Tonight’s brief ranges from Rust-on-CUDA tooling and Swift matrix kernels to privacy side channels, AI labor anxiety, security automation claims, and a few unusually reflective detours through community spaces, old web aesthetics, and computing history.

AI & Tech Policy

Training an LLM in Swift, Part 1: Taking matrix mult from Gflop/s to Tflop/s

Summary: Matt Gallagher begins a series on training neural networks in Swift by hand-optimizing matrix multiplication on Apple Silicon. The article walks through ten implementations, from plain C and Swift to versions that exercise CPU, SIMD, AMX, and Metal GPU paths. Gallagher is explicit that production users should prefer Apple’s established ML frameworks, but uses the no-frameworks exercise to explain what different execution units can really deliver.

HN Discussion: Readers praised the piece as unusually concrete writing on Swift performance, even outside the LLM context. The thread dug into compiler flags for fused multiply-add, especially why -ffp-contract=fast can be a narrower choice than -ffast-math, and used the GPU results to discuss why peak TFLOP figures often mislead.

AMÁLIA and the future of European Portuguese LLMs

Summary: Duarte O. Carmo examines AMÁLIA, Portugal’s 5.5 million euro effort to build an open-source LLM for European Portuguese. The project brings together Portuguese universities and research labs and aims to make European Portuguese a first-class language rather than an afterthought of larger multilingual systems. Carmo notes that AMÁLIA was not trained from scratch, compares the motivation with Italy’s Minerva, and argues that public funding justifies hard questions about design choices.

HN Discussion: Commenters debated whether small language- or country-specific models are the right investment, especially if they remain weak at factual retrieval or translation. Data scarcity was a recurring technical issue, with readers asking whether Brazilian Portuguese corpora could be transformed and whether language-specific models trade local usefulness for access to broader world knowledge.

Interfaze: A new model architecture built for high accuracy at scale

Summary: Interfaze claims a new architecture for computer-level tasks where predictable extraction matters more than open-ended language generation. The company says it outperforms Gemini-3-Flash, Claude-Sonnet-4.6, GPT-5.4-Mini, and Grok-4.3 across benchmarks in OCR, vision, speech-to-text, and structured output. Its core argument is that specialized DNN/CNN-style systems can handle documents, coordinates, translation, and GUI detection with useful metadata such as bounding boxes and confidence scores.

HN Discussion: The discussion focused on structured output as a practical bottleneck, particularly for smaller or local models that become much more useful if they reliably obey schemas. Readers asked whether the architecture can handle code extraction and whether fine-tuning will be available, treating it more as workflow infrastructure than as another general chatbot.

Software engineering may no longer be a lifetime career

Summary: Sean Goedecke argues that even if AI-assisted programming causes some skill atrophy, that is not by itself a decisive reason to avoid AI at work. The older career bargain was unusually convenient: doing software engineering was also the best way to learn software engineering. Goedecke suggests that AI may separate the job from the learning path, making software engineering less of a stable lifetime craft even if the transition resembles earlier abstraction shifts such as assembly to C.

HN Discussion: Commenters pushed back on the idea that developers mainly write code, saying most engineering time goes into understanding systems and shaping solutions. Others argued that experienced engineers using AI as augmentation are already stronger, while the real risk lies with people who let AI replace their reasoning rather than support it.

Students Boo Commencement Speaker After She Calls AI Next Industrial Revolution

Summary: 404 Media reports that University of Central Florida humanities and communication graduates booed Gloria Caulfield after she called AI the next industrial revolution. Caulfield, a Tavistock Group vice president of strategic alliances, was speaking at UCF’s College of Arts and Humanities and Nicholson School of Communication and Media. The crowd reportedly shouted “AI sucks,” making the ceremony a visible snapshot of anxiety among graduates entering fields exposed to generative-AI disruption.

HN Discussion: The thread argued that young people are unlikely to celebrate AI if the promised future looks like lower income, precarity, or corporate concentration. Several commenters framed the issue as distributional: AI may be transformative, but graduates want evidence that ordinary workers rather than trillion-dollar firms will benefit.

A.I. note takers are making lawyers nervous

Summary: The New York Times article itself was unavailable in the compact pack, but the title and discussion point to legal worries around AI meeting note takers. The concern is that a transcription or summarization bot can enter professional conversations that include privileged, confidential, or strategically sensitive material. Commenters identified two concrete risks: attorney-client privilege may be complicated by a third-party service, and casual speech may become a durable record discoverable in later litigation.

HN Discussion: Readers were not opposed to note-taking in principle; they objected to the default SaaS pattern where a random bot joins every call and uploads recordings or transcripts elsewhere. The practical design themes were consent, storage, access control, and whether tools can help people focus without turning every meeting into legal evidence.

What a Japanese cooking principle taught me about overcoming AI fatigue

Summary: Takuya Matsuyama reflects on staying creatively healthy as a developer, content creator, and artist during the AI boom. Instead of building a defensive career moat, he argues for orientation: deciding where one wants to go while strategies adapt. The essay uses Yoshiharu Doi’s “one soup, one dish” cooking principle as a model for a steadier pace, suggesting that simple, durable creative practice may be more sustaining than chasing every AI shift.

HN Discussion: The visible comment connected AI fatigue with novelty-seeking, arguing that intentionally appreciating routine can be powerful. It also pushed back on the idea that fatigue must be overcome, suggesting that the feeling may be a useful signal that the current pace or substance of AI culture is unhealthy.

---

Security & Privacy

Gmail registration now requires scanning a QR code and sending a text message

Summary: A Privacy Guides thread reports that new Google account creation may now show a QR code that opens an SMS message from the user’s phone to Google, instead of Google sending a verification code to the user. The poster frames this as a security measure that also makes anonymous registration and SMS verification services harder to use. The pack supports an important correction: this appears to be an SMS URI flow that asks the user to send a message, not a text silently sent just by scanning.

HN Discussion: Commenters split between privacy concerns and sympathy for Gmail’s abuse problem, noting that Google runs widely depended-on free infrastructure under constant spam pressure. One concrete explanation was SMS pumping: requiring the user to send the SMS can break the economics of schemes that profit from receiving verification texts.

Should you leave red herrings about yourself online?

Summary: Alcazar Security argues that most people should not scatter fake jobs, cities, birthdays, and life details around the web as a default privacy tactic. Broad deception rarely beats systems that combine public records, broker data, and previous leaks, and it can create practical hassles such as confusing account-recovery questions. The post separates pseudonyms and compartments from broad false facts and targeted deception, recommending that any misdirection be tied to a narrow threat model.

HN Discussion: Readers compared the advice with early-2000s internet norms, when avoiding real names, locations, and consistent headshots felt obvious. The thread also cited Derek Sivers’ view that a deliberate fake persona can be better than refusing to share details, while jokes about absurd false biographies underlined how easily red-herring strategies become silly.

Guitar tuner that uses phone accelerometer

Summary: This browser-based guitar tuner detects string vibration with a phone accelerometer rather than a microphone. The page asks users to press the phone against the guitar body, pluck a string, and watch raw X, Y, Z, and combined acceleration traces. Pitch is derived from the strongest axis with alias correction, and the tool includes standard guitar-string presets while noting that high-rate Android IMUs work best.

HN Discussion: Commenters tested browser sensor limits, including a 50 Hz accelerometer rate too low for the lowest guitar string and an improvised foosball-table experiment. Beyond musical usefulness, readers raised a privacy concern: accelerometer access plus frequency analysis can become a microphone-like side channel without normal audio permissions.

Mythos Finds a Curl Vulnerability

Summary: Daniel Stenberg recounts curl’s experience with Anthropic’s Mythos model after claims that it was unusually strong at finding security flaws in source code. Curl was offered access through the Linux Foundation and Alpha-Omega, though the process stalled after contracts were signed. The post emphasizes that Mythos found one curl vulnerability, not the avalanche implied by some surrounding hype, and Stenberg’s excerpted conclusion calls the campaign successful marketing rather than proof of a major leap.

HN Discussion: Commenters quoted Stenberg’s skepticism and treated the initial panic as marketing-driven. Others noted that security scares can still unlock budgets, while a counterpoint held that curl may be unusually well-hardened and contained, so modest results there may not predict performance on larger, messier codebases.

ICE to Develop Own Smart Glasses to ‘Supplement’ Its Facial Recognition App

Summary: 404 Media reports that ICE is exploring smart glasses that would supplement Mobile Fortify, the agency’s facial recognition app. A DHS official and another conference attendee described plans for glasses that could help officers verify citizenship by scanning faces. The article frames the possible device as another technology escalation in the Trump administration’s mass-deportation campaign and connects it to prior reporting on ICE and CBP’s internal use of Mobile Fortify.

HN Discussion: The discussion was dominated by civil-liberties and state-surveillance concerns, including calls to abolish ICE and bar some personnel from public service. One cynical counterthread treated the project as a possible procurement boondoggle, while archive links and Orwell references captured both access concerns and privacy alarm.

Driver accused of DUI tracks missing laptop to Illinois State trooper’s house

Summary: ABC7 Chicago reports that a restaurant executive arrested on DUI charges tracked his missing MacBook to the home of Illinois State Police Trooper Kevin Bradley. The I-Team obtained a 911 call, cellphone video, and internal State Police investigation tied to the incident. Bradley had been recognized as a prolific DUI-arresting “Top Cop,” and the report places the laptop case amid broader accusations from public defenders and civil-rights attorneys that he fabricated evidence and violated drivers’ rights.

HN Discussion: Commenters focused on chain of custody, saying an arrestee’s property should never end up at the arresting officer’s house. Accountability was the sharper theme: readers highlighted the reported one-day suspension and broadened the case into concerns about police corruption, coverups, DUI-enforcement incentives, and overtime pay.

Traces Of Humanity

Summary: The post announces Traces of Humanity, a new blog by a writer returning after seven years of silence. The author references starting Qubes OS in 2009, leading it for nine years, and previously writing deeply technical material on systems, operating-system, and virtualization security. This new project is framed more broadly around tensions between rationality and humanism, pragmatism and beauty, formalism and intuition, freedom and love.

HN Discussion: Commenters supplied context for readers unfamiliar with Joanna Rutkowska, pointing to her Blue Pill work, Xen and Vista virtualization attacks, and influence on security research. Qubes OS drew praise as especially relevant if LLMs accelerate vulnerability discovery, while several readers simply welcomed her return and the more reflective direction.

AI-powered hacking has exploded into industrial-scale threat, Google says

Summary: The Guardian reports Google’s warning that AI-powered hacking has grown into an industrial-scale threat. The headline and URL indicate a rapid escalation over a recent three-month period, but the compact pack does not expose Google’s detailed evidence, examples, targets, or mitigation advice. The supported claim is therefore narrow: the story presents AI capability as changing the scale of cyber operations, while specifics were not available in the allowed source material.

HN Discussion: The compact HN discussion was essentially undeveloped, with the visible comment providing an archive link. No substantive debate about Google’s methodology, threat models, or AI-specific attack workflows appeared in the pack, so the only supported thread theme is access to the article rather than analysis of its claims.

---

Tech Tools & Projects

CUDA-oxide: Nvidia’s official Rust to CUDA compiler

Summary: Nvidia’s cuda-oxide is an experimental compiler for writing SIMT GPU kernels in idiomatic Rust and compiling them directly to PTX. The book stresses standard Rust code instead of a DSL or foreign-language binding, with material on kernels, device functions, memory movement, launches, closures, generics, and error handling. It also describes a “safe(ish)” GPU safety model and compiler internals involving Stable MIR, a rustc CUDA backend, Pliron IR, lowering, intrinsics, fuzzing, and differential testing.

HN Discussion: Commenters concentrated on whether Rust’s ownership and memory model can map cleanly onto CUDA’s execution and memory semantics. Practitioners compared the project with cudarc, CMake/nvcc-based Rust CUDA workflows, and FFI-heavy setups, hoping it could lower friction and improve build times for custom kernels.

Ratty – A terminal emulator with inline 3D graphics

Summary: Ratty presents itself as a GPU-rendered terminal emulator that can display inline 3D graphics. The project site is sparse, but it positions the tool around extending terminals beyond text with downloads, source, and a linked blog post. Its central claim is not just smoother terminal rendering; it is richer graphical output directly in terminal workflows, credited to Orhun Parmaksız as a 2026 experiment.

HN Discussion: Readers debated whether terminals should evolve toward notebook-like mixed media, comparing Ratty with Kitty protocol extensions and older Lisp-machine or Xerox workstation REPL experiences. Practical questions centered on 2D image quality, rasterization, and what GPU acceleration means for SSH sessions, while some reactions treated it as a fun echo of Compiz-era desktop effects.

Scaffold a 1990s Geocities-themed static website

Summary: The PyPI project create-geocities-app appears to scaffold a 1990s GeoCities-themed static website. The compact excerpt was blocked by a client-side challenge, so the supported detail is limited to the package identity and purpose. It fits a nostalgia-driven category of tooling that recreates personal-homepage aesthetics, aiming less at production infrastructure than at quickly generating a playful retro starting point.

HN Discussion: Commenters connected the project to a desire for the web to feel personal and creative again, contrasting older homepages with today’s monetized platforms. Others objected that a faithful GeoCities experience should require hand-written HTML, tables, iframes, and hours of tinkering; one reader called the output more vaporwave re-imagination than historical recreation.

Phel v0.36.0 – Lisp on PHP, now with numeric tower and first-class Vars

Summary: Phel is a Clojure-flavored Lisp that compiles to PHP, and version 0.36.0 is labeled the “Precision” release. It adds exact rational literals such as 1/2, arbitrary-precision BigInteger with overflow auto-promotion, and BigDecimal. The release also introduces first-class Vars with forms such as #'sym, alter-var-root, with-redefs, and watches, plus UUID, Queue, and MapEntry value types and faster REPL and test boot.

HN Discussion: Commenters were curious how a Lisp maps onto PHP’s separate namespaces for classes, functions, constants, methods, variables, and mixed case-sensitivity rules. The Clojure influence was received positively, especially around Vars and namespace syntax, and one reader praised standardizing dot-separated namespaces in the core.

---

Web & Infrastructure

Building a web server in aarch64 assembly to give my life (a lack of) meaning

Summary: ymawky is a small static HTTP server written entirely in AArch64 assembly for macOS/Darwin. The project deliberately avoids libc wrappers, existing parsers, and external libraries, using raw Darwin syscalls as a learning exercise. Despite the intentionally difficult route, it supports GET, HEAD, PUT, OPTIONS, DELETE, byte ranges, directory listings, custom error pages, and hardening measures, while the author says it is not meant to replace nginx.

HN Discussion: Commenters enjoyed the irony of hating string parsing while choosing an assembly HTTP server as a side project. The thread compared assembly’s practical value in embedded or constrained systems with the dubious practicality of a general web server, and some readers wondered how far the assembly-only approach could go toward bare-metal or OS-from-scratch work.

---

Academic & Research

Show HN: TikTok but for Scientific Papers

Summary: Papel pitches a social, full-screen discovery app for researchers with personalized paper feeds ranked by interests, trends, freshness, and community engagement. The product claims more than two million indexed papers and free access without subscriptions. Its AI features include on-device chat over full PDFs using Apple Intelligence or local MLX models, plus a RAG pipeline intended to keep research data on the user’s device; gamification adds quizzes, XP, streaks, and academic-themed ranks.

HN Discussion: Commenters questioned the “TikTok” positioning, arguing that researchers may recoil from the comparison and that the paper-discovery value can stand without polarized branding. Skeptics worried that social feeds around papers could amplify clickbait and decontextualized conclusions, while lighter comments joked about game mechanics and “Flappy Bird with Transformers.”

Classification of amino acids

Summary: The linked Khan Academy resource is an MCAT preparation lesson on classifying amino acids within the topic of amino acids, peptides, and proteins. The pack could not load the actual lesson content because of a client challenge, so the specific classification scheme and examples were not available. From the URL and title, the submission is a foundational biochemistry teaching module for pre-med learners, not a new research result.

HN Discussion: The thread was extremely sparse, with the visible commenter asking why an old Khan Academy class belonged on HN. No substantive discussion of amino-acid chemistry, pedagogy, or MCAT preparation appeared in the compact pack, making submission age and fit the only supported theme.

---

Business & Industry

Microsoft Israel chief leaves amid ethical controversy

Summary: Globes reports that Alon Haimovich, Microsoft Israel’s country general manager, is leaving after four years in the role. The article says the departure followed a Microsoft global-management investigation into alleged unethical use of Azure by Israel’s Ministry of Defense and Microsoft Israel’s work with that ministry. Several managers in Microsoft Israel’s governance department reportedly also left, and the local office has been placed under Microsoft France’s management.

HN Discussion: The compact HN thread was thin, with the visible comment essentially quoting the article’s central claim about Azure, the Ministry of Defense, and Microsoft France. With little discussion developed, the supported theme is corporate governance and accountability around cloud services used by defense customers, not a detailed debate about evidence or contracts.

---

History & Science

Nullsoft, 1997-2004 AOL kills off the last maverick tech company (2004)

Summary: Paul Boutin’s 2004 Slate piece recounts AOL’s layoffs at Nullsoft, reducing the Winamp and Shoutcast maker to only a few employees. The article casts Nullsoft as AOL’s unusually productive and rebellious outpost, led by Justin Frankel after AOL bought his services in 1999. It highlights Winamp and Shoutcast’s role in the MP3 era and describes unauthorized projects that irritated both AOL management and the recording industry.

HN Discussion: Commenters filled in Frankel’s later career at Cockos, especially REAPER, noting the DAW’s tiny installer and reputation as a serious Pro Tools competitor. Several celebrated the audacity of releasing Gnutella from inside a media-company acquirer, while others followed Frankel’s current writings on scraping, Cloudflare, AWS, and decentralization.

Venom and Hot Peppers Offer a Key to Killing Resistant Bacteria

Summary: WIRED reports on research that turns compounds associated with scorpion venom and habanero chile toward possible antibiotics against resistant bacteria. The headline and URL identify the work as Mexican science aimed at antibiotic resistance, using naturally toxic or pungent molecules as the starting point. The compact excerpt did not provide target organisms, clinical stage, dosing, or toxicity details, so the safe takeaway is the antibiotic-development direction rather than a validated treatment.

HN Discussion: Commenters were cautious about lab-stage antibiotic claims, citing past compounds that looked broad-spectrum and low-toxicity but disappeared without clinical follow-up. Others connected venom-based medicine to older healing symbolism or joked that merely killing bacteria is not enough to prove medical usefulness.

Bliss (Photograph)

Summary: The Wikipedia article covers Bliss, the landscape photograph best known as the default wallpaper for Windows XP. The excerpt shows sections on overview, history, reception, legacy, re-creations, references, and external links, and includes the California coordinates of the photographed location. The article treats the image as both a photograph and a cultural artifact made globally recognizable by Windows XP’s distribution.

HN Discussion: Commenters discussed Microsoft’s rights acquisition, including the reportedly very large payment to photographer Charles O’Rear and the unusual logistics of delivering the image. Licensing questions focused on whether Microsoft wanted to prevent non-Windows uses and how earlier stock-image rights may have been handled, while others mentioned alternate wallpapers and using Bliss as a monitor or printer test image.

The Adventure Family Tree (2024)

Summary: The Adventure Family Tree describes itself as an incomplete chronicle of Crowther and Woods’ Colossal Cave Adventure and its descendants. The excerpt calls Adventure the origin point of interactive fiction and a hallowed artifact of computer-gaming history. The page appears to map variants, footnotes, files, and relationships among versions, turning the spread and mutation of an early text adventure into documented computing history.

HN Discussion: Commenters shared memories of modifying Adventure source code, including adding a surreal “Nowhere” room and an inventory item called “nothing.” Others praised the project’s citations and historical value, and one reader noted discovering Donald Knuth’s 2002 Literate Programming port of Adventure.

---

System Administration

Building a Memory Allocator from Scratch in C

Summary: 0xKiire’s article teaches memory allocation by implementing malloc() and free() from scratch in C. It frames allocators as a fundamental systems concept that programmers often hide behind library calls but that matters for efficient, reliable software. The post introduces a typical C program’s stack, heap, BSS, data, and text segments before moving into heap management, fragmentation, optimization techniques, and allocator tradeoffs.

HN Discussion: The compact pack contained no HN comments for this story. As a result, there were no supported commenter themes about allocator algorithms, fragmentation strategy, or comparisons with production malloc implementations.

---

Other

Holding Community Space

Summary: Jesse Evers reflects on running Highside, a roughly 2,500-square-foot rented warehouse community space in Brooklyn that eventually shut down. The essay focuses on the tension between making a space welcoming and generative while keeping it financially and organizationally sustainable. Highside is described as basically anarchic but with one person ultimately responsible, and the post treats trusting people with nice things as part of the space’s magic.

HN Discussion: Commenters with hackerspace experience said rent, cleaning, organizing, and financial support often fall on a small core group. A practical theme was that would-be members may not understand recurring costs until memories and attachment form, while another reader starting a makers’ space in Portland related strongly to the author’s anxiety and imposter syndrome.

Why we lose our friends as we age (2023)

Summary: Isabel Fattal’s Atlantic newsletter reflects on why friendships often thin out after school and early adulthood. It uses college proximity as the contrast: a time when friends may live minutes away and share routines without much planning. The stated thesis is that choosing friendships gives them value, suggesting that adult friendship becomes more intentional when default proximity disappears and the loss is partly structural rather than simply personal failure.

HN Discussion: Commenters emphasized childcare, finite free time, and logistics as major filters on adult social life, especially for parents. Several described workplace friendships, recurring gatherings, and COVID-disrupted routines fading once contact required deliberate effort, while others said middle age also brings a clearer preference for healthier, less negative social circles.