HN Brief: 2026-07-10
Today’s HN was a stew of procedural outrage and language tribe warfare. The EU’s Chat Control 1.0 passed via a parliamentary trick that made rejection harder than approval, igniting a furious debate about democratic erosion and surveillance—with a bitter irony that Europe touts GDPR while legalizing mass message scanning. The Bun Rust rewrite post turned into a raw personality clash, with Andrew Kelley blaming Jarred’s management and “slop” engineering, splitting the room between those who saw a necessary callout and those who read it as a bitter personal attack. OpenAI’s GPT-5.6 release gave the Codex-versus-Claude Code war fresh fuel, while a quieter throughline about minimalism versus bloat emerged in reactions to LazyPi’s all-in-one config and Ghostty’s feature trade-offs.
Threads most worth clicking into: “EU Parliament greenlights Chat Control 1.0” for the procedural trick that let a minority pass a surveillance law everyone thought they’d blocked; “My thoughts on the Bun Rust rewrite” for a rare, unfiltered founder-vs-founder takedown that gets into code quality, VC culture, and personality; “GPT-5.6” because the real action is in the comments war over which coding agent actually works; “Show HN: Getting GLM 5.2 running on my slow computer” for a clever disk-streaming hack that gets 0.05 tokens per second—and the debate over whether that’s useful or just a proof-of-concept; and “Why developers are ditching GitHub for Codeberg” because the article’s headline is oversold but the comments surface real pain around wrongful bans, rate-limiting, and Microsoft’s ICE ties that make the exodus feel less hypothetical.
EU Parliament greenlights Chat Control 1.0 [comments]
1314 points · 613 comments · www.patrick-breyer.de · 21h ago
The EU Parliament has allowed suspicionless mass scanning of private messages (Chat Control 1.0) to pass until 2028, even though a majority of voting MEPs actually opposed it — the motion to reject failed because it needed an absolute majority, thanks to a procedural trick from the Parliament president. The thread zeroed in on that procedural maneuver as a blatant democratic failure, with people pointing out that this is how the EU launders unpopular laws by making rejection harder than approval. A heated split emerged over whether “slippery slope” is a fallacy or a valid warning here, given the EU’s own data showing the scanning hasn’t increased convictions or rescued children. Others noted the bitter irony of Europe touting strong privacy laws like GDPR while simultaneously legalizing warrantless message scanning by US tech giants, making the distinction between government and corporate surveillance feel meaningless. The consensus was that this was less about protecting children and more about Big Tech’s data access and state overreach, with the procedural farce damaging trust in the EU project itself.
GPT-5.6 [comments]
1221 points · 873 comments · openai.com · 15h ago
OpenAI released GPT-5.6, a family of three models (Sol, Terra, Luna) alongside a lengthy safety card claiming improved cybersecurity capabilities and more robust safeguards. Most of the HN discussion immediately bypassed the safety analysis to debate the naming gimmick — people are tired of the decimal creep — and then dove headfirst into a full-blown war between Claude Code and Codex (OpenAI's coding agent). The pro-Codex side argues it's more reliable, less prone to weird restrictions and outages, and doesn't lock you into a moat; Claude Code users counter that Anthropic's model is more proactive and creative for complex tasks. A vocal contingent insists you should just use a model-agnostic harness like OpenCode or Pi and never get locked into either ecosystem. There’s also a persistent undercurrent of skepticism about benchmark graphs, with people calling out deceptive y-axis scaling and pointing out the GPT-5.6 gains over the previous model are marginal at best.
Show HN: 18 Words [comments]
964 points · 315 comments · 18words.com · 19h ago
The linked article wasn’t available to this summarizer; from the discussion, "18 Words" is a daily anagram game where you get 30 seconds per word to unscramble six letters, aiming for 18 correct in a row. The thread split sharply on the timer: many people found it stressful and requested a relaxed mode, while others argued the pressure is the whole point, with some comparing it to timed developer screens they hate. A substantial chunk of the comments was about the word list — players reported valid English words like "tase," "egret," and "braze" being rejected, and the creator acknowledged the issue and reverted to a larger dictionary after saying they originally trimmed it to avoid obscure spamming. Other tangents included a shuffle button request (debated whether it would make brute-forcing too easy), a brief moment of alarm because the number 18 echoes the neo-Nazi "14 Words" slogan, and a deeper observation that the game feels hard because common words can suddenly become invisible to your brain under repeated 30-second pressure.
My thoughts on the Bun Rust rewrite [comments]
691 points · 599 comments · andrewkelley.me · 22h ago
Andrew Kelley’s blog post argues that the real problem with Bun’s Rust rewrite was never about language features—it was a relationship breakdown and a code-quality disaster, with Kelley calling out Bun’s founder Jarred for poor management, “slop” engineering, and a VC-driven culture that clashed with Zig’s long-term vision. The HN thread split hard: a lot of people agreed with Kelley’s diagnosis, saying his emotional tone matched what they’d seen from Jarred’s own public communications, while others felt the post was a thinly veiled personal attack that undermined Kelley’s claim of “no personal criticism”—they pointed to lines like “productivity fantasy fever dream” and argued the whole thing read as bitter and passive-aggressive. A big chunk of the debate circled around whether Rust’s safety guarantees would have actually prevented Bun’s bugs, with one side saying “don’t make mistakes” is useless advice and the other insisting that sufficient engineering resources—which Kelley advocated for—are what really matter, regardless of language. There was also a surprising lack of consensus on whether Bun even matters in production: some commenters insisted it’s niche and rarely used outside hype, while others (including people running sites with millions of users) pushed back and said it’s a flat improvement on Node.js for their workflows.
Show HN: Getting GLM 5.2 running on my slow computer [comments]
605 points · 142 comments · github.com · 23h ago
The article describes Colibri, a single-file C engine that runs the 744-billion-parameter GLM 5.2 model on a laptop with just 25 GB of RAM by keeping only the dense layers in memory and streaming the 370 GB of expert weights from disk on demand. HN was largely charmed by the sheer hacker spirit of the project—people compared it to antirez’s similar work on deepseek and called out the clever use of disk I/O as a genuine alternative to expensive GPUs—but the practical discussion quickly zeroed in on whether 0.05–0.1 tokens per second is actually useful for anything beyond an overnight batch job. Several people pushed back that the approach will chew through SSD write endurance, though the README addresses that directly, and a split emerged between those who see this as a fun proof-of-concept and those who think you’d be better off just using a free cloud API. A notable tangent spun off about whether chat is the wrong interface for such slow models, with multiple people suggesting a ticketing or email-gateway model as a more natural fit for sub-1-tok/s inference.
Hy3 [comments]
452 points · 91 comments · hy.tencent.com · 16h ago
The linked article wasn't available to this summarizer; from the discussion, Tencent released Hy3, a new open-source AI model that punches well above its weight class at 295B parameters. The main split in the thread is whether Hy3 genuinely competes with frontier models like DeepSeek V4 Pro and GLM 5.2, or if its impressive benchmarks are just "benchmaxxed" results that don't hold up in real-world use. Several people who actually tried it say it's remarkably good for its size and price—one person found it stayed on track during coding sessions far more reliably than DeepSeek V4 Flash, which they described as temperamental and prone to building wrong mental models. The hardware crowd quickly got into the weeds on quantization resilience and inference constraints, noting Hy3 lacks DeepSeek's efficient KV cache architecture, making it impractical for long-context local use on anything less than very expensive multi-GPU setups, though its coding benchmarks against much larger models are genuinely impressive for the price point.
The glass backbone: Why the Army's logistics will break in the next war [comments]
365 points · 445 comments · mwi.westpoint.edu · 18h ago
The article argues that the US Army's logistics system, built for uncontested supply lines and static bases in Iraq and Afghanistan, would collapse in a peer conflict where supply convoys and fuel dumps are constantly targeted by drones and precision fires. The thread largely agreed with the thesis, but the real energy came from a tangent about Ukraine's decentralized, almost gamified drone procurement system, where units earn "e-points" for kills and spend them on whatever drones they want from an internal marketplace. That sparked a debate about whether this fragmentation is actually a strength—enabling rapid battlefield innovation and resilience against supply chain attacks—or a logistical nightmare that only works because Ukraine lacks the industrial base to standardize anyway. A few folks pushed back that the US already learned this lesson in WWII with distributed production of Shermans and proximity fuzes, and that the real problem isn't decentralization but the military's cultural preference for funding fancy new weapons over boring things like fuel trucks and ammo depots.
Muse Spark 1.1 [comments]
365 points · 181 comments · ai.meta.com · 17h ago
Meta announced Muse Spark 1.1, a multimodal reasoning model for agentic tasks, alongside a new developer API and aggressive pricing of $1.25/$4.50 per million tokens with cheap cached reads at $0.15. The thread immediately turned into a pricing shootout: many argued Meta is undercutting Grok 4.5 and Claude Opus to claw back mindshare, but others countered that for a “no-name” model the price is still too high, and a few who tried it said quality lags behind Sonnet. A deeper split emerged over whether subsidized token prices signal a healthy commodity market or an unsustainable race to the bottom, with one camp citing DeepSeek V4’s sub-$0.004 cache hits as the only sane benchmark. The conversation then veered into a familiar HN debate about whether AI will increase or decrease demand for software engineers, drawing parallels to 3D printing and Excel—some insisted vibe-coding will create a new class of “fixers,” while others warned good engineers won’t touch the resulting garbage. A side thread argued that China’s software market has already peaked under AI pressure, though another commenter dismissed that as irrelevant to Western enterprise SaaS dynamics.
Why developers are ditching GitHub for Codeberg and self-hosting alternatives [comments]
350 points · 245 comments · www.howtogeek.com · 23h ago
The article argues that developers are ditching GitHub for Codeberg and self-hosted alternatives, citing Ghostty and Zig as high-profile defectors upset by outages, Microsoft’s politics, and creeping AI integration. HN immediately pushed back on the headline—many called it a “why” article that cherry-picks a handful of repos to pretend a mass exodus is happening, while the vast majority of users and projects remain firmly on GitHub. A vocal minority defended the move, pointing to real pain: one org’s CI got turned off for three weeks over a wrongful ban, another user gets rate-limited on every first click, and the ideological case against Microsoft’s ICE ties was argued hard, though that sparked accusations of “propaganda” in reply. The discussion also split on alternatives—Gitea and Forgejo got praise for copying GitHub’s UI well and being fast, while others argued it’s childish to refuse twenty minutes learning a new interface.
ChatGPT Work [comments]
335 points · 172 comments · openai.com · 15h ago
OpenAI announced they're rolling Codex's capabilities into ChatGPT proper, essentially rebranding the whole thing as "ChatGPT" with two modes—Work and Codex—but the thread is mostly people being utterly baffled by the execution. The consensus is that this is just a confusing renaming exercise: the old ChatGPT app was quietly renamed "ChatGPT Classic" and is being deprecated, while the Codex app now downloads as "ChatGPT," leaving users unsure where to go for casual conversation versus serious work. Several people point out that toggling between Work and Codex modes does almost nothing visually, and that the "chat" feature inside the new app is broken or buried—it creates folders in your Documents folder and injects system prompts meant for coding into every thread, which makes it unusable for simple chat. Comparisons to Anthropic's Cowork are common, but most agree OpenAI's rollout is a UX disaster, with multiple commenters calling it a mess and noting that the underlying capabilities were already available, so the change feels like pure brand confusion driven by Microsoft-style naming chaos.
Bonnie Tyler has died [comments]
333 points · 124 comments · www.bbc.com · 22h ago
The BBC obituary for Bonnie Tyler, who died at 75 after intestinal surgery, covers her career from Welsh council house to global fame with "Total Eclipse of the Heart" and "Holding Out for a Hero." HN mostly spent the thread reminiscing about the famously absurd music video for "Total Eclipse," filmed in a Victorian sanatorium, sharing the literal version and debating whether the "Take on Me" or "Safety Dance" literal videos are better. A few people dug into the production details—one commenter claimed to have mixed the song on a pair of NS-10s—while others argued the post is pure off-topic celebrity nostalgia that shouldn't be on the front page, sparking a familiar meta-battle over whether "good hackers" should care about a 75-year-old pop star's death.
A possible future for Damn Interesting [comments]
277 points · 37 comments · www.damninteresting.com · 16h ago
Alan Bellows, the founder of the long-running longform site Damn Interesting, posted a personal fundraiser asking readers to chip in so he can step back from full-time engineering work and actually write again. The HN crowd showed up with donations and a wave of nostalgia, many noting the site basically invented the "generally interesting shit" genre that later fueled podcasts like 99PI and RadioLab. A long thread spun out about platform dependency—Bellows recounted getting burned by Facebook’s algorithm change, and commenters argued over whether he should move to Substack for easier monetization, with Bellows pushing back hard on handing control to any third party. There was also a correction from a reader who admitted they’d assumed the site was blogspam, only to realize DI is often the original source that gets ripped off. Overall, the thread was supportive but sharply aware of the economics: the ask is modest, and people seem to want the old internet back, but there’s no consensus on how to pay for it.
No leap second will be introduced at the end of December 2026 [comments]
269 points · 205 comments · datacenter.iers.org · 17h ago
The IERS announced that there will be no leap second added at the end of December 2026, continuing the current pause. The thread quickly split into two camps: people relieved that the dreaded negative leap second—which would subtract a second from clocks—still hasn't materialized despite Earth's rotation speeding up, and others disappointed that we're missing the chance to witness a "negative leap second" in the wild. A deep technical discussion emerged around how systems like Google Spanner handle leap seconds via smearing, and whether the grid's 50/60 Hz frequency is affected, with a neat tangent about Kosovo-Serbia disputes causing European microwaves to drift six minutes off. The thread also devolved into joyful nerdery about the IERS's delightfully sci-fi titles ("Director of Earth Orientation Center") and whether the next step is a leap hour, though the consensus is that's millennia away and not worth worrying about.
US seeks cheaper hunter-killer drones after Iran destroys $1B worth of Reapers [comments]
256 points · 318 comments · arstechnica.com · 18h ago
The US military, having lost dozens of Reaper drones worth over $1 billion to Iranian air defenses, is now seeking cheaper, mass-producible hunter-killer drones to overwhelm enemy defenses with numbers, inspired by Ukraine's approach. The HN thread quickly pivoted from the article’s acquisition story to a deep, critical look at the Pentagon's procurement and software development culture. Domain insiders who worked on Reaper and Predator control systems described a broken, hyper-bureaucratic process where even a one-line code change could trigger six weeks of mandatory testing, actively discouraging bug fixes and innovation. The top discussion contrasted this with Ukraine's scrappy, existential-pressure-driven drone industry, noting that the US faces no such survival incentive, while several commenters debated whether the problems stem from cost-plus contracting, Congressional oversight rules, or simply a lack of political will to accept the kind of failures that enable rapid iteration.
Interview with Mitchell Hashimoto about Ghostty and Zig [comments]
217 points · 97 comments · alexalejandre.com · 14h ago
The interview covers Mitchell Hashimoto's move from HashiCorp to building the Ghostty terminal emulator in Zig, his philosophy on minimalist but feature-rich software, and his thoughts on terminal protocol evolution. Most of the HN discussion immediately zeroed in on his blunt admission that he doesn't like "Rust culture," which triggered a messy but familiar language tribe war in the comments — several people pushed back saying they've had the exact opposite experience with Zig's community being more aggressively evangelistic and anti-Rust, while others defended Hashimoto's point about taste and unapologetically weird communities being a good thing. A smaller but substantive thread debated whether Ghostty actually delivers on its promises, with some calling it buggier than iTerm2 with fewer features and others insisting it's been rock-solid for them and more performant. The conversation also split on Hashimoto's take that maintainers owe users nothing, with some nodding along and others arguing that attitude doesn't scale past pet projects.
A road to Lisp: Why Lisp [comments]
216 points · 162 comments · scotto.me · 18h ago
The article is a love letter to Lisp, arguing that its unique features—homoiconicity, macros that treat code as data, and a REPL-driven interactive workflow—make it a genuinely more powerful way to program, and that learning it changes how you think about problems. The thread immediately got derailed by a practical bug: the syntax highlighting on the site rendered code blocks black-on-black, making them invisible unless you selected the text; the author jumped in to thank the reporters and fixed it. The real debate, though, swirled around macros: a bunch of people couldn't get past the conceptual hurdle of how macros differ from higher-order functions, and a few commenters pushed back hard, arguing that Lisp's untyped s-expression AST actually makes macros *harder* to write correctly than Rust or Scala’s typed macro systems, and that for most real-world needs macros are unnecessary overkill. Someone also kicked off a playful “All roads lead to Lisp” / “What has Lisp ever done for us?” riff, with a list of concrete successes (ITA flight search, Emacs, Hacker News itself), which got a laugh and served as a counterweight to the navel-gazing.
AI content is everywhere on social media, especially LinkedIn [comments]
214 points · 188 comments · www.pangram.com · 16h ago
A Chrome extension called Pangram analyzed over a million social media posts and found that LinkedIn is the most AI-saturated platform, with more than 40% of longform posts flagged as fully AI-generated, while Reddit had the lowest rate at 4.4% thanks to overwhelmingly human-written replies. The discussion immediately zeroed in on LinkedIn's "enslopification," with multiple people saying they'd deleted the app entirely and now consider a LinkedIn absence a red flag when hiring. Others pushed back on the whole premise, arguing the real question isn't whether content is AI-generated — that's the new default — but whether it's human, given the economics heavily favor slop. A lively tangent emerged around people unconsciously mimicking LLM speech patterns, with some worrying that language evolution is now being driven by the quirks of a particular floating-point architecture, though others pointed out that phrases like "push back" and "circle back" were corporate clichés long before ChatGPT.
GLM 5.2 is nearly as accurate as a human book keeper [comments]
200 points · 115 comments · toot-books.pages.dev · 13h ago
The linked article benchmarks GLM 5.2, an open-weight model, on preparing UK quarterly VAT returns, finding it off by only 7 pence on net position after processing 59 transactions at a token cost of $2.73—compared to human bookkeeping fees of £750–2,100 per quarter. The HN thread immediately zeroed in on liability: even if the model is nearly as accurate, who goes to jail when it books founder shares as “Capital Account” instead of “Unpaid Shares”? Several people who’ve actually run LLMs on their own books chimed in with their own scripts and setups, reporting good results but noting that the real bottleneck isn’t model accuracy—it’s getting clean, structured bank feeds and receipts without two-factor auth headaches. A recurring pushback was that the humans in the benchmark knew they were being tested, so the comparison is against a super-careful bookkeeper, not the average one making routine mistakes. The consensus split between “this is clearly coming and will crush small accounting firms” and “the risk of a single unnoticed catastrophic error undoes any cost savings, because you’re absorbing the liability the human accountant used to carry.”
Why American ambulance rides are so expensive [comments]
183 points · 242 comments · davidoks.blog · 9h ago
The linked article is a deep dive into why US ambulance rides cost thousands of dollars — not simple greed, but a structural mismatch between how ambulances are paid for (per-ride fees) and how they actually operate (like an insurance option, with most costs going to maintaining 24/7 readiness). The HN thread quickly split into two camps: people who skimmed and yelled "private equity," and the much larger group who read the piece and pointed out it explicitly argues the opposite, that even the biggest private operators barely break even. A lot of commenters pushed back against the framing that this is an unsolvable puzzle, essentially saying "just fund it like a fire department through taxes or a universal premium, like every other rich country does, and stop pretending this is complicated." Several people shared personal horror stories — a $1,700 bill for a three-quarter-mile ride, a friend bleeding out who took an Uber instead — while others discussed how California has banned surprise ambulance billing but companies get around it by fraudulently classifying emergencies as non-emergency transports. The most revealing thread argued that the real problem isn't the price tag but the "mental overhead" of being thrown into a bureaucratic nightmare while your family is in crisis, something people in countries with rational systems never have to think about.
TLS certificates for internal services done right [comments]
146 points · 107 comments · tuxnet.dev · 17h ago
The article argues that split-horizon DNS plus an ACME client like acme.sh hitting Let’s Encrypt via the HTTP-01 challenge (with a VPN interface acting as a WAF) is “the way” to get real, auto-renewing TLS certs for internal services without dealing with self-signed certificates. The HN crowd mostly disagreed with the “done right” label, pushing back hard on split-horizon DNS itself — several people called it a perennial headache that creates cache-inconsistency problems and forces you to mirror records, and they pointed out that DNS-01 validation (only a TXT record needs to be public, no A record at all) is far simpler and avoids the whole WAF/listen-on-VPN-IP dance. A strong split argued that managing a small internal CA or using wildcard certs with DNS-01 is less brittle than the author’s HTTP-01 approach, though others countered that distributing a private CA to every client is its own nightmare, especially across phones, Docker containers, and language-specific trust stores. The thread also spent time on the upcoming `dns-persist-01` challenge type, which several people are watching as a potential cleanup that could make DNS-01 even easier than it already is.
Opinionated and easy Pi.dev configuration [comments]
128 points · 66 comments · lazypi.org · 16h ago
LazyPi is a one-command installer that dumps 60+ skills, 67 themes, MCP support, and a bunch of other bells and whistles onto Pi, an otherwise minimal agentic coding tool. The HN crowd is sharply split: many argue this defeats Pi’s entire purpose, since Pi’s design philosophy is "less is more" and loading up the model’s context with a ton of pre-packaged plugins actually hurts reasoning and wastes tokens—some point out that a minimal setup uses 5k tokens while LazyPi’s bloated one uses 20k for the same trivial task. Others counter that not everyone wants to hand‑craft their config, drawing a direct analogy to LazyVim versus vanilla Vim or Neovim, and note that Oh My Pi has been providing a similar "batteries included" experience for ages. A separate thread digs into supply‑chain security: installing random plugins is a risk, and the developer acknowledges the need for a blessed plugin system, while some prefer to just write their own subagents in a couple of hours with Pi itself rather than trust someone else’s code. The clear consensus among the "own your setup" side is that LazyPi is the antithesis of Pi’s minimalism, but the other side is happy to trade that for immediate productivity without the configuration tax.
Parental device use and the adolescent-caregiver attachment bond [comments]
117 points · 93 comments · www.frontiersin.org · 7h ago
A new study in *Frontiers in Psychology* claims a link between how much parents stare at their phones and insecure attachment in their adolescent kids, coining terms like "technoference" and "phubbing." The HN thread immediately tore into the study's methodology, with the top refrains being that it just runs two self-report scales against each other—kids who already feel neglected say their parents are on devices, and vice versa—so the correlation tells you nothing about causation. Several people agreed the finding is intuitive but argued weak studies like this do more harm than good by giving people a "scientifically proven" excuse to moralize about screen time. A smaller camp pushed back that phones are genuinely addictive and that dismissing the research because of imperfect methods lets parents off the hook for their own compulsive behavior, though even they conceded the study's design was garbage. The thread also spun off into tangents about whether paying *too much* attention to kids is also harmful, the hypocrisy of telling kids to get off YouTube while scrolling yourself, and whether phone addiction is remotely comparable to alcohol or drugs.
PostHog FOSS [comments]
91 points · 66 comments · github.com · 17h ago
PostHog published a read-only GitHub mirror of their analytics platform stripped of all proprietary code, offering a 100% FOSS version. The HN thread mostly pointed out this repo is nothing new—it's been around for years and the main repo was always open-core with an `ee/` folder under a proprietary license. Self-hosting got hammered: several people said the hobby deployment script is broken, changes for the cloud offering make local setups impossible, and PostHog abandoned Kubernetes support in 2023, leading to a broader debate about whether the software is truly usable as FOSS or just source-available. A surprising tangent emerged around the `AGENTS.md` rule "Avoid em-dashes like the plague"—users found the file itself full of em-dashes and called it a transparent attempt to hide AI-generated text, with one person arguing they'd rather flag AI writing openly so readers can skip it. There was also a split between longtime users who love the product's simplicity and those who feel it's become bloated with AI features and a confusing navigation overhaul, while a few pushed back saying the AI chat is actually well-implemented.
Launch HN: Context.dev (YC S26) – API to get structured data from any website [comments]
91 points · 64 comments · www.context.dev · 16h ago
The thread is about Context.dev, a YC startup offering an API that turns any URL into clean Markdown, brand metadata, or custom structured JSON for AI agents and product integrations. A lot of the pushback centered on whether it’s just another web scraping service with a nicer API — the founder leaned into that openly, saying "yes but we do a lot more" around brand data and caching, while several commenters immediately flagged it as something they’d need to block server-side, with one quipping about an "AI-service-blocklist-as-a-service" business model. There was a direct head-to-head with Firecrawl (the incumbent in the space), where the founder claimed better error rates and no hidden credit multipliers, but the response was basically "we’ll benchmark it" and the consensus seems to be that competition is healthy and teams blend providers. A more technical split emerged around reliability: some argued that new frontier models already do this natively (but slower and pricier), while others appreciated the nuts-and-bolts honesty about residential proxies, caching layers, and the tricky ethics of deciding where to draw lines on auth-walled pages.
Show HN: Analog Watch [comments]
90 points · 86 comments · analog.watch · 17h ago
The submission is a web-based tool for practicing reading analog clocks. The HN thread quickly turned into a debate over whether teaching analog clock reading is a waste of time — one camp argued it's as obsolete as a slide rule, while the other pushed back hard that analog faces give an intuitive sense of time as a progress bar, not just a number, and that the skill is still taught in schools for mathematical reasoning. A tangent about the phrase "half ten" meaning 9:30 in German versus 10:30 in British English drew a lot of confused corrections. The creator also got a feature request to display the clock hand positions in the results, which they said they'd add.
Show HN: FableCut – A browser video editor AI agents can drive (zero deps) [comments]
90 points · 58 comments · github.com · 18h ago
The linked article wasn't available to this summarizer; from the discussion, FableCut is a browser-based non-linear video editor that exposes its entire timeline as a single JSON file, letting AI agents (via MCP, REST, or direct file writes) edit the video while the UI hot-reloads in under 150ms. The thread loved the architectural choice of using JSON+SSE instead of a traditional API, calling it a simple and effective way to let agents and humans edit the same timeline simultaneously. Several people pushed back on the name "Fable," arguing it sounds tacky or is capitalizing on Anthropic's product, though others defended it as fitting for storytelling. The author got called out for using AI-assisted replies, apologized, and promised to follow guidelines, but also delivered a demo video showing an agent editing a real reel in one shot. Some commenters wanted client-side rendering via Mediabunny or a hosted demo, while others debated whether the project even needs AI for simple TikTok edits.
Focus [comments]
89 points · 49 comments · boz.com · 7h ago
Andrew Bosworth, Meta's CTO and an early Facebook engineer, published a piece arguing that startups thrive on ruthless focus, and that Meta lost its way by saying "yes" to too many small, individually reasonable things—like charitable giving—that collectively created drag. The Hacker News thread is overwhelmingly hostile: most responses treat the essay as a stunning display of self-awareness failure, pointing out that Bosworth himself helped drive the Metaverse distraction that burned tens of billions, then laid off 20,000 people, and that the company's "focus" was always on maximizing engagement and ad revenue regardless of societal harm. A few defenders argue the article is really about bloat and cognitive load, and that the charitable-giving example was just a proxy for a broader point, but the dominant take is that this is survivorship bias from someone who lucked into early Facebook and now blames employees for the leadership's own strategic chaos. Several commenters also note that the timeliness—the piece was originally from 2023—changes the reading, since Bosworth subsequently pivoted to leading Meta's AI efforts after the Metaverse flop, further undermining the "focus" argument.
How to Write an Email [comments]
89 points · 49 comments · blog.dannycastonguay.com · 16h ago
The article by Danny Castonguay gives blunt, no-nonsense rules for writing emails that prioritize speed and clarity—lead with the point, put bad news first, keep it to one screen, and make the next step obvious. The Hacker News crowd largely agreed with the core method, especially the "BLUF" (bottom line up front) strategy, but they split hard over tone: several people argued the recommended subject lines like "Action needed today: approve revised offer" sound like barking orders and risk coming across as rude or bossy, especially in cultures that expect more politeness. A long thread debated a clever technique—telling the recipient you'll proceed with a default action unless they object, even proposing a deliberately bad default to force a response from unresponsive colleagues, though some warned that tactic could make you look incompetent. Others pushed back on the "one email per topic" rule, pointing out that sometimes you need connected questions, and recommended numbered lists for multiple asks. A few commenters veered into workplace culture complaints, particularly about France's formal email style and the plague of meeting invites labeled only "Point" with no agenda.
SimPolitics: America’s quest to solve politics with computers [comments]
89 points · 24 comments · direct.mit.edu · 17h ago
The linked article wasn't available to this summarizer; from the discussion, it’s an open-access book about historical attempts to model and simulate politics with computers, from the Limits to Growth to rational-choice theory. The author showed up in the thread and fielded questions, which shifted the conversation toward the tension between computational ideals and messy human reality—many commenters flatly rejected the premise that politics can be “solved” by simulations, insisting that self-interest, short-term thinking, and corruption are the real drivers. A few threads spun off into sortition (random selection of leaders) as a more radical alternative, and one person argued that the real problem isn’t modeling but scaling human empathy. The book’s own framing, according to the author, traces how computer reasoning became the yardstick for evaluating political thought, but the crowd was sharply divided between those intrigued by the history and those convinced the whole project is naive.
Building a real-time AI tutor for 5-year-olds [comments]
83 points · 121 comments · www.ello.com · 11h ago
Ello published a detailed technical post about their real-time AI tutor for 4–9 year olds, walking through how they ditched the standard agent loop to avoid multi-second latency that makes kids tune out, using a streaming interpreter and an asynchronous planner to make pedagogical decisions during natural pauses. HN split hard: a vocal camp argued that five-year-olds need human interaction, not AI, calling it “brainrot” and comparing it to dumping kids in front of YouTube, while others pushed back that the current system is failing 60% of US fourth-graders and that a well-built AI tutor can supplement teachers and give every child something close to 1:1 tutoring. The founders waded into the comments, emphasizing they’re a public benefit corp with a free tier in emerging markets and that the goal is to free up teachers for human connection, but critics weren’t buying it — especially around hallucination risks (someone cited Gemini suggesting freezing your hands in an ice bath for a recipe) and the developmental importance of peer play at that age. A few commenters who said they were that kind of curious, underserved kid growing up were genuinely excited, framing it as the “private tutor for everyone” promise finally realized.
Generated 2026-07-10 08:29 UTC
Generated by Sauron from Hacker News discussions and linked articles.