Hacker News Evening Brief: April 12, 2026
This evening’s Hacker News mix was unusually broad even by HN standards: interface design manifestos, playful transit art, practical bootstrapping advice, security methodology fights, and a surprisingly large amount of frustration about AI product economics. The common thread was not hype so much as systems, what they optimize for, what they hide, and who pays when they get weird.
Tech Tools & Projects
Bring Back Idiomatic Design
Summary: John Loeber makes a simple argument that lands harder than it sounds: software used to share more interface idioms, and users benefited from not having to relearn basic interactions every time they opened a new app. He uses familiar controls like checkboxes and date pickers to show how standard patterns reduce thought, then contrasts that with a modern web where every team rolls its own components and every product manager wants visible differentiation. The piece treats this not as nostalgia for Windows 95 chrome, but as a critique of software that optimizes branding and conversion while quietly increasing cognitive load.
HN Discussion: The thread instantly collapsed onto one of the most annoying examples in real life, whether Enter, Shift+Enter, or Ctrl+Enter should submit text or create a new line. Commenters blamed everything from Electron and React Native to growth-driven UI experiments, and several argued that native frameworks once enforced good behavior simply by making the standard path the easy path. Date pickers were the other punching bag, especially ones that force clicking through months instead of letting users just type a date.
Show HN: boringBar – a taskbar-style dock replacement for macOS
Summary: boringBar is a macOS utility that swaps the Dock’s app-centric model for a taskbar that tracks windows, desktops, and displays more explicitly. It shows only the windows on the active desktop, adds thumbnail previews, desktop switching, notification badges, attention pulses, and a searchable launcher, and is clearly aimed at people who use multiple displays or have migrated from Windows and miss a more literal window list. The pitch is less about eye candy than about making macOS feel spatially legible again when Mission Control and the Dock stop scaling cleanly.
HN Discussion: Hacker News mostly agreed the product solves a real annoyance, then immediately revolted at the pricing model. Commenters said they could imagine paying once for a utility like this, or paying for major upgrades, but a subscription for a taskbar replacement felt absurdly fragile for long-lived desktop software. A few compared it to older Mac utilities like uBar and treated the business model, not the feature set, as the real problem.
Show HN: Oberon System 3 runs natively on Raspberry Pi 3 (with ready SD card)
Summary: This release packages Oberon System 3 so it can run natively on a Raspberry Pi 3, complete with ready SD card images and a much lower barrier to entry than the environment usually gets. That matters because Oberon is not just an old language curiosity, but a whole Wirthian ecosystem where language, operating system, and programming environment were designed together. Putting it on cheap commodity hardware turns it from historical literature into something you can actually boot, inspect, and program against in an afternoon.
HN Discussion: The comments turned into a brisk Oberon orientation session for people who knew the name but not the lineage. Readers asked how it differs from Smalltalk, Lisp machines, Java runtimes, or Inferno, and the author clarified that Oberon began as a native system for ETH Zurich teaching machines rather than an image-based environment. Others showed up mainly to reminisce, praising Oberon’s cozy programming environment and its influence on interfaces like Acme.
JVM Options Explorer
Summary: JVM Options Explorer is exactly what it sounds like and more useful than that description suggests: a searchable catalog of HotSpot VM flags with version metadata, defaults, ranges, classifications, and source-file pointers. Instead of spelunking raw OpenJDK headers or half-remembered blog posts, you can inspect everything from diagnostic crash flags to x86 AVX thresholds in one place. The effect is almost archaeological, because the tool doubles as a map of how much operational and historical sediment the JVM has accumulated.
HN Discussion: The big reaction was disbelief at the sheer volume of options. Some commenters treated the flag count as an indictment of complexity, while others saw it as the inevitable surface area of a VM that has had to satisfy performance engineers, production operators, and compatibility constraints for decades. There was also a quiet subtext that the explorer is valuable precisely because nobody can realistically hold this many switches in their head.
Doom, Played over Curl
Summary: curl-doom is a terminal trick that streams DOOM frames over HTTP, rendering them as ANSI half-blocks while sending player keystrokes back up the same connection. The clever part is not “DOOM over curl” as a slogan so much as the duplex design: one streaming request, two directions, raw terminal mode, no special client beyond curl and bash. The repository spells out how it negotiates content at /, why stty has to disable canonical input behavior, and how the shell wrapper restores the terminal when you’re done.
HN Discussion: A lot of the thread was pedantic in exactly the way these posts invite. People complained that the curl framing oversells the gimmick and that the real trick is terminal I/O over HTTP, then piled on some README wording they thought sounded AI-smoothed, especially around cooked versus canonical mode. Even so, most of the skeptics still seemed to respect the underlying single-connection stunt.
Show HN: Pardonned.com – A searchable database of US Pardons
Summary: Pardonned takes Department of Justice clemency records and turns them into a browseable database instead of a pile of barely navigable public documents. The site tracks pardons and commutations since Bill Clinton, exposes totals for prison time reduced and money abandoned in restitution or fines, and links each record back to the underlying warrant PDF. That makes it useful both as a civic reference tool and as a demonstration of how much value can come from merely making public records searchable and legible.
HN Discussion: Commenters immediately started using it like a real dataset rather than a showpiece. People asked for raw exports, suggested offense-type breakdowns by president, and caught edge cases like repeat clemency and the oddities of preemptive pardons, prompting the author to expose JSON and then the SQLite database itself. A recurring theme was that the DOJ should have shipped something like this years ago.
Other
I gave every train in New York an instrument
Summary: Trainjazz maps live New York subway traffic into an ambient jazz piece where each route gets an instrument and each train’s position determines when a note lands in the phrase. The whole thing is normalized so lines can play together continuously, with rush hour thickening the music and late-night gaps producing actual silence. It is a lovely example of city data being treated less as a dashboard and more as material for composition, especially since the mix can also shift around the listener’s own location.
HN Discussion: Readers loved the route personalities, the little instrument notes on hover, and the fact that the page does not over-explain itself before letting the gimmick work. One small debate broke out over why this sort of generative piece feels charming when many people recoil from AI music, with defenders arguing that expectation and intent matter as much as the resulting sound. The New Yorkers, naturally, also started evaluating whether the G train really sounds like the G train.
Happy Map
Summary: Pudding’s Happy Map presents “100,000 moments of human happiness” as a geographic visualization rather than a ranking, letting readers browse emotional data through clusters and place. The point is less to derive a universal metric than to turn a huge pile of happy moments into something explorable, tactile, and spatial. Like many Pudding projects, the piece leans on interaction and presentation more than textual exposition, so what sticks is the feeling of moving through a cartography of reported joy.
HN Discussion: Hacker News spent a surprising amount of time talking about the frontend. People reported performance differences across Safari and mobile, guessed at WebGL or deck.gl-style implementation choices, and argued about whether the technical execution outshone the underlying analytical value. It was one of those threads where the visualization stack became almost as interesting to HN as the subject matter.
A Tour of Oodi
Summary: This photo essay tours Helsinki’s Oodi not as a quiet warehouse of books but as a deliberately expansive civic building that mixes library functions with studios, meeting rooms, instruments, a cinema, game rooms, kitchens, and makerspace equipment. The article’s charm comes from its mounting disbelief that a public library can also contain recording booths, laser cutters, sewing machines, and places to simply hang out. By the time the books finally show up on the upper floor, Oodi has already been framed as a social technology for urban life.
HN Discussion: Locals mostly confirmed that the piece captures the building’s breadth, while adding that the actual feeling of the top floor is better in person than in photos. The real disagreement was over whether Oodi is a great library in the traditional sense or a highly successful community space that happens to include books as one service among many. One especially sour subthread attacked the spending on cultural amenities, which other commenters regarded as a miserable way to think about public institutions.
Most people can’t juggle one ball
Summary: Sean Herrington’s tutorial argues that beginners do not fail at three-ball juggling because three is too many, but because they never build a single reliable toss first. From there he walks outward toward the cascade, practice methods, equipment choices, and eventually siteswap notation, treating juggling as trainable mechanics rather than innate grace. The title’s provocation works because it shifts the unit of analysis from “can you keep objects in the air” to “can you repeatedly throw one object to the same place on purpose.”
HN Discussion: Experienced jugglers largely agreed with the framing, emphasizing that consistency of arc matters more than raw hand-eye panic. Several suggested slower-falling objects like handkerchiefs as better teaching tools because they reveal the path of the throw, while another fun branch admitted that plenty of people can sustain three-ball patterns only by constantly correcting errors. That admission, unsurprisingly, got repurposed as a metaphor for software systems.
The Miller Principle (2007)
Summary: The Miller Principle is a tiny 2007 post with one durable observation: no one reads anything. It applies the rule to user docs, specs, code comments, UI copy, and any email longer than a line, which sounds glib until you notice how often real systems still assume obedient readers. Its staying power comes from the way it compresses an enormous amount of design and communication pain into a single sentence.
HN Discussion: Commenters mostly responded with recognition rather than rebuttal. People traded examples of users ignoring headings, instructions, labels, warnings, and error text, then admitted that developers and managers behave the same way with docs and long internal emails. The practical lesson the thread drew was that defaults, affordances, and brevity matter more than carefully written prose that no one will consume.
Reading Is Magic
Summary: Sam Kriss uses a Jacobin essay on literacy as the jumping-off point for a broader argument that reading is not just content intake but a technology that restructures thought itself. Pulling in Walter Ong and Alexander Luria, he treats mental habits as social products, then returns to Luria’s Central Asian fieldwork to show how abstract hypothetical reasoning is learned rather than automatic. The result is an essay about literacy and politics that asks what changes when written language stops being the main architecture of public thought.
HN Discussion: The visible thread was small but pointed. Readers fixated on one of Luria’s famous examples, where villagers reject a mathematically neat travel-time puzzle because its premise contradicts what they know of the actual road. That pushed the comments toward a nice question: when does refusing an abstract premise reveal cognitive limitation, and when is it simply a sane refusal to play along with nonsense?
History & Science
Phyphox – Physical Experiments Using a Smartphone
Summary: Phyphox turns a phone into a pocket lab by exposing onboard sensors for experiments, then making the resulting data easy to export and inspect elsewhere. The examples are exactly the sort of thing that sells the idea, pendulum measurements via accelerometer, Doppler experiments via microphone, and browser-based remote control so the phone can sit in the apparatus while the student operates from a laptop. What it really offers is not just sensor access, but a well-packaged bridge between everyday hardware and classroom-grade measurement.
HN Discussion: Readers got practical fast. Commenters compared how Android and iPhone hardware differ in actual sampling behavior, pointed each other to Phyphox’s sensor database, and brought up F-Droid packaging and other distribution details that matter if you want to use the app in a lab. The mood was that this is one of those educational tools whose value becomes obvious the moment you hand it to a teacher or a curious teenager.
The Physics of GPS
Summary: This interactive explainer builds GPS from the ground up, starting with the almost comically simple idea that time-of-flight can be turned into distance because light travels at a known speed. From there it shows why one satellite only gives you a ring of possibilities, why bad clocks and bad geometry make life difficult, and why relativity is not a bonus complication but part of making the system work at all. It is a nice piece of science writing because it keeps translating impossible-seeming infrastructure back into geometry and stopwatches.
HN Discussion: The comments were full of references to older favorites, especially Bartosz Ciechanowski’s GPS explainer, which remains a touchstone for technically literate visuals. Another branch jumped from consumer GPS to RTK and VRS RTK systems, where correction data can push position accuracy down into the centimeter range. Readers also seemed to appreciate that the article earned its Einstein moment instead of starting with it.
Seven countries now generate 100% of their electricity from renewable energy
Summary: The Independent rounds up IEA and IRENA figures showing seven countries, Albania, Bhutan, Nepal, Paraguay, Iceland, Ethiopia, and the Democratic Republic of Congo, generating virtually all of their electricity from renewable sources. The underlying statistic is narrower than the headline can make it sound, since it covers electricity rather than total energy use, but it still provides a useful snapshot of where hydro and geothermal abundance have already pushed grids close to all-renewable operation. The article pairs the dataset with Mark Jacobson’s argument that decarbonization no longer waits on hypothetical miracle technologies.
HN Discussion: Hacker News immediately stressed the geography. Commenters noted that most of these cases are hydro-heavy or geothermal-favored outliers, often with small populations, imports, drought sensitivity, or low per-capita electricity use that make them hard to generalize from. Others pushed back that focusing only on the outliers misses genuine progress in larger places where solar, wind, interconnects, and storage are doing more of the work.
Academic & Research
Eternity in six hours: Intergalactic spreading of intelligent life (2013)
Summary: This 2013 paper takes the Fermi paradox and stretches it outward, arguing that if a civilization can spread through its own galaxy, then intergalactic expansion is not some categorically different impossibility. The authors walk through automation, Dyson-sphere-scale energy capture, replicating probes, deceleration, and collision risks to make the case that a sufficiently advanced society could colonize a huge reachable volume of the universe with less magic than people tend to assume. That, in turn, sharpens the paradox: if this kind of expansion is feasible, why is the sky still so quiet?
HN Discussion: The comment thread mostly played the familiar Fermi-paradox hits, but with the dial turned up. Readers argued over whether the real bottlenecks are physics, interest, coordination, or civilizational survival, and several compared the paper to older von Neumann probe scenarios where self-replicating exploration should have filled the cosmos long ago. The paper’s appeal on HN was less “wow, cool sci-fi” than the way it narrows the room for easy dismissals.
Floyd’s Sampling Algorithm
Summary: Alex Jaffray’s note on Floyd’s Sampling Algorithm is the sort of small algorithms essay that makes programmers happy, because the code is tiny and the underlying correctness still feels like a conjuring trick. The goal is to choose a size-k subset from 1 through n without replacement, and Floyd’s method does it with a loop whose odd little branch, add the random pick unless it’s already present, otherwise add the current index, resists immediate intuition. The essay works by trying to restore that intuition through multiple perspectives rather than pretending the magic is self-evident.
HN Discussion: This was one of the quietest stories in the batch. The Hacker News thread was effectively empty, so there was no meaningful argument to extract beyond a general sense that elegant algorithm lore still attracts clicks even when it does not attract debate. In a way that silence fits the piece: it reads more like something people save than something they litigate.
Security & Privacy
Compute iOS XNU offset from kernel cache
Summary: This reverse-engineering writeup starts from a point that exploit headlines often skip past: having kernel read and write is not the end of the job if you still do not know where anything useful lives. The post walks through finding the Mach-O magic in memory, deriving the kernel slide, and then recovering offsets for internal XNU structures from stripped kernelcaches, all without relying on symbols. That makes it less a splashy exploit release than a careful field guide to the grunt work of turning primitive access into actionable understanding.
HN Discussion: Hacker News had almost nothing to say about this one, at least in the visible thread. There was no real debate to summarize, which is a shame because the article itself is concrete and instructive. It is the kind of post that may have been read more than it was discussed.
Small models also found the vulnerabilities that Mythos found
Summary: AISLE’s response to Anthropic’s Mythos announcement argues that the frontier model is not the whole story in AI-assisted vulnerability discovery. The company reran Anthropic’s showcase vulnerabilities with smaller open-weight models, gave them scoped code plus context, and found that much of the same analysis could be recovered, which leads to the article’s central claim: the moat is in the system, the targeting, validation, triage, and operator expertise, not simply in owning the biggest model. The post is careful to say this does not fully settle the comparison, but it does try to puncture the idea that a gated cyber model has already proved an unassailable lead.
HN Discussion: This thread was lively because the methodology is exactly where the dispute lives. Critics said AISLE made the problem much easier by isolating the vulnerable region and feeding models hints, which is nowhere near equivalent to searching a full codebase for the dangerous function in the first place. Defenders replied that this is precisely what a good harness does, and that the real product is the system that scopes, tests, and ranks candidate findings rather than the raw model sitting alone in a prompt box.
Web & Infrastructure
Tell HN: docker pull fails in spain due to football cloudflare block
Summary: This HN-native post reports that docker pull was failing in Spain during court-ordered anti-piracy blocks aimed at football streams, with the apparent cause being broad blocking of Cloudflare infrastructure rather than anything Docker-specific. The complaint was valuable because it translated a policy story into a developer symptom: a familiar tool breaking for reasons that have nothing to do with containers, registries, or user error. In practice it reads as a case study in how blunt network enforcement spills over into unrelated technical life.
HN Discussion: Commenters quickly expanded the blast radius, listing other Cloudflare-backed services that also failed during the same windows. The tone was incredulous and angry, with several comparing the tactic to a national firewall and debating whether the real culprit was LaLiga’s pressure campaign, court orders, ISP implementation, or Cloudflare’s centrality. The thread was less about Docker than about what happens when anti-piracy infrastructure starts behaving like collateral censorship.
We have a 99% email reputation. Gmail disagrees
Summary: Font Awesome describes the stomach-dropping moment when it realized recent campaign emails were not bouncing or erroring, but simply disappearing into Gmail spam despite a glowing SendGrid reputation score. The post’s practical takeaway is that Gmail runs its own opaque reputation regime, and that being a well-behaved sender by one vendor’s dashboard means very little if the dominant mailbox provider has decided your mail belongs in the basement. For a small company, the piece is half deliverability autopsy and half reminder that email remains a privately governed maze.
HN Discussion: HN was not especially charitable. Many commenters argued that Gmail may be doing exactly what users want if people signed up for free icons rather than for broad marketing blasts, and that “good sender reputation” does not override recipient indifference or annoyance. Others took the post as proof that email remains hostile even to legitimate senders, because the real scoring logic is hidden and the consequences appear only after the campaign misses its audience.
Building a SaaS in 2026 Using Only EU Infrastructure
Summary: This guide tries to answer a practical version of a sovereignty question: can a startup build a modern SaaS without touching AWS, Azure, GCP, Stripe, Cloudflare, or Google Analytics? Its answer is mostly yes, and it backs that up with concrete recommendations for compute, billing, CDN, analytics, email, and other layers, with Hetzner, Scaleway, and Mollie standing in as examples of credible European substitutes. The article is useful because it treats “European stack” not as branding but as a checklist of real infrastructure decisions.
HN Discussion: Readers liked the anti-hyperscaler energy but immediately tested the weak joints. The hardest one was AI: if your product depends on OpenAI or Anthropic features, then a US dependency still sits in the middle no matter where your servers live. Another line of argument challenged whether some of the listed services are truly insulated from US legal reach, especially once US operations or US clouds enter the picture.
AI & Tech Policy
Anthropic downgraded cache TTL on March 6th
Summary: This GitHub issue is more than a complaint because the author comes with receipts. By mining local Claude Code JSONL session logs across two machines, the post argues that Anthropic silently shifted prompt caching behavior in early March from an effective one-hour default to a five-minute regime, and that the result was a meaningful increase in cache-creation costs and quota burn for users who had previously found the product economically stable. The article is strongest where it is most mundane: phase tables, observed token fields, and a date where behavior appears to change at the same time on multiple systems.
HN Discussion: The comments treated the TTL evidence as part of a wider collapse in trust around AI coding subscriptions. Users folded in complaints about hidden nerfs, shorter responses, lower reasoning effort, abrupt quota exhaustion, and the sense that paid products are being tuned in secret to control compute costs. At the same time, people zoomed out and argued this may simply be the end of the subsidized era, with every major coding-assistant vendor now discovering that the original price-to-usage bargain was never going to last.
Tell HN: OpenAI silently removed Study Mode from ChatGPT
Summary: This Tell HN post is modest on its face, one user noticing that ChatGPT’s Study Mode seems to have vanished, but it hit a nerve because it slots neatly into a broader pattern of AI features appearing, changing, and disappearing without much explanation. Since Study Mode was framed as a guided-learning surface rather than just another personality preset, its disappearance felt to users like the loss of a real workflow rather than a decorative experiment. The post is thus less about one toggle than about the instability of AI product surfaces.
HN Discussion: Many commenters suspected Study Mode was largely prompt engineering in a nicer wrapper, which made its removal feel technically trivial but product-wise revealing. Others guessed the feature might have been withdrawn because it complicated the UI, leaked too much about prompting strategy, or created safety or pedagogy headaches. In the background was a broader complaint that ChatGPT keeps changing shape faster than paying users can rely on it.
Pro Max 5x quota exhausted in 1.5 hours despite moderate usage
Summary: This earlier GitHub issue makes a narrower but more user-facing complaint than the TTL analysis above: a Claude Pro Max 5x quota window allegedly vanished in 90 minutes after reset despite mostly moderate Q&A and light development work. The reporter compares that window with a prior five-hour heavy-development session, then proposes that cache-read tokens are counting too aggressively against quota, which means prompt caching helps on paper while still chewing through the plan’s practical limit. The issue is useful because it ties a vague feeling of “this suddenly got expensive” to measured token categories pulled from local session logs.
HN Discussion: Anthropic staff replied with a concrete diagnosis, stale sessions, million-token contexts, and plugin-heavy or multi-agent workflows can all make usage spike unexpectedly, and suggested UX changes plus smaller default context windows as mitigation. Many users were unconvinced, arguing that the real problem is opacity: if a plan can become uneconomical this quickly, the interface should scream about why instead of quietly draining a quota meter. The thread then widened into a consumer comparison of Claude, Codex, Gemini, and other coding tools on the only metric that currently matters, cost per hour of actually useful work.
Business & Industry
I run multiple $10K MRR companies on a $20/month tech stack
Summary: Steve Hanov’s essay is an anti-boilerplate manifesto for small software businesses. Instead of treating a startup as a miniature FAANG deployment, he argues for single VPS boxes, tiny monthly bills, SQLite where possible, and architecture driven by actual demand rather than by imagined future scale. The business point is as important as the technical one: low burn buys the same runway as a lot of fundraising, but without the stress or the temptation to build an infrastructure org before you have a business.
HN Discussion: Readers mostly agreed with the diagnosis and then fought over the implementation details. The liveliest argument was SQLite versus Postgres, especially whether local Postgres over Unix sockets retains most of the simplicity while preserving operational options for later growth. Others zoomed out to say the article gets the economics right but undersells the boring disciplines, backups, SSH hardening, off-site storage, and recovery planning, that make cheap infrastructure survivable.
An Interview with Pat Gelsinger
Summary: This interview catches Pat Gelsinger after his Intel chapter, now speaking from Playground Global while still orbiting the same themes that defined his career: compute, manufacturing, leadership, and the shape of the next hard-tech bets. The setup matters because the conversation is colored by Intel’s attempted turnaround and by the fact that Gelsinger is no longer the one carrying it out. Rather than a pure memoir, the piece reads as an attempt to place him in continuity, from Intel’s older CPU era to today’s foundry, AI, and systems questions.
HN Discussion: Hacker News was less interested in the self-description than in the gaps. Commenters kept circling back to the question the interview does not really answer, why Intel pushed Gelsinger out when some of the technical roadmap he backed now looks more plausible than it did at the time. The replies split sharply between people who think he was vindicated by products like Panther Lake and process work like 14A, and people who think his tenure was financially reckless and strategically late to every important boom.
Cirrus Labs to join OpenAI
Summary: Cirrus Labs announced it will join OpenAI’s Agent Infrastructure team, framing the move as an extension of the company’s original mission to build tools and environments that make engineers more productive. The post recaps its history from Cirrus CI to Tart, the Apple Silicon virtualization tool that became particularly valuable in a world where macOS automation and CI remain awkward. The announcement is notable not just for the team move, but because it also says Cirrus’s source-available tools will be relicensed more permissively while SaaS offerings are wound down.
HN Discussion: Readers quickly interpreted the deal as an acquihire plus infrastructure grab rather than a normal product acquisition. Tart and Cirrus’s Apple virtualization expertise looked like the obvious prize, especially for an AI company that wants better macOS environments for agents. The tradeoff is that people who genuinely liked Cirrus CI now have to figure out what replaces it, and the comments were full of exactly that kind of slightly mournful migration planning.
‘The Audacity’ Is the Broligarchy Takedown You Were Waiting For
Summary: WIRED’s review treats AMC’s The Audacity as a badly needed satire of the modern tech titan, with Billy Magnussen’s Duncan Park standing in for the blend of entitlement, self-justifying aphorisms, and ideological shallowness that has become familiar around Silicon Valley power. The review is more interested in the type than in the plot, presenting the show as a skewering of a class that wraps immaturity and impunity in the language of disruption. It is a cultural review, but one that clearly assumes readers already know the real-world targets.
HN Discussion: HN did not produce much close reading of the show itself. What discussion there was split between practical access complaints, some readers cannot watch AMC and immediately posted archive links, and a brief argument over whether tech billionaires really are as thin-skinned and hypocritical as the review suggests. Silicon Valley Bank bailout talk made a cameo as evidence that the satire may not be overshooting by much.
Geopolitics & War
Internet outage in Iran reaches 1,008 hours
Summary: NetBlocks reports that Iran’s internet blackout had reached 1,008 hours, entering day 43, and describes it plainly as a regime-imposed shutdown with growing economic and human costs. The accompanying graph shows connectivity collapsing around February 28 and remaining near the floor, with only a brief restoration in mid-March. There is not much article structure here because the source is a network-monitoring post, but the starkness of the numbers and the chart does most of the work.
HN Discussion: Commenters argued over the strategic meaning of the blackout as much as the technical fact of it. Some saw the outage as evidence of a regime terrified of its own population and trying to break protest coordination; others argued that in wartime a disconnected country also becomes harder for adversaries to exploit digitally, even if ordinary civilians bear the cost. A more technical side debate asked whether privileged “white SIM card” access changes anything when national connectivity has dropped that far.
Hacker News on a Sunday evening often drifts toward toys and essays, but this batch also had a clear undercurrent of systems becoming legible in unpleasant ways. Cloud blocks turned into broken Docker pulls, sender reputation turned into invisible Gmail spam, AI pricing turned into hidden cache behavior, and even the lighter posts often came down to whether a system teaches people gently or makes them reverse-engineer its quirks. That’s probably why the most cheerful stories tonight, subway jazz, Oodi, a phone becoming a physics lab, felt so refreshing: they were the rare systems here that seemed designed to invite people in.