Welcome to today’s HN morning brief. Here are the top stories shaping the conversation across security, AI policy, infrastructure, and the long tail of interesting tech history.

AI & Tech Policy

Evals Will Break and You Won’t See It Coming

Summary: Lun Wang’s blog post argues that evaluation infrastructure is the true bottleneck for the next LLM capability jump. Standard benchmarks and safety evals implicitly assume the next model is a stronger version of the current one, but when models cross into new capability regimes — through emergent abilities or grokking-like transitions — the entire evaluation infrastructure breaks silently. Wang claims eval development lags behind advances in training, architecture, and data.

HN Discussion: Commenters pushed back on the framing: evals test boundaries and steer LLMs toward specific goals rather than predict capabilities in advance. One concern was raised about conflicting evals — whether desirable emergent behaviors get accidentally penalized by benchmark designs that optimize for different axes. Several noted that benchmarks are nearing saturation at current model sizes, making further evaluation refinement increasingly difficult.

Skills in Web, iOS, and Android

Summary: xAI launched “Skills” for Grok — persistent expertise that carries across conversations on web, iOS, and Android. Built-in skills generate Word documents, presentations, spreadsheets, and PDFs with full formatting. A custom Skill Creator lets users define their own capabilities through conversation, with a priority override system ensuring personal skills take precedence over xAI defaults.

HN Discussion: No HN comments were posted for this product announcement.

Testing MiniMax M2.7 via API on Three Real ML and Coding Workflows

Summary: Andrey Lukyanenko ran MiniMax M2.7 through Claude Code’s agentic loop against three concrete workflows: scaffolding a Kaggle competition entry, drafting and auditing knowledge-base notes for an Obsidian vault, and updating an old PyTorch project. Compared to Claude Opus 4.7 as baseline, M2.7 performed well when constraints were explicit but stumbled on implicit context — some of the same gaps appeared in Opus too. Runs cost $40/month on MiniMax’s Plus tier.

HN Discussion: No HN comments were posted for this workflow comparison.

AI, “Humanity”, and Dr. Manhattan Syndrome: A Communications Intervention

Summary: Jim Prosser’s essay critiques how AI leaders invoke Capital-H “Humanity” while making specific partisan political donations. He points to Greg Brockman’s $25 million MAGA donation justified as serving humanity at large, arguing that abstract, floating-above-the-fray rhetoric obscures specific policy impacts on real people — those with healthcare anxieties, job displacement fears, or differing political views.

HN Discussion: Commenters noted that CEOs face multiple audiences: consumers vs. investors often have conflicting messages. One shared a Chesterton quote satirizing “philanthropy” as love of anthropoids rather than humans. Another critiqued how idealists creating utopias historically lead to the worst crimes, leaving pragmatists to clean up.

Architect of the UK Online Safety Act Calls for Its Complete Repeal

Summary: Former Culture Secretary Nadine Dorries — who formally introduced and championed the UK Online Safety Act in Parliament — has called for its repeal “in its entirety.” The shift followed TikTok using the OSA framework to censor Reform Party Shadow Home Secretary Zia Yusuf over a post on immigration policy. Dorries appeared on GB News arguing the act had gone too far.

HN Discussion: Commenters were deeply skeptical: some pointed out this follows the familiar pattern of voting for something and complaining when leopards eat your face. Preston Byrne’s counter-argument was that he may be pivoting to promote his think tank’s white paper advocating US-style free speech standards for the UK, making Dorries’ statement a convenient vehicle. Questions remained about whether any replacement legislation could work better than the current OSA.

The Last Six Months in LLMs in Five Minutes

Summary: Simon Willison’s PyCon US 2026 lightning talk condenses LLM developments since what he calls the November 2025 inflection point — a critical month for coding models where the “best” model title changed hands five times between Claude Sonnet 4.5, GPT-5.1, Gemini 3, and others. He uses his pelican-riding-bicycle SVG test to illustrate generation differences, noting zero AI labs would train on such an absurd task.

HN Discussion: Commenters questioned whether the “inflection point” is real or marketing framing. One user reported coding agents still struggle to produce production-quality applications despite tool-calling improvements. Another noted they’re excellent for answering questions about large codebases but haven’t solved the actual problem of generating shippable code without heavy steering and babysitting. A few pointed out the pelican test originated in Microsoft’s original GPT-4 report three years ago.

Security & Privacy

GitHub Compromised

Summary: A Twitter/X post reported unauthorized access to GitHub’s internal repositories triggered by a poisoned VS Code extension on an employee device. The malicious extension version was removed from distribution. Attackers claimed access to approximately 3,800 internal repositories — a figure GitHub said was “directionally consistent” with their own investigation. No evidence yet of impact on customer enterprises, organizations, or public repositories.

HN Discussion: HN discussion focused on whether attackers were trying to help fix availability problems or truly exfiltrating data. Commenters asked which specific extension was compromised and whether there’s visibility into what the poisoned payload actually did.

GitHub Is Investigating Unauthorized Access to Their Internal Repositories

Summary: GitHub announced an investigation of unauthorized access to its internal repositories through a Twitter/X post, with no official blog announcement or status page update. The company stated they have “no evidence of impact to customer information stored outside of GitHub’s internal repositories” and are monitoring for follow-on compromise activity.

HN Discussion: The primary thread of discussion centered on why Twitter/X was chosen as the security incident channel — commenters noted the absence from github.blog and githubstatus.com. Some debated whether the attackers’ motive was constructive (trying to fix outage issues) or malicious exfiltration. A detailed follow-up tweet claimed ~3,800 repositories were accessed.

Remove-AI-Watermarks – CLI and Library for Removing AI Watermarks from Images

Summary: A GitHub repo offers a CLI tool and library designed to remove visible watermarks (Gemini) and invisible ones (SynthID, C2PA, EXIF) from images — targeting both visual overlays and embedded provenance metadata. It’s an open-source project for stripping AI-generated content identifiers.

HN Discussion: Commenters split on the ethics: some argued watermarks erode trust by forcing barcoding of every digital move, aligning with a privacy-focused hacker ethos. Others defended watermarks as necessary tools to protect against misuse of copyrighted training data and to provide markers that help people identify AI output. One commenter noted SynthID and C2PA are problematic precisely because they embed identifiers into images generated from stolen copyrighted material.

System Administration

Railway Blocked by Google Cloud

Summary: Railway experienced a major outage triggered by issues on Google Cloud Platform, with workloads hosted on GCP continuing to experience intermittent recovery problems. Multiple data centers across US East, US West, and EU West were affected. The status page showed gradual workload restoration but warned that intermittency could continue during the full recovery.

HN Discussion: Commenters blamed both Railway’s incident handling and GCP’s recurring reputation for sudden service disruptions. One drew a parallel to the UniSuper outage of May 2024, when Google had to issue a joint public apology. Several expressed concern about automated account actions taken without human contact escalation — one commenter asked if this was “some AI automation gone wrong” because GCP seems allergic to contacting humans for responses.

The TTY Demystified (2008)

Summary: Linus Åkesson’s classic 2008 explanation of the Linux TTY subsystem covers everything from real teletypes in the 1940s through modern kernel terminal handling. It explains how LF-to-CRLF conversion, line editing, and session management are split across the terminal itself, the kernel, and the application code — with serial line driver complications layered on top.

HN Discussion: Long-time readers credit the page as one of the best deep dives into TTY internals available online. A recent HN discussion prompted its re-sharing. Commenters criticized the awkward design where terminal, kernel, and application concerns are poorly separated. One noted the redundant “index.php” in the URL.

Tech Tools & Projects

The Mercury Logic Programming System

Summary: The Mercury logic programming system by Zoltan Fuzezesgyörgy has surfaced on HN after years of quiet existence. Described as a modern alternative to Prolog with full type inference and pattern matching, the last release was in 2023. The repository contains files dating back 32 years, raising questions about current development activity.

HN Discussion: Commenters expressed nostalgia — one remembered being taught Mercury at University of Melbourne. Others noted it’s “effectively dead” given the stale dates and lamented that no modern alternative to Prolog has emerged. Several asked whether there was a reason for the sudden reposting.

Gemini CLI Will Stop Working from June 18, 2026

Summary: Google announced that Gemini CLI — an open-source terminal AI assistant under Apache 2 — will stop receiving updates after June 18, 2026. Development effort is shifting to “Google Antigravity,” described as an agent-first development platform for multi-agent workflows. The Antigravity CLI repository appears proprietary rather than open-source.

HN Discussion: Simon Willison highlighted that Gemini CLI was Apache 2 while Antigravity appears closed-source. Community frustration focused on Google’s pattern of discontinuing publicly-used tools through internal reorganizations. Early impressions described Antigravity as buggy, with one user reporting it couldn’t use mDNS despite having network access enabled in its sandbox — a fundamental oversight suggesting rushed development.

Polypad

Summary: Mathigon’s Polypad offers virtual manipulatives for mathematical exploration: fraction bars, 3D polyhedra, balance scales, function machines, dice, coins, spinners, and data science tools. The platform is free to use with teacher features including real-time student work viewing. Behind the scenes, it’s built without traditional frontend frameworks — using jQuery-style DOM wrappers combined with Vue-like reactivity and web components.

HN Discussion: The annual Polypad Art and Music Contest draws creative submissions. A former Amplify employee shared tangram project experience implementing multi-solution detection. Commenters appreciated the no-framework approach but noted Polypad itself isn’t open source, even though its underlying libraries are published under the Mathigon org.

Fixing the Most Dangerous Dam in the World

Summary: Wesley Crump’s Practical Engineering examines Mosul Dam in northern Iraq — one of the tallest dams in the Middle East at 370 feet above the Tigris River. Built in the 1980s on geological formations that dissolve when reservoir water reaches them, the dam has faced seepage issues since filling: within a year it was leaking 200 gallons per second (one Olympic pool every hour). Forty years of engineering solutions have been complicated by Iraqi politics, bureaucracy, and armed conflict.

HN Discussion: The Practical Engineering channel received praise for its accessible explanation style, compared favorably to Technological Connections. Viewers speculated about future coverage of the Aswan Dam and Three Gorges Dam. One comment reflected general enthusiasm rather than substantive technical debate.

Kv4p HT – A Homebrew 1W Radio (VHF or UHF) that Plugs into an Android Phone

Summary: The kv4p HT is an open-source 1W VHF/UHF ham radio that plugs into Android phones via USB-C, transforming them into fully-featured handheld transceivers. GPL3 licensed with the Android app, ESP32 firmware, PCB designs, and 3D printer files all public. Features built-in APRS for SMS-style text messaging and position beaconing. Draws power from the phone — no internal battery.

HN Discussion: Commenters noted that cheap Chinese shanzhai phones with PTT functions have existed for a decade but remain closed-source. A previous HN thread from October 2024 drew 191 comments, showing sustained community interest. Several discussed analog’s decline in favor of DMR repeaters and asked whether KiCad schematics should be exported as PDF for easier review without opening the CAD software.

Java: Rethink Domain Primitives with Valhalla

Summary: Davide Angelocola argues that Project Valhalla’s value classes finally make domain-safe type primitives practical in real Java systems without performance penalty. Normally, a wrapper class like PositiveInt costs 16 bytes on HotSpot — four times the memory of the int it wraps — plus each becomes a heap object causing cache misses and GC overhead. Value classes flatten wrappers into array slots with no header or indirection. Experiments run on OpenJDK 27 JEP 401 early access.

HN Discussion: The author solicited feedback from Valhalla experimenters on modeling domain types in Java. The practical argument centers on eliminating the historical performance tax on tiny domain objects while maintaining strong compile-time invariants. Production readiness timeline remains uncertain with the early access build.

History & Science

FiveThirtyEight Articles on the Internet Archive

Summary: Ben Welsh created a searchable index of archived FiveThirtyEight articles via the Internet Archive, preserving over 21,000 pages from fivethirtyeight.com dating back to March 2008. It’s browseable by year, byline author, and headline. The project emerged after Disney shut down FTE and removed its articles from the web entirely — Nate Silver himself wrote about the loss.

HN Discussion: Commenters noted that many important interactive visualizations are broken or missing in the archived pages, including gun deaths data and P-hacking explorers. Still, the preservation work is appreciated despite these gaps. One commenter reminisced about enjoying FTE’s Lionel Messi piece as a non-soccer fan. Others reminded readers why this matters — thousands of articles seemingly vanished from the internet.

In 1979 Engineer Hugh Padgham Discovered “Gated Reverb” – By Accident

Summary: Hugh Padgham discovered gated reverb while working on Peter Gabriel’s third solo album in 1979. The technique involves placing a noise gate on a reverb return to cut the decay short, producing bright, crisp, punchy drum sounds that became quintessentially ’80s. Examples include Prince’s “Kiss” (1986) and Phil Collins’ signature drum sound — the antithesis of the dry drums from the 1970s.

HN Discussion: No HN comments were posted for this article about recording technique history.

Japan Is Gripped by Mass Allergies. A 1950s Project Is to Blame

Summary: Japan’s hay fever crisis traces back to a post-war reforestation project that planted only two tree species across vast areas. Every spring, massive pollen releases cause symptoms in an estimated 43% of the Japanese population — compared to 26% in the UK and 12-18% in the US. Viral videos showed waves of what looked like smoke blowing off evergreen forests but were actually pollen clouds. Climate change is making allergy season arrive earlier each year.

HN Discussion: No HN comments were posted for this BBC Future article about a public health crisis rooted in mid-century policy decisions.

The Two Oldest Printing Presses

Summary: The Museum Plantin-Moretus in Antwerp houses the world’s two oldest printing presses, dating from around 1600. Christoffel Plantin’s workshop employed 56 people across at least 16 presses by 1575 and was the largest of its kind in the world. The presses produced 1,250 sheets per day during 14-hour shifts — printers were paid by output. Today seven presses remain in the printing room, five still operational. The museum holds UNESCO World Heritage status unusual for a functioning business site.

HN Discussion: Visitors described the museum as exceptional, with one noting they could operate a real press and had their print framed at home. Historical linguists discussed how the printing press froze written German before natural declensional simplification could occur. A commenter asked whether any extant movable-type presses survive from Bi Sheng’s 1040 invention in China.

Academic & Research

Growing Neural Cellular Automata

Summary: Distill’s 2020 paper presents Growing Neural Cellular Automata — a differentiable model of morphogenesis developed by Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin. Models were trained to generate patterns that can persist over time, regenerate from damage (especially when subject to pattern damages during training), and rotate their perceptual field without changing the output pattern. The interactive demo lets users erase parts of a pattern and watch it regenerate.

HN Discussion: Commenters expressed nostalgia for Distill as an inspirational ML learning resource, particularly praising “The Building Blocks of Interpretability.” One drew a connection to ongoing research using neural cellular automata for pretraining token generation (arXiv 2026). Biological parallels were cited — electric fields acting as signaling mechanisms for axolotl cell differentiation.

Lisp in Web-Based Applications (2001)

Summary: Paul Graham’s 2001 BBN Labs talk excerpt argues that server-side deployment removes the OS-language coupling constraint that historically tied desktop software to its operating system’s language. With web-based applications, developers control both the server and the compiler, enabling them to write in any language — including Lisp. The freedom is double-edged: more choice means more thinking required, but also greater leverage.

HN Discussion: Comments compared PG’s BBN talk redundancy with his published essays online. Several discussed modern Lisp web stacks (HTMX, Datastar, Mito ORM) as contemporary relevance. One contributor shared experience building large systems in Scheme that outlived competitors despite smaller team sizes.

Programming as Theory Building (1985) [PDF]

Summary: Per Naur’s 1985 paper argues that programming is fundamentally a scientific activity of constructing and refining theories about computation, not merely writing instructions for a machine. The original PDF hosted on Gwern.net resurfaced on HN after years of quiet availability. Naur, a Turing Award winner, frames program development as iterative theory-building rather than engineering implementation.

HN Discussion: Readers found the paper’s arguments applicable beyond software engineering, with one discovering Gilbert Ryle’s “The Concept of Mind” sparked by Naur’s philosophical references. The single comment was appreciative rather than debating — calling it a “gem.”

Geopolitics & War

India’s Hottest District Shuts at 10 AM as Mercury Breaches 48°C Mark

Summary: Banda district in Uttar Pradesh hit temperatures of 48°C, forcing schools and businesses to shut down by 10 every morning. Sustained extreme readings place it among India’s most heat-stressed locations, competing with Rajasthan’s Churu and Jaisalmer. Forest loss and climate crisis factors compound the situation in the Bundelkhand region. Electrical infrastructure strain led to measures including pouring water over transformers to keep them cool.

HN Discussion: Commenters referenced Kazuo Ishiguro’s “The Ministry for the Future” as speculative fiction about where this is headed. One raised wet-bulb temperature concerns, noting the approaching threshold for mass-casualty climate events. UK commenters noted their infrastructure remains unprepared even for more modest heat spikes compared to India’s extremes.

Even by Trumpian Standards, a $1.8B Fund for Friends Is Bad

Summary: An Economist article criticized what it describes as a $1.8 billion fund directed toward political allies — extreme even by Trump administration standards of cronyism. The full article could not be fetched (HTTP 403), so discussion content is limited to the headline and source publication’s framing.

HN Discussion: No HN comments were available due to the article fetch failure.

Web & Infrastructure

HTML-in-Canvas Demos

Summary: Google Chrome Labs released demos showing native rendering of HTML elements inside a canvas context — effectively enabling standard HTML pages rendered within <canvas> elements rather than the DOM. The feature currently only works in Chrome with the experimental flag chrome://flags/#canvas-draw-element. Safari returned an explicit “HTML IN CANVAS NOT SUPPORTED” message when tested.

HN Discussion: Commenters drew comparisons to Flash, questioning whether this is genuine web innovation or a Google extension play (Embrace, Extend, Extinguish). The Chrome flag requirement severely limits practical utility, and the nested HTML-in-Canvas-in-HTML architecture prompted jokes about being stuck in a tech recursion. Safari’s rejection highlights fragmentation concerns for any emerging standard that requires per-browser flags.

Other

Tool Mapping 90 Companies in the Photonics and CPO Supply Chain

Summary: An interactive map covers 90 companies in the Co-Packaged Optics ecosystem, visualizing silicon photonics, InP lasers, hybrid bonding, optical switches, ASICs, and fiber supply nodes. Layers reveal demand drivers, raw element monopolies, and the relationships between the silicon and III-V/InP manufacturing chains. Built as a field guide for tracking where capital flows in photonic infrastructure.

HN Discussion: The only HN comment criticized the UI: on large screens, interactive tooltips cover the underlying map with no scroll mechanism to recover visibility. Otherwise, technical interest centered on understanding supply chain concentration and monopoly positions within the CPO market as data center bandwidth demands intensify.

Unusual Uses of OEIS Sequences on GitHub

Summary: Jeremy Kun discovered OEIS sequences used in two live-coding music frameworks — mercury (which uses Fibonacci, Pell, Lucas, and Fibonacci-like sequences to modulate parameters) and ziffers, a Sonic Pi extension incorporating de Bruijn sequences, Recamán’s sequence, Thue-Morse, Dress’s sequence, and even 10-adic decimal expansions like A225410. These sequences generate rhythmic patterns, melody selection, and track overlap strategies in algorithmic composition tools.

HN Discussion: Commenters questioned whether music based on number sequences actually sounds good — one argued raw sequences produce “random garbage” that requires a musician’s arrangement to become musical. Max Cooper was cited as an artist who uses OEIS sequences effectively (tracks “Aleph 2,” “Fibonacci Sequence”). Tool’s “Lateralus” mentioned as mainstream exposure of Fibonacci themes in popular music.

Museum of Imaginary Musical Instruments

Summary: The MIMI website catalogs fictional instruments from literature, history, and art — including J.G. Ballard’s Singing Statues from a 1962 novel, Carlos Salzedo’s 1927 Polyharp concept, La Nature’s 1883 Cat Piano (Katzenklavier), and Jonathan Ulman’s 2024 AI-generated “Snare Drums Reimagined” series depicting percussion instruments made of bricks, televisions, and quilts.

HN Discussion: Commenters suggested AI music videos featuring these invented instruments would be entertaining to watch. One referenced P.D.Q. Bach’s tromboon — a “real but shouldn’t exist” instrument — as a real-world parallel. Disaster Area’s experimental setup was mentioned as fitting the category. The overall sentiment was appreciation for the collection itself.