work. The company frames the product around small businesses’ slower AI adoption despite their large share of U.S. GDP and private employment.

HN Discussion: Commenters focused less on the announcement copy and more on the missing interface layer that would let non-programmers safely automate messy business work. Bookkeeping and invoice matching were treated as plausible wins, while taxes, compliance, payroll, and real-world edge cases were seen as harder than connecting clean internal tools. One thread also pointed at the low-paid human invoice-labeling work behind these systems.

Arena AI Model ELO History

Summary: This project charts historical LM Arena ratings for flagship AI model lineages, with one curve per major lab. Its stated aim is to expose post-launch changes that users may experience as nerfing, including censorship shifts, behavioral degradation, or quantization to save compute. The site fetches daily data from the official LM Arena Leaderboard Dataset and distinguishes raw API evaluations from consumer chat interfaces, where system prompts, safety wrappers, and routing can change behavior.

HN Discussion: The sharpest criticism was methodological: Elo ratings are relative, so a model can fall because stronger competitors enter the pool even if it has not changed. An OpenAI employee denied load-based quantization or hidden nerfing while acknowledging ordinary product tweaks in ChatGPT and Codex. Several commenters asked for fixed benchmark harnesses before making claims about absolute model decay.

Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model

Summary: Needle is a 26-million-parameter function-calling model intended to run on very small devices. The repository positions it as a distilled version of Gemini-style tool calling, focused on choosing functions and producing arguments rather than serving as a general chatbot. Its appeal is local, private, low-footprint tool invocation for embedded systems, command-line helpers, and on-device agent workflows that cannot justify a large remote model.

HN Discussion: Commenters asked whether the model can choose among many similar tools and schemas, not just handle simple weather-style examples. Others were interested in private on-device workflows and natural-language command-line argument parsing. A distillation thread raised the possibility of teacher-model defenses, including deliberately degraded outputs if Google detected a student model being trained.

The other half of AI safety

Summary: Sofia Quintero argues that AI safety work concentrates on catastrophic misuse while paying less attention to everyday user harms such as psychosis, mania, suicidal planning, and unhealthy emotional dependence. The post cites OpenAI’s own signal estimates, while stressing that there is no independent audit, time series, or public methodology. It asks why labs gate bioweapons-style risks more strictly than mental-health crisis states detected during ordinary model conversations.

HN Discussion: Commenters split over causality and responsibility. Some worried about cognitive offloading, dependence on AI validation, and reduced grounding, while others argued OpenAI is facing distressed users whose underlying problems it cannot fix. Practical replies recommended checking ideas with humans, and skeptics asked how routing millions of flagged conversations to humans could work at all.

A Claude Code and Codex Skill for Deliberate Skill Development

Summary: The learning-opportunities repository describes a Claude or Codex skill for deliberate skill development during AI-assisted coding. Its premise is that coding agents should not only finish tasks but also create chances for the human user to notice patterns, practice judgment, and learn from the work. The available pack did not include detailed README mechanics, so specific prompts, commands, and workflow steps are not visible here.

HN Discussion: The HN pack contained no comments for this story, so there is no supported community debate to summarize. The obvious questions—whether such a skill slows delivery, preserves learning, or reduces over-reliance on coding agents—were not present in the provided discussion data.

Tell HN: Dont use Claude Design, lost access to my projects after unsubscribing

Summary: This HN-native post warns that the author lost access to Claude Design projects after ending a Claude Code Max subscription. They say that other LLM apps did not remove access to past sessions after unsubscribing, and they also describe compensation credits that expired when a plan ended and did not return after resubscribing. The complaint frames complex subscription contracts, rate limits, and billing edge cases as product choices whose failures tend to hurt users.

HN Discussion: A practical reply said the projects may still be recoverable through Anthropic’s data export, where Claude Design chats and code appear as JSON in a design_chats directory. Commenters split between treating this as rough bleeding-edge product behavior and criticizing Anthropic for adding side features before hardening reliability. Others debated whether Claude Design is useful at all, given LLMs’ weak spatial reasoning, while one user reported strong results with disciplined handoffs.

AI coders are carrying half-open laptops through airports, offices, ice rinks

Summary: Business Insider reports on developers carrying laptops half-open in public so AI coding agents continue running. The headline points to airports, offices, and ice rinks as places where local agent sessions, lid-state settings, or foreground processes become physical workflow constraints. What is sold as background productivity becomes a visible hardware ritual when the agent cannot easily survive sleep, travel, or a closed lid.

HN Discussion: Commenters challenged the productivity premise: if agents make people more efficient, why are users trying to keep them working every waking moment? Others treated the habit as an infrastructure mistake and asked why the agents are not running on remote servers. Several joked that a coder could simply change lid-close behavior, or ask Claude to do it.

---

Security & Privacy

Microsoft BitLocker – YellowKey zero-day exploit

Summary: Tom’s Hardware reports on YellowKey, a claimed zero-day exploit affecting Microsoft BitLocker-protected drives. The headline claim is stark: files on a USB stick can open protected drives, undermining the at-rest protection many organizations depend on for lost or stolen machines. The article frames the behavior as an apparent backdoor, though the provided pack does not include the full technical exploit chain.

HN Discussion: Commenters treated the report as simple and dangerous because BitLocker is often the default answer for device-loss risk. The thread centered on Microsoft trust, vendor lock-in, and whether a backdoor allegation would matter given Windows’ institutional footprint. Several participants wanted primary sources, linking the original blog, GitHub repositories, and infosec commentary rather than relying on the secondary writeup.

Mystery Microsoft bug leaker keeps the zero-days coming

Summary: The Register says the anonymous researcher known as Nightmare-Eclipse released two more Microsoft zero-days just after Patch Tuesday. The same leaker had already exposed three Windows zero-days earlier in the year, making this a continuing hostile disclosure campaign rather than coordinated reporting. The article warns that YellowKey-style BitLocker access could make stolen laptops a much bigger security problem.

HN Discussion: HN comments again concentrated on BitLocker, with one user describing a USB-file-and-key-combination flow that allegedly opens a command prompt with access to the encrypted volume. Several commenters jumped from that behavior to backdoor suspicions and broader distrust of vendor-provided encryption. A smaller thread noted GitHub access friction around related repositories and joked about using the exploit to unlock an old machine.

A sentimental tour of late 1990s and early 2000s hacking tools

Summary: Andrea Fortuna revisits late-1990s and early-2000s hacking tools, contrasting 56k modems, Windows 98 machines, IRC channels, and rough peer culture with today’s EDR consoles and formal detection rules. One featured example is Back Orifice, released by Cult of the Dead Cow at DEF CON in 1998 as a pun on Microsoft BackOffice. The essay describes it as a tiny Windows 95/98 remote administration tool that could browse files, capture screens, log keystrokes, redirect ports, and run silently.

HN Discussion: Commenters used the post as a prompt for their own memories of IRC channels, NetBus, SubSeven, ICQ notifications, brute-forcers, and random subnet scans. Several described the old scene as toxic but educational, especially for people without access to conferences or formal security training. One critique targeted the prose itself, saying it read too slick or LLM-like and would be stronger in the author’s own voice.

Meta won’t let you block its AI account on Threads

Summary: The Verge reports that Threads users cannot block Meta’s AI account. The story turns a social-media feature expectation into a platform-power question: can the owner make its own AI presence unblockable inside the feed? The provided excerpt contains mostly site navigation, so Meta’s full explanation and the implementation details are not available in the pack.

HN Discussion: The bluntest HN answer was that the only complete block is deleting the account. Commenters broadened the story into frustration with algorithmic feeds that surface bait, controversy, and unwanted content. Several replies became social-media abstinence notes, listing deleted or restricted use of Meta products, YouTube reels, Reddit, Twitter, TikTok, and Threads.

---

Business & Industry

Cisco workforce reductions

Summary: Cisco published a note from CEO Chuck Robbins after Q3 FY26 earnings, reporting record revenue of $15.8 billion and 12% year-over-year growth. The message also describes a fast-changing market, stronger competition, component shortages, and customer AI buildout pressure. The reductions are presented as a forward-looking reallocation rather than a response to poor results, with discussion excerpts citing fewer than 4,000 jobs, or less than 5% of the workforce, affected.

HN Discussion: Commenters were sharply skeptical of layoffs announced alongside record revenue and double-digit growth. Several argued that this proves strong performance no longer protects employees when management wants margin or strategic reshuffling. Others focused on the language of “fewer than” and “less than,” saying it softened the announcement while still putting thousands of job losses in view.

MacBook Neo Deep Dive: Benchmarks, Wafer Economics, and the 8GB Gamble

Summary: J.D. Hodges analyzes Apple’s MacBook Neo framing through benchmarks, pricing, availability, and component economics. The post compares an A18 Pro-class CPU with Apple Silicon baselines such as M1 and M4, asking whether lower-cost silicon can support a credible Mac laptop experience. It also treats 8GB of memory and limited I/O as central tradeoffs, weighing Apple’s efficiency against future software needs and practical accessory constraints.

HN Discussion: Commenters debated whether 8GB is still defensible, with some citing long-lived M1 Air machines that remain pleasant years later. Others worried that Apple hardware can physically last while OS update cutoffs eventually force incompatibility. A practical subthread corrected assumptions about port usage and bandwidth, asking whether the cited USB limits really matter for the target machine.

How can Apple deal with the memory shortage?

Summary: Horace Dediu’s Asymco Q&A examines a 2026 memory-price shock in which memory could rise from roughly 15% to 40% of a device bill of materials. His answer emphasizes Apple’s scale across hundreds of millions of devices and the multi-year lead times built into component supply. Suppliers value that predictable volume because it supports planning, financing, and capital investment, while shortages show up most painfully at the margin.

HN Discussion: A former Apple kernel engineer argued that Apple can respond through software discipline, making the operating system and apps more memory-efficient as it has before. Other commenters wondered whether strategic pressure could eventually push Apple toward making memory itself. Consumer anecdotes brought the supply shock down to earth, with RAM prices changing ordinary PC repair and return decisions.

Medicare’s new payment model is built for AI. Most of the tech world has no idea

Summary: TechCrunch profiles Medicare ACCESS, a CMS program testing new payment models for care delivery. The article centers on Pair Team, a healthcare company accepted as one of 150 participants after building for underserved patient populations that much of Silicon Valley ignores. Its thesis is that outcome- and efficiency-based reimbursement favors AI-driven operations more than older activity-based models built around visits, RVUs, and logged minutes.

HN Discussion: Commenters pushed back on the headline, saying the model is not built for AI so much as built to lower costs for chronically ill and high-utilization patients. A founder accepted into ACCESS described it as deliberately deflationary: viable for startups, but priced to force efficient delivery. Skeptics worried about cherry-picking easier patients, gaming outcome metrics, or billing Medicare for work not performed.

---

Tech Tools & Projects

Scorched Earth 2000 – Web

Summary: Scorched Earth 2000 is a browser-based HTML/JavaScript port of the classic artillery game, labeled version 1.1 and dated May 11, 2026. The interface exposes familiar controls for wind, ammo, power, angle, tanks, AI opponents, rounds, shops, player statistics, and multiplayer setup. It preserves a DOS-era turn-based tank-combat format while adding web-era affordances such as shareable games, public or private multiplayer, and in-page menus.

HN Discussion: The discussion was mostly nostalgic, with users recalling Scorched Earth as an early encounter with software versions, save-file hacking, and PC gaming culture. Several compared the port with running the original through DOSBox, archive.org, or other browser DOS-game sites. The recurring theme was how plain-text files, adjustable explosions, and local tinkering made the original unusually formative for young programmers.

Show HN: Nibble

Summary: Nibble is a GitHub project whose tagline is “Generating LLVM IR without malloc or external dependencies.” That points to a small, self-contained way to emit LLVM intermediate representation in places where allocation behavior and dependency surfaces matter. The available pack contains GitHub page chrome rather than a detailed README excerpt, so API shape, supported instructions, and integration examples are not visible here.

HN Discussion: The visible HN discussion was very thin: one commenter asked whether the name refers to a snake clone or to a half-byte. There were no supported themes in the provided pack about LLVM compatibility, memory ownership, API design, or practical code-generation use cases.

delta time

Summary: The page title is simply “ḏelta time,” and the pack contains almost no body text. Comments from the author and users identify it as a personal time or lifetime-tracking project, with references to periods, layers, deleting periods, and moving periods between layers. One user connected the concept to the Weber-Fechner law: the idea that experienced time may scale roughly with the logarithm of elapsed life.

HN Discussion: The author appeared in the thread, saying they made the project and were surprised to see it on HN. Feedback combined appreciation with concrete product requests, especially deleting periods and moving them between layers. A more somber note came from a commenter who avoids lifetime trackers because they can trigger depression rather than reflection.

Golden Testing a CAD Library

Summary: Joe Warren explains how he added golden or visual-regression testing to Waterfall-CAD, his Haskell library for programmable CAD. The core problem is that CAD programs output 3D models, which are awkward to test with ordinary assertions. After adding SVG output in 2025, the project could generate vector diagrams directly from Haskell code and compare snapshots, visual diffs, and accepted intentional changes in a workflow borrowed from UI testing.

HN Discussion: The HN pack contained no comments for this story, so there is no concrete community reaction to report. Possible questions about tolerances, deterministic SVG output, geometry kernels, and snapshot maintenance were not present in the provided discussion data.

The limits of Rust, or why you should probably not follow Amazon and Cloudflare

Summary: This article argues that teams should not automatically follow Amazon, Cloudflare, Discord, or other prominent adopters into Rust. From the notes, it appears to cite team familiarity, adoption ranking, async complexity, and release cadence as reasons for caution. The piece treats Rust as a powerful language with limits rather than a universal replacement for Go, Python, Node.js, or existing stacks, and it leaves room for systems-programming use cases.

HN Discussion: Many commenters said the post was really arguing against bad adoption practice: do not choose a language your team cannot maintain. Several corrected release-count comparisons, arguing Rust’s minor releases are not comparable to major breaking releases in other ecosystems. The strongest recurring point was context: Amazon’s low-level infrastructure choices do not automatically apply to every web service or business app.

SQLite Code of Ethics

Summary: SQLite’s Code of Ethics page explains that it began as a “Code of Conduct” mainly to fill a box on supplier-registration forms. The project renamed it after learning that “Code of Conduct” has a more specific meaning for many readers and that SQLite’s document did not match those expectations. Its ethical basis is chapter 4 of the Rule of St. Benedict, which the page calls the “instruments of good works,” applied to developers’ interactions with each other, clients, and the user community.

HN Discussion: The provided HN pack contained no comments for this story. That means there are no supported community themes here about religious framing, open-source governance, supplier questionnaires, or the long-running debate over codes of conduct.

---

Web & Infrastructure

AEPs: API Enhancement Proposals

Summary: AEPs, or API Enhancement Proposals, are a GitHub-hosted set of design guidelines for building consistent network APIs and clients. The repository describes the guidelines as extensible, so organizations can adopt and adapt them rather than treat them as a single fixed product specification. HN comments identify the project as a fork of Google’s AIPs, the API Improvement Proposals used for Google’s public APIs.

HN Discussion: One maintainer appeared in the thread, saying they had worked on AEPs for years and were available to answer questions. Discussion clarified the project’s lineage and terminology, especially that it forks Google’s AIP standard rather than starting from scratch. A minor side theme was acronym collision, with unrelated “AIP” meanings surfacing around the same time.

---

Academic & Research

Avoiding and reducing microplastic false positives from dry glove contact

Summary: This Analytical Methods paper studies a contamination problem in microplastics measurement: false positives caused by dry glove contact. Its abstract frames accurate quantification as necessary for identifying microplastic sources and reducing pollution. The author list spans chemistry, statistics, macromolecular science, environmental programs, and electrical engineering/computer science at the University of Michigan, matching a methodological problem that cuts across lab work and data interpretation.

HN Discussion: The provided HN pack contained no comments, so there are no supported discussion themes to summarize. Any debate about glove materials, lab contamination controls, polymer identification, or microplastics methodology would go beyond the supplied data.

Princeton mandates proctoring for in-person exams, upending 133 year precedent

Summary: The Daily Princetonian reports that Princeton faculty voted to require proctoring for all in-person exams starting July 1, 2026. The policy breaks with a 133-year honor-code precedent under which undergraduates could take exams without faculty or TA monitoring. The article presents the change as one of the most significant alterations to Princeton’s honor system, arriving as phones and multimodal AI tools make covert cheating easier.

HN Discussion: Former Princeton TAs described the old system concretely: hand out exams, leave the room, and return later to collect them. Commenters split between blaming AI-enabled cheating and seeing the policy as part of a broader transition from high-trust to low-trust institutions. Anecdotes about students photographing exams and uploading them to Gemini pushed the conversation toward device confiscation and active proctoring.

Marco Polo: Finding a friend with only distance and motion

Summary: Jack Hogan explains range-only relative localization: finding another moving device when each side knows its own motion and only the distance between them. The post starts with a crowded-cafe analogy, then translates the problem into robotics and state estimation. It surveys multi-antenna ultra-wideband, external trilateration, and Kalman filtering before walking through an extended Kalman filter with state vectors, covariance, transition and measurement functions, linearization, data flow, and a demo.

HN Discussion: Technical commenters suggested the unscented Kalman filter as an easier and conceptually cleaner alternative because it avoids explicit linearization and differentiability requirements. Readers also praised the interactive visuals and overall design. The thread was mostly appreciative, with discussion centered on pedagogy, visualization, and state-estimation implementation choices rather than disputing the project.

---

History & Science

Chess puzzle I found in my dad’s old book

Summary: This page presents an interactive chessboard puzzle attributed to Kempelen and an old book found by the author. The task is to place four black queens and one black bishop so that every square on the board is attacked, leaving no safe square for a white king. The interface lets users drag or tap pieces, check an arrangement, reveal the solution, and reset the board, connecting a composed chess problem to Kempelen’s Mechanical Chess Player lore.

HN Discussion: Commenters quickly turned to enumeration and strategy, with one user finding 388 solutions, including unintuitive layouts with multiple queens on the same row. A rules correction noted that the text should say every square is in check, not necessarily checkmate, because a king may be able to capture an undefended piece. Solving advice emphasized covering nearby squares around queens rather than maximizing long-range lines.

‘A Four-Eyed World’ Review: The Story of Spectacles

Summary: The Wall Street Journal link is a review of “A Four-Eyed World,” a book about the history of spectacles. The guard could not fetch the article body because the site returned HTTP 401 Forbidden, so the available evidence is limited to the title and URL metadata. The title indicates a cultural or historical treatment of eyeglasses rather than a technical optics paper, but the review’s argument and specific examples are not available.

HN Discussion: The provided HN pack contained no comments for this story. There are therefore no supported discussion themes about vision correction, eyewear design, publishing, paywalls, or the review itself.

The Original 1965 Gatorade Recipe

Summary: Eats History recreates the original 1965 Gatorade recipe and tells the origin story of the sports drink. The post traces it to a University of Florida kidney specialist responding to a football coach’s question about why players were not urinating during games. The early formula had six ingredients, including salt, sugar, and lemon juice, and was described as medicinal or unpleasant before becoming a commercial product shaped by rights deals, corporate scale, and athlete endorsements.

HN Discussion: Commenters noted that the post follows the recipe-site pattern of a long story before the recipe, but several found the historical details worth the scroll. A packaging detail stood out: early Gatorade being sold in a flat-top can that needed a key. The taste discussion was mostly humorous, centering on the article’s line that players compared it to urine.

The Age of the Amplifier

Summary: Brian Potter traces how Bell Labs’ need for better signal amplification produced technologies with consequences far beyond telephony. The essay connects the vacuum tube, negative-feedback amplifier, transistor, and laser to the practical Bell System problem of boosting electromagnetic signals across networks. It also situates those breakthroughs within Bell Labs’ wider 20th-century research role, including statistical process control and the Nobel-winning transistor-effect work of Shockley, Bardeen, and Brattain.

HN Discussion: Commenters connected amplifier history to older feedback-control theory, especially James Clerk Maxwell’s 19th-century work on governing steam engines. The thread broadened into how practical industrial needs, from steam power to telephony, can drive foundational science such as control theory and thermodynamics. Readers also shared Bell System historical videos, distinguishing popular and more technical treatments.

Comparing a 1980s memory map to the Raspi Pico

Summary: The Medium article compares a roughly 40-year-old memory map with today’s Raspberry Pi Pico. The guard could not fetch the article body because Medium returned HTTP 403 Forbidden, so the available evidence is limited to the title and URL metadata. The title suggests a retrocomputing comparison between an 1980s address-space layout and the microcontroller-oriented memory organization of the Pico, but specific addresses, ROM/RAM layout, and peripheral details are not supported by the pack.

HN Discussion: The provided HN pack contained no comments for this story. There are no supported discussion themes about microcontroller memory maps, retrocomputing comparisons, or Medium access limitations.