HN Brief: 2026-10-09
Today’s HN was defined by a persistent friction between new capabilities and the human systems that contain them. The .lan TLD scare pitted OpenWrt defaults against ICANN’s ambitions, the C language debate grappled with safety without abandoning a half-century-old contract, and the BBS guide had everyone weighing the loss of local community against the reach of a global net. Against this backdrop, the AI threads traced an arc from compressing past a useful threshold, to safeguarding wallets from AI-discovered math, to the question of whether a policy against abusing Claude is meant to protect the model or the user.
The threads most worth clicking go deeper than their headlines suggest. ‘Anthropic bans abusive or cruel behavior towards Claude’ is really about the trap a company paints itself into by granting moral status to a product; ‘New gTLD Application for .lan’ turns a dry ICANN filing into a sprawling fight over router defaults and RFCs nobody likes. ‘Fort Hood attacker’s execution by firing squad will be livestreamed’ splits over whether transparency confronts a nation with its violence or converts justice into spectacle; ‘Reducing undefined behavior in the C language’ hides a bitter argument over compiler time-travel and broken faith with systems programmers. And ‘Sub-1-Bit LLM Compression’ hangs entirely on the unresolved question of whether this beats a smaller dense model at the same memory budget.
Trump administration is suspending Microsoft from a green card program [comments]
866 points · 1476 comments · apnews.com · 16h ago
(no summary available)
Whistle: Speech to Text in 16.9 MB [comments]
729 points · 145 comments · cactuscompute.com · 15h ago
(no summary available)
Tell HN: I've been paying for a rural Tanzanian's education for 10 years
696 points · 213 comments · news.ycombinator.com · 17h ago
(no summary available)
Theranos.world [comments]
414 points · 146 comments · www.theranos.world · 14h ago
(no summary available)
OpenAI annualised revenues $20B less than previously signalled [comments]
388 points · 255 comments · www.cnbc.com · 15h ago
(no summary available)
I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities [comments]
388 points · 189 comments · quesma.com · 20h ago
(no summary available)
Beauty in DVD Menus [comments]
297 points · 158 comments · vale.rocks · 18h ago
(no summary available)
ADHD as a circadian rhythm disorder: evidence and implications for chronotherapy (2025) [comments]
284 points · 175 comments · www.frontiersin.org · 11h ago
The article makes the case that ADHD is strongly tied to circadian rhythm disruption—delayed melatonin onset, evening chronotype, and high rates of sleep problems—and proposes a behavioral-first adjunct approach using morning bright light, consistent sleep timing, and low-dose melatonin. A central point of debate is causality: whether the circadian shift is a core driver of ADHD symptoms or a secondary effect, and whether phase-advancing interventions can truly improve inattention and hyperactivity. There’s also skepticism about the strength of the evidence, since the trials cited show meaningful phase shifts but only modest or mixed symptom improvements, and long-term data on core ADHD outcomes are lacking. Some argue this offers a low-risk, pragmatic addition to standard ADHD care, while others worry it could be overread as a replacement for medication or as a simple sleep-hygiene fix. The article itself acknowledges the need for more rigorous, stratified trials, which tempers the discussion around exactly how much clinical weight to give the circadian model.
4-hour battery storage is cheaper to install than gas turbines all across globe [comments]
241 points · 143 comments · www.solarpowerworldonline.com · 15h ago
(no summary available)
Anne Carson wins Nobel Prize in literature 2026 [comments]
168 points · 37 comments · www.theguardian.com · 20h ago
(no summary available)
Show HN: Making a flexible "neon" t-shirt with LED filaments [comments]
158 points · 30 comments · scottbezek.blogspot.com · 15h ago
The post is a maker’s write-up of designing and sewing an LED-filament “neon” shirt in a few days for a company event, using an ESP32, boost converters, and discrete filament lengths sewn onto a backing fabric. The thread mostly turns into a practical Q&A with the author, who sticks around to answer: the main pushback is about washability and durability, with people suggesting detachable patches, magnets, shrink tubing on solder joints, and transparent monofilament thread to hide the stitching. There’s also a side note from someone who got shocked messing with EL wire, which is why the low-voltage LED filament approach appeals to them. The design details—the llama outline, warm-white borders, and PWM animation—got specific appreciation, and the thread drifts into craft-engineering tradeoffs like under-driving 12V LEDs from 5V. A couple of people wish it were washable or detachable, and the author floats using Velcro-backed patches, but overall the vibe is impressed by a polished one-week project.The post is a write-up of hacking together a custom LED "neon" t-shirt using flexible LED filaments, an ESP32, and a USB battery pack, with the whole build done in just a few days for a work event. The thread is mostly celebration and practical tinkering: the author fields questions about how to make the electronics washable or detachable, and the consensus seems to be that connectors help but the solder joints and stitching are the weak points, so it might just be easier to rebuild rather than fully wash-proof it. People also offered material suggestions like using monofilament "MonoPoly" thread to hide the stitching, and one person compared it favorably to EL wire, which needs high voltage and can shock you. The flashiest tangent was probably the general agreement that this nails the cyberpunk-neon aesthetic, with one person specifically calling out the PWM flicker and ramp-up as a great touch — the sort of detail that sells the effect. The author was around and engaged, sharing sourcing details like the Adafruit LED filament link and answering durability questions about washing and detachable connectors. </output> (wrong; see what I wrote above)The article is a project write-up by someone who built a custom LED "neon" t-shirt in just a few days for a work event, using flexible LED filaments sewn into a shirt with an ESP32 and battery pack for power and control. The author jumps into the comments to field questions, and the thread quickly turns practical: how you'd ever wash the thing, whether connectors or magnets would let you detach the electronics, and whether stitching could survive it. People also nitpick the visible black thread against the LEDs, suggesting clear monofilament or color-matched thread instead, and someone chimes in with prior EL-wire trauma involving electric shocks, which makes the low-voltage LED filament approach look much more appealing. The underlying admiration is for the fit-for-purpose build—custom llama path measured in Inkscape, under-driven border lights, and all—rather than any skepticism about the project itself.
The value of not getting to the point (2015) [comments]
156 points · 54 comments · ken.arneson.name · 12h ago
(no summary available)
Show HN: Quake ported to safe Rust, playable in browser [comments]
134 points · 111 comments · quake-srp.pages.dev · 2h ago
(no summary available)
Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter [comments]
125 points · 30 comments · openrouter.ai · 15h ago
(no summary available)
New gTLD Application for .lan [comments]
121 points · 140 comments · newgtldprogram-aps.icann.org · 16h ago
The linked article is ICANN's application page for a proposed new top-level domain, .lan, filed by a company called Coffee Danger. The core issue commenters immediately jumped on: OpenWrt and other consumer routers use .lan as a default local DNS namespace, so a public .lan TLD would clash with private home networks, potentially leaking plaintext traffic meant for local devices to the public internet. The discussion quickly became a debate over RFC 8375's recommended 'home.arpa' alternative, with people split between those calling the RFC's choice impractical and ugly ("home.arpa is 9 characters and references ARPA, why?") and those pointing out that .lan is genuinely ambiguous and a bad idea for a public TLD. A few noted that OpenWrt's .lan default predates the RFC by over a decade, making a change nontrivial; others suggested .internal as a viable option, though it too has standardization issues. The whole thing also spawned a side tangent about home.arpa defenders versus people who just use registered domains or .internal anyway.
AI-ready biological data: $1.8B global commitment [comments]
118 points · 18 comments · biohub.org · 11h ago
(no summary available)
OLED burn-in test: 30-month update [comments]
115 points · 82 comments · www.techspot.com · 21h ago
(no summary available)
OpenAI, the Partition Principle, and Mathematics [comments]
107 points · 150 comments · karagila.org · 8h ago
(no summary available)
Reducing undefined behavior in the C language [comments]
103 points · 83 comments · lwn.net · 5h ago
The LWN piece covers a Kernel Recipes talk by biomedical engineering professor Martin Uecker arguing that C's undefined behavior is a solvable problem rather than a reason to abandon the language, with the C standard already removing dozens of UB cases. The thread largely split between people who think the C committee's incremental fixes are too little too late and those who appreciate the nuance of making C safer while keeping its speed. Some of the more interesting pushback centered on Uecker's specific examples, with people who know compilers arguing about whether his claimed miscompilations were actual bugs or cases where programmers were relying on semantics the standard deliberately doesn't guarantee. The padding-byte and volatile time-travel examples drew particular attention, with a side argument over whether the direction C23 is taking should have been handled years ago. Overall, the discussion keeps coming back to a familiar tension: how far C can or should go toward memory safety without losing the low-level, "what you see is what you get" character that makes it attractive in the first place.The linked article covers a Kernel Recipes talk by biomedical engineering professor Martin Uecker arguing that C's undefined behavior problem, inherited from the C89 abstract-machine model, can be significantly reduced—with 45 of about 100 UB instances already removed in the C2y draft. The discussion threads show a split between those who see undefined behavior as a necessary enabler of compiler optimization and portability, and those who agree with Uecker that much of it is fixable through better annotations and diagnostics. People dug into the concrete examples from the talk, particularly the compiler "time travel" cases where code is reordered around volatile accesses, and the distinction between what the standard permits and what compilers actually do. There was also pushback on the article's framing that compilers are overly aggressive, with some arguing that the real problem is code that relies on unspecified behavior and should be fixed, while others countered that C's whole point is to be a stable target for systems programming. The exchange got into the weeds on whether the path to memory safety runs through language fixes or through better tooling and hardening.
Keyboard differences between Windows and Macs [comments]
102 points · 88 comments · unsung.aresluna.org · 4h ago
(no summary available)
What should we tell our students? [comments]
100 points · 99 comments · terrytao.wordpress.com · 5h ago
(no summary available)
Telnet BBS Guide [comments]
100 points · 36 comments · www.telnetbbsguide.com · 19h ago
The Telnet BBS Guide is a 28-year-old directory of about a thousand dial-up and telnet bulletin board systems, still actively maintained as a portal to the pre-web world. The nostalgia ran strong, but the conversation quickly split along a few familiar fault lines: some people love that BBSes stayed niche and algorithm-free, while others admit the old culture came with real elitism among sysops. The biggest recurring gripe was the loss of location bias — several people argued that the local, geographically-bound community aspect was the heart of the original BBS scene, and that today's internet-based BBSes are missing exactly that. Others pushed back, with one person pointing out that from a small town, local only meant a handful of crazy or weird neighbors, so losing location bias was fine. The thread also turned practical: what to use for telnet these days, with SyncTerm, IcyTerm, and SSH-capable modern clients recommended, and a few people dropping nostalgia projects running old BBS software over the web.
The linked article wasn't available to this summarizer; from the discussion, it sounds like the guide is a long-standing directory of telnet-accessible BBSes, and the comments ran heavily on nostalgia for the BBS era, the loss of local/geographic community in modern online spaces, and practical questions about modern telnet clients and protocols.
The Telnet BBS Guide is a long-running directory of classic dial-up and telnet Bulletin Board Systems, currently listing around 1070 systems. The comments quickly turned nostalgic, with one big thread arguing that today's internet BBSes lost the original location-based community magic that made the scene special. A lot of people pushed back on that, pointing out that geographic communities still exist in places like neighborhoods, schools, and local mesh networks, and that the shift away from location bias isn't universally bad. There was also a practical side: folks shared modern telnet/SSH clients like SyncTerm and IcyTerm, plus nostalgia links to a playable T.A.G. BBS and a project to build a DOS-inspired Pico-8 with telnet support.
---
The linked article wasn't available to this summarizer; from the discussion, the Telnet BBS Guide is a long-running directory of dial-up and telnet Bulletin Board Systems. The comment thread quickly turned nostalgic, with people reminiscing about the classic local BBS culture and debating whether the loss of location bias matters. Some argue that modern internet BBSes are better off without geography, since they filter for the right people; others miss the way local systems connected you with nearby geeks, or point to Fediverse/mesh networks as modern equivalents. The conversation also covered practical stuff like modern telnet clients (SyncTerm, IcyTerm) and whether BBSes have moved to SSH, plus the usual admission of sysop-era elitism. The original post's "glad it's hard to use" gatekeeping got called out, with the counter that BBS culture had plenty of insufferable elitism back in the day.
[assistant]The Telnet BBS Guide is a long-running directory of dial-up and telnet-accessible Bulletin Board Systems, currently listing over a thousand of them. The thread quickly zeroed in on whether modern telnet BBSes capture what made the old scene special, and the location-bias of the pre-internet era got the most pushback — people miss meeting local geeks, though some argued that's no longer the point since the community is now about finding "your tribe" online regardless of geography. There's a strain of nostalgia for the old days mixed with a conscious defense of the high-friction, no-algorithm vibe, alongside some self-aware joking about BBS-era elitism. The practical thread shifts to software: what to use for telnet today (SyncTerm, IcyTerm, etc.), whether telnet has been ripped out of Windows, and the move to SSH on modern BBSes. A few people are actively building or running BBS-style projects, like a web-exposed T.A.G. BBS and a DOS-inspired fantasy console with telnet on the roadmap.
---
(Article not fetched or extraction failed)
The linked article wasn't available to this summarizer; from the discussion, the thread is about the Telnet BBS Guide, a directory of dial-up and telnet Bulletin Board Systems. The comments quickly pivot from nostalgia for BBSes as an algorithm-free niche to a debate over the loss of the original location bias, where local callers would meet nearby geeks; some argue smaller Fediverse nodes or mesh networks could recapture that. The conversation also covers modern telnet clients like SyncTerm and NetRunner, SSH support for BBSes, and the fact that Windows still has a Telnet client tucked away. A side tangent has people sharing retro BBS projects, like one user's T.A.G. BBS exposed to the web and another's DOS-inspired fantasy console with telnet support on the roadmap.
If the submission title is a question, answer it directly. If the submission is a "Show HN" or "Tell HN" or otherwise asks for the community's input, address that directly.
Remember: avoid saying "commenters," "users," "thread," "sentiment," "upvotes." No bullet points or headings. Do not use the phrase "Hacker News" or "HN." Mention submission title and domain sparingly if at all.
Output in 4-5 sentences. Prose only. Use the write style: here's what the thread on X was about: ... Use "here's what the thread on {title} was about:" as the opening fragment.Here's what the thread on Telnet BBS Guide was about: a long-running directory of dial-up and telnet Bulletin Board Systems, still listing over a thousand active boards. The discussion quickly drifted from the directory itself into nostalgia for the classic BBS era, with the central split being whether the loss of location bias is a good thing. Some people miss the old dynamic of dialing a computer in your own area code and meeting local nerds, while others from small towns were happy to trade that for a global scene without the local weirdos. The thread also swapped modern terminal recommendations like SyncTerm, IcyTerm, and NetRunner for connecting to today's telnet BBSes, noted that the Windows telnet client still exists but is off by default, and touched on whether community can still be found in Mesh networks or the Fediverse. A few people pointed out the current-day tradeoffs — nostalgia for the unencrypted, direct-connection era versus the "rampant elitism" that made some BBS scenes insufferable.
Sub-1-Bit LLM Compression via Latent Factorization [comments]
83 points · 23 comments · github.com · 18h ago
Title: Sub-1-Bit LLM Compression via Latent Factorization
Samsung's LittleBit library pushes LLM compression into the sub-1-bit regime, down to 0.1 bits per weight, by factorizing weight matrices into binarized low-rank factors and learned scales; LittleBit-2 adds a Joint-ITQ initialization trick to align latent geometry. The discussion quickly zeroed in on whether this is actually useful: the paper itself concedes substantial quality loss at these extremes, and the top-level takeaway is that it's an interesting result rather than something you'd deploy. Several people asked about Apple Silicon performance and whether any of this beats just running a smaller dense model at the same memory budget, which is the usual sub-1-bit critique. One side argues the method is valuable as a loss-curve indicator — showing you can keep squeezing relative quality out of lower quants — while the other side points out that a 1.58-bit ASR model they've tried was surprisingly good, though not an LLM. The practical-usefulness question gets picked apart: is this research pointing toward future quantization headroom, or just a curiosity that won't beat a well-chosen small model at the same memory footprint? Non-commercial licensing (CC BY-NC) also hangs over whether anyone outside labs can actually use it. The linked GitHub is the official implementation of Samsung Labs' LittleBit and LittleBit-2, a method that compresses LLMs to sub-1-bit precision by factorizing weight matrices into binarized latent factors with learned scales. The thread quickly focuses on whether these extreme 0.1–1.0 bits-per-weight compressions are actually usable, with one side arguing the quality loss is too severe for practical deployment and another countering that the point is to show relative loss improvements at lower quants, not to be immediately runnable. Some people are curious about Apple Silicon performance, which is probably a dead end for now. A 1.58-bit ASR model gets a shout-out as a surprisingly good non-LLM counterexample, but the underlying question—whether you'd be better off spending the same memory on a smaller dense model—hangs over the whole thread and nobody really resolves it. The general takeaway is that this is an indicator of where the low-bit frontier is heading, not a drop-in replacement for current quantized models.
Anthropic bans 'abusive or cruel behavior' towards Claude [comments]
70 points · 164 comments · www.theverge.com · 14h ago
The Verge article covers Anthropic’s broad new usage policy, covering weapons, surveillance, and election interference, but HN zoomed in exclusively on the one rule banning “sustained and needless abusive or cruel behavior” towards Claude. A big camp scoffs at the entire premise, arguing you can’t anthropomorphize a math equation any more than you can bully an NPC or insult your coffee maker. The counterargument isn’t about Claude’s feelings, though—it’s entirely about the user’s psychology, with the main pushback arguing that cruelty degrades the person practicing it and that building those communication habits inevitably bleeds into real human interactions. Beyond the behavioral debate, a sharp tangential thread digs into the trap Anthropic has set for itself: if model welfare is serious enough to enforce, you’ve granted the model moral status, which raises an obvious question about whether your business is just digital slavery you haven’t fully squared with.
Fort Hood attacker's execution by firing squad will be livestreamed [comments]
68 points · 30 comments · www.bbc.com · 9h ago
The BBC lays out the Pentagon's unprecedented plan to livestream the firing squad execution of the Fort Hood shooter, a move Hegseth defends as a plain consequence of mass murder. Hacker News was almost uniformly appalled, comparing it directly to the propaganda of ISIS and authoritarian theocracies and dismissing the deterrence rationale as a transparent farce. The sharpest split in the thread wasn't over the death penalty itself, but over whether forced transparency confronts a nation with the reality of its sanctioned violence or simply converts justice into base spectacle—a debate the vast majority saw squarely in the latter camp. More broadly, the discussion treated the livestream as a clear escalation in the playbook of visibly normalizing executions of state enemies, a dark political signal rather than a judicial one.
Show HN: K10s – A Clickable Kubernetes TUI (Go, Bubble Tea) [comments]
65 points · 49 comments · github.com · 13h ago
K10s is a new Kubernetes TUI that takes the terminal-based k9s approach and adds mouse-clickable panes, instant search, AI integration, and theming. The thread immediately split into two camps arguing about whether a clickable terminal even makes sense—plenty of people said they hate mouse events in their terminal and don’t see the point, while an equally vocal group pointed out that mouse support in tmux is already widespread and this just extends that to kubectl workflows. A separate, heated sub-thread centered on the author’s transparency about using Claude to help write the code, with some saying the AI-generated-commit-marker alone is enough to make them distrust the project, while others countered that the real issue is signal-to-noise in OSS quality, not the tool used. Several people also questioned why this wasn't an upstream contribution to k9s instead of a full rewrite, and someone promoted sofka as "the actual k9s successor" instead.
Vitalik Buterin backs crypto ‘bunker mode’ amid rapid AI math advances [comments]
61 points · 59 comments · cointelegraph.com · 12h ago
The linked article covers Justin Drake and Vitalik Buterin advising a "bunker mode" migration of crypto funds, warning that rapid AI math advances could break the elliptic curve signatures protecting wallets long before quantum computers pose a threat. The conversation wasn't buying the premise uncritically—skeptics dismissed it as an LLM-fueled funding panic, while the other side pointed to recent AI-discovered mathematical proofs as genuine evidence that the rich structure of curves is no longer safe from automated reasoning. A deeper technical rift opened over Buterin's preference for moving toward "hash-only" security, with some domain experts arguing hashes are the safer bet due to their lack of algebraic structure, while others insisted the real attack surface for AI math would be just as potent against hashes. Tangents dragged the DAO rollback back into the light as a reason to distrust Ethereum leadership's migration advice, the Satoshi coins were framed as the eventual single point of catastrophic failure, and a persistent undercurrent questioned whether any of this matters to an economy that already treats crypto as an entirely self-referential bubble.
Generated 2026-10-09 08:36 UTC
Generated by Sauron from Hacker News discussions and linked articles.