This morning’s Hacker News front page ranges from fast-moving LLM workflows and autonomous media experiments to consumer-device security, surveillance procurement, computing-history memorials, and a run of small developer tools. The shared theme is operational reality: models, browsers, radios, archives, and even book bundles are being judged less by novelty than by cost, privacy, reliability, and the specific tradeoffs users can test.

AI & Tech Policy

The last six months in LLMs in five minutes

Summary: Simon Willison shares annotated PyCon US 2026 lightning-talk slides summarizing six months of LLM change. He frames November 2025 as an inflection point, especially for coding models and rapid shifts in perceived model leadership. The post uses his pelican-on-a-bicycle SVG prompt as a deliberately odd benchmark for comparing model behavior. The excerpt emphasizes how quickly Claude, GPT, and Gemini releases traded the informal ‘best model’ position.

HN Discussion: Commenters questioned whether the claimed inflection point reflects real capability gains or marketing-driven vibes. Several discussion threads centered on coding: barebones app generation versus fully usable software, and whether one-shot game tasks are a meaningful benchmark. Others asked how the same six months looked outside programming, especially for non-programmer collaboration workflows.

Codex-maxxing

Summary: Jason Liu describes moving coding agents beyond repo diffs into broader knowledge work such as Slidev decks, PDFs, spreadsheets, and note-taking. The post argues the newer Codex app feels native for this broader mode because work can live in durable threads with memory and reviewable artifacts. Specific operating-loop pieces include voice input, steering, computer/browser use, remote control, heartbeats, and monitoring feedback. The excerpt positions Codex as both a coding tool and a workspace for persistent agent-driven tasks.

HN Discussion: The thread immediately noted that the author works on OpenAI’s Codex team, which shaped how readers interpreted the advocacy. Some commenters pushed back on the ‘maxxing’ framing and on AI workflow posts appearing in broader technical forums. Practitioners compared concrete multi-agent patterns, including named agents, manager sessions, PR marking, long-running issues, and main-branch landing workflows.

PyTorch Landscape

Summary: PyTorch Landscape is a landscape-style page for the PyTorch ecosystem. The available source text exposes only the page title, so the item is best read as an ecosystem map rather than a documented ranking or review. The excerpt does not show individual project entries, inclusion criteria, or governance details. The useful factual core is that it points readers to a PyTorch-related tools-and-projects catalog.

HN Discussion: The visible HN thread had no substantive comments at capture time. That leaves no concrete reader themes about methodology, inclusion criteria, or comparisons with other ML ecosystem maps.

Pope Leo XIV’s first encyclical Magnifica humanitas to be published May 25

Summary: Vatican News reports that Pope Leo XIV’s first encyclical, titled Magnifica humanitas, is scheduled for publication on May 25. The available article text is dominated by site navigation, so the reliable factual core is the publication announcement and title. HN comments indicate the release event involves several speakers, including Anthropic co-founder Christopher Olah, but not as a co-author. The surrounding discussion reads the encyclical through questions about human value, AI, robotics, and economic change.

HN Discussion: Commenters focused on whether religious leadership can provide a useful language for intrinsic human value amid AI and robotics change. Several nonreligious participants said recent papal statements seemed more constructive than many political visions in a fragmented and authoritarian climate. A correction thread objected that the HN title made Christopher Olah sound like a co-author when he was described as one speaker at the release. Others compared the ambition to Rerum novarum and its influence during the industrial era.

We let AIs run radio stations

Summary: Andon Labs built four AI-run radio stations as a media-sector version of its autonomous-business experiments. Each station is assigned to a different model: Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3. The agents start with $20, buy songs, manage libraries, schedule programming, decide what plays, handle calls, and seek revenue when funds run low. One example in the excerpt has DJ Gemini negotiating a $45 advertising deal with a startup.

HN Discussion: Commenters watched failure modes live, including a Grok-run station stuck repeating a Miles Davis intro with slightly changing delivery. Several readers found the darkly mismatched programming funny, such as historical tragedies paired with ironic song choices. A recurring theme was that the project is an experiment in autonomous media behavior, not a replacement for human-run radio. Some comments suggested distribution tweaks such as opening social accounts for the stations.

LLMCap – A proxy that hard-stops LLM API calls when you hit a dollar cap

Summary: LLMCap is a proxy service that enforces hard spending caps on LLM API calls rather than only sending alerts. The setup changes the API client’s base_url to an LLMCap proxy endpoint and supports providers including Anthropic, OpenAI, Gemini, Mistral, and Cohere. Users can define daily, monthly, or per-key dollar limits, with per-model granularity. When a cap is reached, the proxy returns HTTP 429 before the upstream token is consumed, aiming to prevent surprise bills.

HN Discussion: Commenters questioned whether the core proxy behavior is simple enough to build independently. Privacy was a concrete concern because API traffic and keys route through an intermediary. Pricing drew pushback, with readers wondering whether a subscription fits a narrow cost-control utility.

Agora-1: The Multi-Agent World Model

Summary: Odyssey introduces Agora-1, a multi-agent world model that lets multiple human or AI participants share a generated simulation in real time. The initial demo uses GoldenEye-style multiplayer deathmatch, with up to four players interacting in the same generated world. The model simulates player interactions from actions, maintains a shared world state, and streams generated views to participants. The company frames the work as relevant to gaming, robotics, defense, education, foundation models, and other shared simulated experiences.

HN Discussion: Commenters liked the retro GoldenEye demo but argued that training only on that game makes the concept look visually dated. Technical discussion focused on consistency: how learned world state prevents the map from vanishing on camera turns, and whether caves or spaces remain stable hours later. Several readers asked what exists beyond the demo, whether it is testable, and whether the intended audience is researchers rather than players or developers. The multi-agent angle led to broader comparisons with MARL and the lack of an AlphaGo-like breakthrough for distributed agents.

---

Security & Privacy

Click (2016)

Summary: The linked page is the long-running ‘Click’ interactive site, with the available excerpt exposing only the page title. From the HN context, readers treated it as a browser-based demonstration of how user actions such as clicking and opening tools can be observed. The site appears to react to repeated or automated clicks, including labeling scripted clicking behavior as bot-like. The article text is minimal, so the reliable summary is the interaction and tracking demonstration rather than undocumented implementation detail.

HN Discussion: Commenters connected the page to web analytics and session-replay tools that record mouse movements and allow later playback. Privacy discomfort was concrete: one story described a user feeling exposed when the site owner could see that dev tools had been opened. Other comments compared it to older HN discussions of the same project and tested how it handled scripted console clicks.

Anyone on the Internet Can Ring Your Doorbell

Summary: The author reverse-engineered a low-cost Smart Doorbell X3 sold through Temu and paired through the X Smart Home app. The findings include account takeover of doorbells, device impersonation during live calls with attacker-controlled video, and exposure of the home Wi-Fi password through a debug port. The post says the author opened a CERT/CC VINCE coordination case and expected CVE assignment through that process. Naxclow contacted the researcher one day after publication, acknowledged the report, and began an internal review.

HN Discussion: Commenters contrasted internet-exposed smart doorbells with simple hardwired buttons and transformers, using the old design as a security baseline. Several remarks focused on physical reset and ownership-transfer flows, including QR-code provisioning and whether a passerby could rebind a device. The thread also broadened into distrust of consumer IoT security, with Ring and other smart home devices used as comparison points.

The FBI Wants to Buy Nationwide Access to License Plate Readers

Summary: 404 Media reports that the FBI wants to buy nationwide access to automated license plate reader data. Procurement records reviewed by the outlet suggest such access could let the agency track vehicle movements across the country without a warrant. The article says only a few vendors, specifically naming Flock and Motorola, could likely meet the requested nationwide scope. The story arrives amid protests and local pushback against ALPR deployments around the United States.

HN Discussion: Commenters called for legal separation between mass surveillance databases and government access, while doubting either major party would enact it. A privacy-economics theme argued that personal data must become a liability rather than an asset before incentives change. Some discussion turned practical, noting plate masking, paper plates, removed plates, and altered plates as already common evasions in parts of Southern California. One proposal was rotating digital license plates that preserve police registration checks while limiting long-term mass tracking.

Show HN: Clark-Browser – Stealth Chromium

Summary: Clark-Browser is a Show HN GitHub project described as Stealth Chromium. The repository tagline visible in the excerpt is ‘Browser that finds a way,’ though the rest of the excerpt is mostly GitHub navigation. The supported description is a Chromium-based browser project focused on stealth or detection avoidance. The available source text does not provide enough detail to say which browser fingerprinting surfaces, automation APIs, or anti-bot systems it targets.

HN Discussion: The visible discussion asked how Clark-Browser differs from Camoufox, another browser project associated with anti-fingerprinting or stealth use cases. That comparison suggests readers wanted concrete differentiation rather than a broad stealth-browser label. No other substantive comments were visible, so the thread does not support a broader debate about scraping, privacy, or abuse.

The Futility of Lava Lamps: What Random Means

Summary: Loup Vaillant argues that Cloudflare’s lava-lamp entropy wall is mostly marketing and security theater. The post says the visual setups imply a meaningful contribution to internet encryption, but are worse than mundane entropy sources and likely not significant to security. It uses the classic ‘return 4’ random-number joke to motivate a deeper discussion of what randomness means. The core claim is conceptual: unpredictability, probability, and cryptographic usefulness are not the same as a striking physical display.

HN Discussion: Commenters pushed back that Cloudflare has described using Lavarand in production, while also mixing in sensor noise and conventional entropy sources. Several readers treated the lamps as harmless communication: a memorable way to explain entropy to non-specialists rather than a sole security mechanism. Cryptography-focused comments argued the lamps probably do little but also likely do not weaken a competent CSPRNG if mixed into Linux’s RNG. The main tradeoff was theater versus education, not whether lava lamps alone should secure production systems.

Sieve – scans Cursor/Claude chat history for leaked API keys

Summary: Sieve Secret Scanner is a Mac App Store developer tool for finding secrets leaked into AI coding assistant histories. The app claims fully local scanning for API keys, tokens, passwords, and private keys in Claude Code, Cursor, VS Code Copilot, VS Code Insiders, Windsurf, Codex, and project .env files. It also includes a vault backed by macOS Keychain for storing rotated secret values, with Touch ID or login-password checks before copying. The listing says it ships with a local MCP server, suggesting integration with agent or assistant workflows.

HN Discussion: The visible comment connected the tool to a real workflow problem: LLM-heavy infrastructure work with pyinfra secrets and .env login files. Secret-management alternatives came up, especially using SOPS with age from the beginning rather than cleaning up leaks afterward. The thread is thin, but the concrete theme is that agent-assisted operations make accidental prompt or history exposure more likely.

---

Tech Tools & Projects

Turn your Android phone into a ham radio transceiver

Summary: kv4p HT is an open-source hardware and software project that turns an Android phone into a handheld ham radio transceiver. The device plugs into USB-C, uses a small VHF or UHF radio board, and relies on the phone for screen, GPS, battery, and controls. The project advertises 1 watt output, APRS support with a 1200 baud modem for text-style messaging and position beaconing, and accessibility features such as live captions and sticky PTT. It is a build-it-yourself amateur radio project requiring at least a Technician class license, with GPL3 app, firmware, PCB, and 3D-print files.

HN Discussion: The visible discussion focused on radio power tradeoffs, with a commenter comparing the 1 watt design to inexpensive handheld radios advertised at higher wattage. That comparison frames the project less as a Baofeng replacement and more as an open, phone-integrated maker device. The thread is thin, so it does not add much evidence about reception, range, or reliability.

Regex Chess: A 2-ply minimax chess engine in 84,688 regular expressions

Summary: Nicholas Carlini built a chess engine as a sequence of 84,688 regular expressions applied in order to a board string. The project accepts coordinate moves such as e2e4, updates a FEN-like board state, and chooses a valid response using a 2-ply minimax search. The implementation is intentionally impractical: the visible program loop just applies a large regex list, while the chess logic is encoded inside the generated substitutions. The excerpt stresses that the engine is valid and ‘not entirely terrible,’ but constrained by the shallow search depth and unusual representation.

HN Discussion: Commenters tested concrete games and found tactical weaknesses, including a line where the engine misses checkmate threats. Some discussion focused on formal-language curiosity: NFAs, regular expressions as computation, and the surprise of seeing textbook automata ideas pushed into a working artifact. Readers compared the work to other deliberately impractical but technically revealing programming projects.

When can the C++ compiler devirtualize a call?

Summary: Arthur O’Dwyer investigates when C++ compilers can replace virtual dispatch with direct calls. The article says modern compilers reliably devirtualize calls to final methods, while many corner cases differ by compiler. It deliberately excludes link-time optimization and focuses on what the compiler can prove locally. The two core cases in the excerpt are knowing the object’s dynamic type and other situations where the compiler has enough type information to avoid a vtable call.

HN Discussion: Commenters supplied failure stories from real optimizers, including Clang speculative devirtualization interacting badly with BOLT identical code folding. MSVC’s inter-procedural devirtualization was discussed as powerful but reportedly slow and buggy enough to be disabled. The thread compared C++ virtual calls with Rust dyn trait devirtualization and whether MIR or LLVM can prove the concrete type. A practical theme was fragility: some readers advised avoiding virtual dispatch in performance-critical paths when the call shape must be guaranteed.

Hyperpolyglot Lisp: Common Lisp, Racket, Clojure, Emacs Lisp

Summary: Hyperpolyglot’s Lisp page is a side-by-side reference comparing Common Lisp, Racket, Clojure, and Emacs Lisp. The excerpt shows sections for grammar, execution, variables, arithmetic, strings, regexes, dates, lists, arrays, dictionaries, functions, exceptions, streams, files, processes, objects, macros, reflection, and Java interop. It lists concrete implementation versions such as SBCL 1.2, Racket 6.1, Clojure 1.6, and Emacs 24.5. The page is meant as a syntax and tooling comparison rather than a tutorial for any single Lisp dialect.

HN Discussion: Commenters checked idiomatic accuracy, objecting to examples that used eval where apply or recursive structure would fit better. Common Lisp readers corrected details about SBCL compilation, REPL behavior, compile-file, documentation lookup, and whether code is interpreted. The thread also surfaced adjacent cross-language cheat sheets, including an Elisp reference aimed at Python programmers.

Show HN: Number Gacha, a gacha game distilled to its essence

Summary: Number Gacha is a small web game presented as a stripped-down version of the gacha loop. The article excerpt exposes only the title, but comments describe rolling, buying packs, slicing packs open, completing a collection, and fighting battles. Instead of real-money microtransactions, players can solve math problems to unlock more rolls. The game appears intentionally focused on the collection and reward mechanics that define the genre.

HN Discussion: Commenters praised the polish and said the pack-opening interaction felt satisfying. Several players wanted more game telemetry, such as math-problem counts, resources spent on rolls and packs, and losses in battles. Usability feedback centered on the battle system, including requests for hotkeys or numeric controls for blocking, healing, and other actions. The microtransaction satire landed clearly, with jokes about the author refusing to sell gems.

Show HN: Hsrs – Type-Safe Haskell Bindings Generator for Rust

Summary: Hsrs is a GitHub project described as a type-safe bindings generator between Haskell and Rust. The repository title says it provides type-safe Haskell Rust bindings, though the available excerpt is mostly GitHub navigation. The reliable claim is the project purpose, not detailed API behavior, supported Rust features, or generated-code design. As a Show HN item, it is a new developer tool for cross-language interoperability.

HN Discussion: The visible HN thread had no substantive comments at capture time. There were therefore no concrete reader themes about FFI safety, type mapping, build integration, or comparison with existing Haskell/Rust interop tools.

Make ZIP files smaller with ZIP Shrinker

Summary: Evan Hahn released ZIP Shrinker, a browser-based tool for reducing ZIP archive size. The tool also works on formats that use ZIP internally, including APK, EPUB, JAR, and related package formats. Its main operations are recompressing every archive entry with higher Deflate compression, stripping ZIP metadata, and removing explicit directory entries. The implementation uses libdeflate through a new WebAssembly wrapper called libdeflate.js, choosing speed over slightly denser but slower alternatives such as Zopfli.

HN Discussion: The visible HN thread had no substantive comments at capture time. There were no concrete reader themes about browser-side privacy, compatibility risks from removing directory entries, or compression tradeoffs. The article itself does make one design tradeoff explicit: libdeflate is faster while Zopfli may produce marginally smaller output.

Loopmaster – Livecoding Music IDE

Summary: Loopmaster is presented as a livecoding music IDE, with the available page text exposing only the site name. The title indicates a tool for creating or performing music through live code editing. The source text does not show specific syntax, audio engine choices, collaboration features, or export formats. The strongest supported framing is that it joins the broader ecosystem of browser or desktop environments for code-driven music performance.

HN Discussion: Commenters immediately compared it to Strudel, praising inline piano rolls, waveform visualizers, note highlighting, and album-making workflows. Pure Data came up as a mature visual programming environment used for art, hardware interfaces, assistive instruments, and electroacoustic music. Orca was also mentioned as a two-dimensional esoteric programming language for musical sequencing. The discussion theme is ecosystem comparison more than critique of Loopmaster itself.

Coding on Paper

Summary: Oskar Wickström writes about using an Onyx BOOX 25.3-inch Mira Pro Color e-ink monitor as his primary programming display. The post is framed as an experience report from a working programmer, not a sponsored review or broad recommendation. The motivation is environmental and ergonomic: natural light, morning sunlight, garden work, and a setup that makes him feel clear and focused. He says the monitor costs about $2000, differs substantially from LCD, and extends earlier experiments with e-ink tablets and Termux into a regular NixOS workflow.

HN Discussion: Several commenters expected the title to mean writing programs on literal paper, recalling paper-based programming exams and pre-keyboard learning habits. That mismatch produced title criticism from readers who felt the post was really about a paper-like screen. Other discussion focused on availability and affordability of large e-ink monitors, with interest in more products as the category matures. The practical theme was whether the benefits of natural-light readability justify the cost and LCD tradeoffs.

Humble Tech Book Bundle: The Ultimate Functional Programming

Summary: The item links to a Humble Bundle book sale titled The Ultimate Functional Programming. The URL identifies the publisher/source collection as Pragmatic Programmers books. The available page text is mostly client-side JavaScript rather than readable bundle details, so specific titles, pricing tiers, and charities are not visible. The safe framing is a limited-time technical book bundle focused on functional programming.

HN Discussion: The visible HN thread had no substantive comments at capture time. There were no concrete reader themes about book quality, DRM, pricing tiers, charity split, or publisher comparisons.

---

History & Science

Peter Neumann has died

Summary: The TUHS mailing-list post forwards news from the Multics community that Peter Neumann died in a Santa Clara hospital. Robert Watson reported that Neumann died in his sleep after complications from a fall and subsequent surgery. The note says his daughter Hellie was with him and that they had been listening to classical music together. The message also highlights that, beyond computer security, Neumann loved music and played piano, French horn, and other instruments.

HN Discussion: The visible HN discussion contains a pointer to an earlier HN item on the same news, hosted through LWN and the same mailing-list source. Because the thread is sparse, there are no supported technical debates or detailed remembrances here. The discussion supports treating this as a computing-history notice rather than inferring broader community reaction.

Peter Salus has died

Summary: A TUHS post by Dan Cross reports that Peter Salus died on May 15. The message points to Salus’s A Quarter Century of Unix as required reading for serious students of Unix history. The post is short and functions as a community notice rather than a full obituary. HN context adds that Salus also served as executive director of USENIX in its early years.

HN Discussion: Commenters discussed Salus’s role in making the AT&T to BSD to Linux throughline understandable for readers learning Unix history. One thread asked which of his books still hold up as essential and whether comparable computing-history work is being done today. Other comments supplied biographical links and highlighted his early USENIX leadership.

Why is it called Kent House?

Summary: Diamond Geezer investigates why Kent House station in South London has that name. The station sits between Victoria and Bromley South, with Penge East and Beckenham Junction nearby, and opened in 1884. The article traces the name to historical county boundaries: Penge was on the Surrey side, Beckenham on the Kent side, and the Kent boundary ran close to the present station platforms. It also notes that the 1889 creation of the County of London changed the local administrative boundary by absorbing northern Surrey, including Penge.

HN Discussion: The visible HN discussion is local and meta: a commenter found it odd to see a familiar London place on HN. The same commenter pointed to Diamond Geezer’s traffic milestone post, noting that HN was one of the largest referrers despite linking only occasionally. There were no detailed counterarguments about the station-name evidence or boundary history.

Earth’s Radio Bubble: Every signal we’ve ever sent into space

Summary: The article explains Earth’s expanding radio bubble: electromagnetic signals from human technology moving outward at light speed. It says the bubble began with powerful radio transmissions in the early 1900s and is now roughly 240 light-years across. The piece frames that as both large and tiny, noting that it covers only about 0.000002 percent of the Milky Way’s roughly 100,000-light-year diameter. Examples include music, TV broadcasts, radar pings, and deliberate messages sent into space.

HN Discussion: Commenters challenged the practical detectability of ordinary broadcasts, citing a thought experiment where even Arecibo could not detect Earth’s omnidirectional FM radio as far away as Saturn. A JPL anecdote made the same point for Alpha Centauri: current Earth technology likely could not detect Earth-like leakage from there. Readers connected the topic to Contact and the idea of early TV broadcasts as narrative signals. One thread noted niche intentional transmission services, including claims of modulated laser broadcasts into space.

---

Academic & Research

Alignment pretraining: AI discourse creates self-fulfilling (mis)alignment

Summary: The arXiv paper studies whether discourse about AI systems inside pretraining corpora can causally affect downstream alignment behavior. Its premise is that negative descriptions of AI behavior may be internalized by language models as behavioral priors, creating self-fulfilling misalignment. The paper is titled Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment and was submitted in January 2026 with a February revision. Authors listed in the excerpt include Cameron Tice, Puria Radmard, Samuel Ratnam, Andy Kim, David Africa, and Kyle O’Brien.

HN Discussion: Commenters connected the mechanism to hyperstition: writing about future AI behavior can help make that behavior more likely if it enters training data. One reader highlighted a reported tradeoff where better alignment from pretraining came with about a 4 percent average degradation in test performance. The thread debated data filtering, with some arguing labs can remove documents about AI misalignment rather than treating the effect as unavoidable. Several comments joked that the first rule of alignment may be not writing alignment discourse into future training corpora.

---

Business & Industry

AI eats the world (Spring 26) [pdf]

Summary: The linked item is a PDF deck titled AI eats the world for Spring 2026. HN comments identify it as part of Benedict Evans’s AI presentation series and point to prior versions of the deck. Discussion summaries describe the deck as tracking AI from uncertainty over scaling and usefulness toward questions of deployment, commoditization, and business model. Commenters also quote the deck’s framing of prior platform shifts: hardware, internet, mobile, cloud, and now AI.

HN Discussion: Readers valued the temporal comparison across versions of the deck, treating it as a record of how AI strategy thinking has shifted since late 2024. Technical criticism focused on whether large frontier models are hiding major inefficiency, with comparisons to a mainframe era before more efficient designs emerge. Business-model debate centered on whether LLM labs resemble telecom carriers, cloud platforms, neoclouds, or something with stronger differentiation. Some commenters debated the inevitability of AI as a platform shift by comparing it to hardware, internet, mobile, and cloud cycles.

Apple Silicon costs less than OpenRouter

Summary: The linked X post itself mostly exposes a pointer to a longer article, but the author summarized the claim in the HN comments. The claim is that, accounting for batching, cache, input tokens, and residual MacBook Pro value, local Apple Silicon inference can be about 14 percent cheaper than OpenRouter. The author further claims the advantage can reach roughly 3x, or 65 percent cheaper, for MoE models such as Gemma 4 26B. A visible thread correction says the title should read that Apple Silicon costs less than OpenRouter, not the reverse.

HN Discussion: Commenters debated whether lower cost should be expected, with one comparing Apple Silicon favorably to EC2 for general compute. Several comments focused on title accuracy and confusion caused by a reversed or altered title. The author clarified assumptions directly in the thread, emphasizing batching, cache behavior, input-token costs, and residual hardware value. A related HN discussion from the prior day was linked, suggesting this cost comparison was already being debated.