📰 Tech Trends Daily — Monday, November 16, 2026

🔥 Today’s Focus

Two failure paths worth remembering today, both of the “nobody expected it to break this way” kind: a Microsoft 365 outage left alert emails from a German hotel’s sauna chlorination system stuck in transit, and 15 guests inhaled excessive chlorine gas; ULA’s Vulcan return-to-flight failed, taking 40 Amazon Leo satellites down with it. The former entrusted the last link of a life-safety chain to email, the latter bet constellation deployment on a freshly repaired rocket — alerts and redundancy were both in place, but both were built on the same assumption. The loudest AI debate is “models consumed all of humanity’s code — should their output be forced open source?” Two legal precedents are pushing the answer toward “pure AI output has no copyright; human-involved portions remain protected.”

🤖 AI & Models

  • All AI-generated code should be Open Source — All AI-generated code should be Open Source. 1019 points / 565 comments (HN); 25/9 (Lobsters). The argument: models trained on all of humanity’s code corpora, so their code output shouldn’t be privatized either.
    • 💬 Comments: The US Copyright Office’s stance was cited repeatedly — purely AI-generated content isn’t copyrightable, while parts involving human expression remain protected. Opposition centers on the “tool argument”: code written with open-source frameworks was never required to be open-sourced because the framework is open; the developer bears the responsibility and losses, so the responsible party owns the output. The most concrete story comes from an AGPL author: his code was stolen by a company, and after a few rounds of infringement notices they simply had a model rewrite it — “like using ChatGPT to rewrite Harry Potter: same plot, terrible dialogue.” Others anchor on precedent: Doe v. GitHub held that model output doesn’t inherit open-source licenses, and Thaler v. Perlmutter held that 100% AI output can’t get copyright.
  • What positional embeddings in LLMs actually do — Understanding Positional Embeddings in Modern LLMs. 185 points / 23 comments (HN). Raschka walks through mainstream approaches from absolute positions to RoPE — good as a cheat sheet.
  • What thousands of conversations look like inside a language model — What thousands of conversations look like inside a language model. 106 points / 29 comments (HN). The author projected Qwen3-1.7B’s internal conversation representations into an interactive map, checking whether similar topics land in adjacent regions and whether the same idea expressed in different languages converges.
  • Eval Awareness vs. Eval Paranoia — Eval Awareness vs. Eval Paranoia. 39 points / 17 comments (HN). Distinguishing “the model knows it’s being tested” from “researchers interpret every model behavior as deception” — the latter has become an unfalsifiable argumentative habit.
  • Your Agent Will Pick the Hackable One — Your Agent Will Pick the Hackable One. 50 points / 33 comments (HN). The author starts with the $70 M-VAVE FM-1 synth: almost nobody kept the stock firmware, users have been writing their own since September 2025 (mostly with AI coding tools), and by October the vendor simply open-sourced the firmware under GPLv3 with an SDK and browser simulator. The inference: openness will become a purchasing criterion — nobody consciously compares openness, but agents picking hardware will favor what they can modify and repair.
  • Claude Code and Codex reverse engineered my 1994 Jeep’s ECU over a $3 Arduino — Claude Code and Codex reverse engineered my 1994 Jeep’s ECU over a $3 Arduino [video]. 284 points / 149 comments (HN). The owner let two coding agents probe the bus directly to figure out ignition and fuel logic, and the car actually ended up being driven.
    • 💬 Comments: The most-quoted moment is when an agent proposed driving the fuel pump itself, then repeatedly pressurized the fuel rail until the owner yelled stop. Some say this is AI’s real value — returning control to device owners; others cite reports that Anthropic is tightening reverse-engineering capabilities, noting vendors could cut off this path at any time.

🔒 Security & Privacy

  • Old Bitly Links Are Being Hijacked — Old Bitly Links Are Being Hijacked. 34 points / 18 comments (HN). Short links in years-old tweets from Burger King, McDonald’s and Trump now point elsewhere. Bitly confirmed and suspended the accounts, official wording: “someone exploited old inactive links and changed their targets” — once a short link gives up control, citations in historical content become a resellable attack surface.
  • Keylogger Detection and Containment (thesis) — Keylogger Detection and Containment [pdf]. 5 points / 1 comment (HN). A full PhD thesis from Vrije Universiteit Amsterdam on detecting and constraining input-stealing components from the system side — more concrete than tool marketing.
  • Aspidistra: A British wartime radio transmitter used for MitM attacks — Aspidistra: A British wartime radio transmitter used for MitM attacks. 10 points / 1 comment (HN). The method: overpower enemy stations on their own frequency with higher power — signal-layer man-in-the-middle hasn’t changed principles in decades.
  • Preventing Cross-Service UDP Loops in QUIC — Blog: Preventing Cross-Service UDP Loops in QUIC. 5/0 (Lobsters). Protocol-level details from Google’s security team: how to avoid two services drowning each other in UDP.
  • Disk encryption on NetBSD achieving secrecy — Disk encryption on NetBSD achieving secrecy. 12/1 (Lobsters). The author lays out cgd’s threat model, key storage locations, and “what level of encryption actually counts” item by item — the kind of document where you draw your own boundaries.

🏢 Companies, Incidents & Money

  • Microsoft 365 outage leads to chlorine gas poisoning of 15 hotel guests — Microsoft 365 outage leads to chlorine gas poisoning of 15 hotel guests. 281 points / 148 comments (HN). The automatic chlorination system at Phantasialand’s hotel sauna area overdosed, and nobody received the alert emails sent by the safety system — Microsoft 365 was down at the time. German broadcaster WDR confirmed the chain with the company.
    • 💬 Comments: A reader who builds chlorination systems himself explained the failure mode — low sensor readings cause overdosing, so commercial pools must be manually titration-tested daily before opening, with handwritten records. Top comments blame the hotel’s architecture: staking the last link of safety alerting on email with no second confirmation channel; an Azure employee also commented “not defending the outage, but this is a design problem.”
  • Vulcan rocket fails during return-to-flight launch, 40 Amazon satellites lost — Vulcan rocket fails during return-to-flight launch, 40 Amazon satellites lost. 64 points / 27 comments (HN). ULA’s Vulcan had a problem during its return-to-flight, and the Amazon Leo constellation satellites aboard never reached orbit.
  • Boeing-Lockheed rocket fails mid-flight, destroys payload of Amazon satellites — Boeing-Lockheed rocket fails mid-flight, destroys payload of Amazon satellites. 25 points / 11 comments (HN). The wire-service version of the same incident, with a clearer timeline and loss list. Amazon spreads constellation deployment across multiple rockets, and this time lost 40 satellites at once.
  • Ask HN: Stripe closed us after one $1,500 invoice and never asked for evidence — Ask HN: Stripe closed us after one $1,500 invoice and never asked for evidence. 41 points / 9 comments (HN). The founder’s business charges contractors based on actually completed bookings; Stripe flagged it as unauthorized transactions and closed the account without even requesting supporting materials.
    • 💬 Comments: Two technical replies point to the same thing — a “small charge, then large charge” sequence is exactly the signature of card-testing fraud, and banks will block it outright. The advice: use Setup Intents for off-session authorization, with the first charge being held by risk controls as a normal cost. Another commenter says Stripe already took $4.5k from them, and emails to Patrick and John went unanswered.

🛠️ Tools & Infrastructure

  • Protomaps: Self-hosting map tiles has gotten cheap and easy — Protomaps: Self-hosting map tiles has gotten cheap and easy. 73 points / 16 comments (HN). PMTiles packs an entire tile set into one file served via Range requests; drop it in object storage with static serving and it can replace commercial map APIs — the author calculated Mapbox costs $5 per thousand loads past the free tier.
  • Eclipse Mosquitto – An open source MQTT broker — Eclipse Mosquitto – An open source MQTT broker. 52 points / 19 comments (HN). The broker most often chosen by default in IoT projects: single process, simple config, TLS and ACL self-configured. For learning MQTT it beats chewing through the protocol spec.
  • Emacs Is My Browser — Emacs Is My Browser. 163 points / 83 comments (HN). The author moved browsing into EWW, claiming it covers 80–90% of use cases; the motivation is that browsers drag you into unrelated rabbit holes too easily. The cost: sites requiring JS are mostly abandoned.
  • Clice v1.0 Released: Welcome to Modern C++ — Clice v1.0 Released: Welcome to Modern C++. 71 points / 43 comments (HN); 3/0 (Lobsters). A clang-based LSP implementation whose selling point is indexing and completion responsiveness in large projects.
  • A datetime library for Rust — A datetime library for Rust. 6/1 (Lobsters). BurntSushi’s work; timezone database handling and timezone arithmetic are the main selling points — another option besides chrono.
  • FreeBSD bhyve Live Migration: Moving VMs Between Hosts — FreeBSD bhyve Live Migration: Moving VMs Between Hosts. 23/0 (Lobsters). Still at RFC stage, filling the operational gap bhyve has long lacked versus KVM.
  • Please Implement A Quorum Database — Please Implement A Quorum Database. 2/0 (Lobsters). Notes from an HPTS 2026 talk; the author argues that industry treating quorum replication as impractical is an outdated view. The piece also includes a distribution chart of production database replication algorithms — Raft has absorbed both the primary-backup and quorum camps.
  • cursor:pointer on buttons — cursor:pointer on buttons. 27/12 (Lobsters). Chasing “should buttons get the hand pointer” all the way to the spec and browser defaults — the kind of piece that digs deep into something small.

💻 Languages & Programming

  • Memory management in 40 major C projects, a survey — Memory management in 40 major C projects, a survey. 43/10 (Lobsters). eatonphil surveys these projects’ allocation strategies, error handling, and cleanup paths; you come away with an empirical distribution rather than just one project’s preferences.
  • Things That Turbo Pascal Is Smaller Than (2011) — Things That Turbo Pascal Is Smaller Than (2011). 180 points / 89 comments (HN). Turbo Pascal 3.02’s compiler and IDE together total 39,731 bytes — smaller than the minified jQuery 1.6 of 2011. This old piece resurfacing on the front page shows the comparison still holds.
  • A NES emulator written in Odin — A NES emulator written in Odin. 98 points / 35 comments (HN); 13/1 (Lobsters). The author started in June 2024, paused for over half a year, and the most painful phase was chasing “complete accuracy” — there’s no shortcut for things like PPU timing.
  • 20 Years of Pyrolog — 20 Years of Pyrolog. 66 points / 4 comments (HN). A project embedding logic programming into the Python runtime; the author thanked contributors in the comments, hinted at possible revival, and name-dropped another Python Prolog (Clausal Prolog).
  • Parsing nested lists in BQN — Parsing nested lists in BQN. 16/1 (Lobsters). Array language style differs greatly from imperative; this one lays out the full derivation, good as an introduction.
  • Rust Declarative Macros: Organization, Scoping, and Common Use-Cases — Rust Declarative Macros: Organization, Scoping, and Common Use-Cases. 1/1 (Lobsters). Hygiene and scoping are the two easiest places to trip with macro_rules!; this piece is organized by use case.
  • Some problems of recursive descent parsers (2009) — Some problems of recursive descent parsers (2009). 1/0 (Lobsters). How left recursion and backtracking become pitfalls was written clearly over a decade ago; hand-written parser authors should read this first.
  • Reintroducing Spaces — Reintroducing Spaces. 16/1 (Lobsters). The atproto team adding non-public data capability to the protocol, positioned for social scenarios rather than “bookmarks or locked posts”; the protocol is currently alpha.
  • Periodic-Oriented Programming for FIRST Robotics — Periodic-Oriented Programming for FIRST Robotics. 20/2 (Lobsters). An author with 17 years of FRC programs argues the official libraries have long encouraged over-engineering, advocating a return to the simple model of “read inputs, compute state, write outputs each cycle.”
  • Anthropomorphic Software — Anthropomorphic Software. 21/11 (Lobsters). A survey of programs that speak in first person — git’s “Please tell me who you are” and “I’m stopping.” The author excludes conversational AI, but the ending still has to discuss it.

🔬 Hardware, Graphics & Retro

  • The MacBook Pro’s OLED PWM problem — The MacBook Pro’s OLED PWM problem. 376 points / 203 comments (HN). A brighter screen is a regression for flicker-sensitive people: 240Hz-class PWM makes some people get headaches and lose focus after extended viewing.
    • 💬 Comments: Someone verified the frequency with an oscilloscope and argued the 240Hz figure is incomplete — the OLED’s two stacked layers may each be driven at 480Hz, 180 degrees out of phase, for an effective 960Hz across most of the usable brightness range. A reader with a concussion history said this describes their daily experience with all screens, and a red filter helps; others noted this is an often-forgotten accessibility issue.
  • 25 of my favourite modern NES games — 25 of my favourite modern NES games. 63/9 (Lobsters). The console stopped production long ago, but new NES games keep being made; the author lists a batch by play value.
    • 💬 Comments: One veteran player said reading it felt like walking back into an eighties cartridge aisle; the author replied “that’s exactly the effect I hoped for,” adding that neither the blog post nor the games mentioned contain AI-generated content.
  • WebGL Flight Simulator with seed generated terrain — WebGL Flight Simulator with seed generated terrain. 34 points / 7 comments (HN). A WebGL flight simulator with seed-generated terrain, throttle, flaps, brakes, and both keyboard/touch controls — just open and play.
  • Handwriting, as a Font — Handwriting, as a Font. 50 points / 9 comments (HN). Write letters in the browser following prompts, and it generates a font that belongs only to you — the letters never leave your machine.

🎮 Light & Life

  • ‘Birdbrain’ Is a Misnomer. Crows Make Decisions Faster Than We Can — ‘Birdbrain’ Is a Misnomer. Crows Make Decisions Faster Than We Can. 122 points / 90 comments (HN). The research puts crows’ reaction speed above humans; the contrast in the headline alone is worth a read.
    • 💬 Comments: The most valuable comment points out that intelligence has evolved independently at least three times on Earth — mammals, corvids, octopuses — with mutually different brain structures, so “intelligence has only one way of being done” doesn’t hold. Others caution against getting smug too fast: crows must fly at 40 mph, so spatial and temporal perception has to be extremely strong.
  • Why did some Apollo missions carry a grenade launcher? (2019) — Why did some Apollo missions carry a grenade launcher? (2019). 213 points / 141 comments (HN). An old Q&A back on the front page; the answer quotes NASA reports directly: the detonators are triggered by ground commands arm grenades and fire grenades through command gating, using 72V signals to charge the detonation capacitors before firing them in sequence.
    • 💬 Comments: A reader dug into the asker’s account pattern — from November 2 to 22, 2018, that account asked exactly one question per day, all phrased to elicit quotable primary-source answers. Another added the Soviet side: crews returning to Siberia did carry guns, against wolves and bears.
  • Your room can be as bright as the outdoors (2019) — Your room can be as bright as the outdoors (2019). 64 points / 44 comments (HN). Boston winters go dark at 4:30 pm; the author gave up waiting for natural light and used bulbs bright enough to need thermal design to push the room to daylight intensity.
  • A short explainer on how El Niño works [video] — A short explainer on how El Niño works [video]. 134 points / 41 comments (HN). Explains the relationship between sea temperature anomalies and atmospheric circulation in a few minutes — good to stockpile as outreach material.
  • Bullet Dodged — Bullet Dodged. 35 points / 29 comments (HN). Milk kept spoiling in the fridge; a thermometer showed over 50 degrees. Using an inarticulate repair video, the author identified the compressor inverter module for the same model, double-checked with a multimeter, ordered the part, and rescued a seventy-year-old fridge.
  • Notes on Mathematics for Engineers (1959) [pdf] — Notes on Mathematics for Engineers (1959) [pdf]. 60 points / 15 comments (HN). University of Michigan course material, running from limits and series to Laplace transforms, written more directly than modern textbooks.

📌 Still on the Charts: Notes

8 of the 25 items on the Lobsters front page overlap with yesterday’s two editions — How Not to Build a Music Player (a vibe-coding music player failure log), The need for at least a third option (mobile left with only two choices), What happened to 3com, Solving the fox game, Creating a self booting ZIP archive, Terminal borders that don’t cost a cell, Someone’s impersonating this blog, Let Me Let You Own Your Data — not repeated here; see vol-148 and vol-149 for links. On the HN side, SpaceX Starship’s first catch by the launch tower was already covered yesterday and serves only as background today: the same batch of satellites was lost entirely on Vulcan, making the value of two-stage reuse more vivid by contrast.

🧭 Summary

Today’s mood is half maintenance logbook, half theoretical positioning: when equipment breaks, measure the sensor first; when a rocket explodes, check the timing first — while at the other end, people are finding a place for AI-generated code in copyright law. The three must-reads are the complete causal chain of the M365 outage and chlorine poisoning (especially the self-built chlorination reader’s explanation of failure direction in the comments), the two precedents laid side by side in the AI code open-source debate along with the AGPL author’s hands-on experience, and the “agent will pick the hackable hardware” piece — which turns openness from a moral issue into a procurement parameter. The cross-cutting signal is that “maintainable, replaceable” is becoming the default standard: self-hosted tiles replacing map APIs, firmware costs crushed by AI into a weekend, browsing and notes folding into Emacs — and short links’ historical citations biting back is the other side of the same coin: whoever you give control to can rewrite what you wrote.