HN Brief: 2026-09-22

MiMo v2.6 [comments]

845 points · 378 comments · mimo.xiaomi.com · 11h ago

Xiaomi’s MiMo v2.6 hit HN overnight, and the big takeaway was a rare bit of praise for a Chinese lab publishing model release charts that don’t fudge the numbers—people specifically liked seeing it benchmarked honestly behind DeepSeek, not just cherry-picked top scores. The comment section split into a few camps: some dove deep into the technical artifacts, like the live RL dashboard Xiaomi posted during training showing “staleness” and KL divergence metrics, which one person called the best teaching tool they’ve seen and promised to write a blog post about, while others zeroed in on the political angle, arguing the OpenAI/Anthropic “slow down” push is just regulatory capture to lock out cheaper Chinese open weights from US enterprises—though the counterpoint came up that American firms don’t touch these models mostly due to data-protection liability, not pure politics. A fun tangent nobody expected: people noticed the models keep generating that “01 · UPPERCASE LABEL” frontend motif, wondering if it’s emergent RLHF behavior or just scaffolding, and one user tried to prove a conspiracy theory that the older MiMo got “dumber” before the release, only to get a shrug in return. The biggest genuine debate was about whether open-weight Chinese models like this will actually disrupt US Big AI, with one side saying the margin squeeze is already happening and the other pointing out that neutral hosting and legal risks make it a non-starter for most American companies regardless of how good the model is.

Attention is all you have [comments]

755 points · 223 comments · alicegg.tech · 17h ago

The article, "Attention is all you have," argues that recommendation algorithms and AI-generated content are hijacking our attention, contrasting today's curated feeds with a more intentional, bookmark-based internet. HN immediately seized on the title's nod to the "Attention is all you need" ML paper, sparking debate over whether the author’s omission was a deliberate snub or an innocent miss, with some claiming the connection is obvious regardless. The discussion then broadened into a familiar split: some defended algorithmic curation as just another navigational tool, while others romanticized the pre-algorithm web, though a few pushed back that Yahoo! and MSN were already clickbait machines. A major tangent veered into the impossibility of human verification online—with one commenter predicting a doom-loop where bots force humans toward cryptographic proofs—before detouring into a surprisingly heated debate about whether Starbucks' failure in Australia is proof of quality coffee culture or just marketing. Ultimately, the thread lands on a sobering consensus: the problem isn't just algorithms, but that most people don't care enough about what they consume, making human curation a niche luxury rather than a viable default.

I don't want to read what you didn't write [comments]

608 points · 220 comments · blog.colinbreck.com · 9h ago

The linked article wasn't available to this summarizer; from the discussion, Colin Breck’s essay argues that AI-generated writing, especially design docs and PR summaries, forces readers to wade through context-free detail, and that machine-generated prose eliminates the vulnerability and voice that make human communication meaningful. The HN crowd largely split between those who see value in using LLMs for personal synthesis and those who refuse to receive AI text from colleagues entirely. A recurring pushback centered on the idea that the reader often lacks the LLM’s context, so sending raw code or a personal note beats a polished but bloated summary. Others pointed to real-world failures, like journalists misrepresenting academic findings through sloppy summarization, and there was a fascinating digression on Shannon information theory versus semantic meaning, with some arguing LLMs only add redundant bits that actually obscure transfer. The strongest agreement came around the notion that static AI output is a poor substitute for interactive explanation—if you’re going to make someone read machine prose, at least give them the source so they can query it themselves.

What Sun got wrong [comments]

574 points · 327 comments · bcantrill.dtrace.org · 18h ago

The linked article is Bryan Cantrill's retrospective on Sun Microsystems, centering on the idea that Sun became "bored with the mechanics of running a business," using a 2005 anecdote where a startup (which later became Oxide) couldn't get Sun to return a phone call while Dell swooped in overnight. The HN discussion largely pivoted from Cantrill's operational-failure thesis into a broader, heated debate about enterprise hardware sales culture, with multiple people sharing war stories of negotiating with Dell, bemoaning the "value-based pricing" games and the ritual of pushing back on quotes. A significant split emerged: some argued Dell has itself become the new Sun, complaining about inconsistent pricing and long lead times (with one user jokingly suggesting AI will replace, not remove, sales middlemen), while others defended the big OEMs for their engineering, logistics, and long-term part availability, contrasting them with cheaper whitebox alternatives. A separate, deep technical tangent challenged the article's implied nostalgia for Sun's SPARC architecture, arguing x86 was far more proprietary and that Sun's real failure was a lack of software ecosystem (like not getting Mozilla to build on Solaris), with some contending that DTrace and ZFS, while great, were "just Blub" without the ability to run the software people actually needed.

Grok 4.7 [comments]

555 points · 476 comments · x.ai · 16h ago

Grok 4.7 launched as xAI’s latest coding and knowledge-work model, touting a larger base model, longer reinforcement learning, and a new safeguard stack, all at the same price as 4.6 with a faster, pricier variant available. The HN crowd immediately split into two camps: one dismissing Grok as lazy and bland—citing it ending tasks early or refusing to critique Elon Musk’s circle, including a tangent about his mother scrubbing an AI-generated photo label—while others swore by it for coding, claiming it outperformed Trae and rivaled Claude in real-world frontend work after launches. A recurring theme was frustration with rival models’ prose (“Claudish”), with several arguing Grok’s plain-English output was a feature, not a bug, even as one user predicted the model’s RL training would regress that clarity. Benchmarks took a beating too, with skepticism about whether long-autonomy scores translate to real utility, and some noted token price alone ignores efficiency—though others speculated xAI’s delayed release and margin squeeze signal cold feet ahead of a rumored Opus 5.5. The thread drifted into broader rants about oversimplified explanations, neurodivergent communication styles, and whether AI’s tendency to overcomplicate is a sign of fraud, but the loudest takeaway was that Grok 4.7’s value hinges less on specs and more on how it handles messy, multi-step tasks without hand-waving.

ZuckOff is a free app that sees Meta glasses before they see you [comments]

388 points · 340 comments · www.wired.me · 21h ago

The linked article covers ZuckOff, a free app by Polish developer Pawel Szydlowski that detects Meta smart glasses via their Bluetooth fingerprints, addressing privacy concerns over unsolicited filming. HN ran with the name first, joking about Meta’s legal team while debating whether Zuckerberg would even notice, then pivoted to technical nitpicking—spoofing concerns, signal-strength beeps, and a proposed "do-not-film-me" Bluetooth standard that sparked a heated split between tech-fix idealists and those insisting it’s a social problem needing laws, not workarounds. The thread also veered into Disney copyright abuse by police, GDPR’s Article 22 for automated decisions, and a lot of back-and-forth on whether social norms or enforcement actually change behavior. The sharpest pushback argued that any tech solution inverts the onus onto the filmed, not the filmer, with one blunt take that Meta’s camera is designed to be unnoticed and no app fixes that.

Fable 5 – Median thinking declined in August [comments]

385 points · 272 comments · x.com · 15h ago

The article claims Anthropic's Fable 5 model saw a dramatic drop in "thinking tokens" and overall capability in August after being made widely available, sparking a thread split between genuine suspicion of deliberate downgrades and skepticism of anecdotal complaints. A lot of the pushback centered on the impossibility of verifying such claims—users can't measure thinking tokens directly, and the company insists it never intentionally degrades models, though commenters point out that leaves huge loopholes around quantization or prompt tweaks. Others argued the perceived decline is just hedonic adaptation or the model being routed to cheaper hardware, with several people sharing similar week-to-week drops in other models like GPT-5.6-luna, building a case that this is either a widespread industry pattern of "release hot, serve cold" or confirmation bias amplified by closed black boxes. A recurring sub-theme was the "slot machine" comparison—both for the UX encouraging repeated attempts and for the lack of seed control making any single output impossible to audit, leaving users to trust or abandon providers on vibes.

Spymarks, Not Watermarks [comments]

383 points · 91 comments · brand.io · 9h ago

The article argues that "watermarks" have quietly morphed into "spymarks"—hidden, identity-mapping signals embedded by companies like Google and OpenAI into AI-generated content, which can trace back to your personal data without your knowledge. HN pushed back hard on the proposed terminology, with several people suggesting "tracking watermark" or "steganomark" as clearer alternatives, while others debated whether the fear is overblown since invisible watermarks can have legitimate provenance uses. The discussion split between those who see SynthID et al. as a net positive for identifying AI slop—one person even dreams of a browser extension that highlights LLM-written text—and those who frame it as a privacy nightmare, with Apple’s on-device processing getting dragged in as a counterexample and then dismissed as still an ad company at heart. A particularly sharp tangent emerged around how anyone can actually trust a provenance claim if the key is public, and the commenters also wandered into steganography, printer tracking dots, and even whether Apple’s 15-second audio buffer means your conversations are already being summarized. A developer even showed up with an open-source alternative to SynthID, arguing transparent, independently-decodable watermarks are the only way to keep provenance in the hands of the people, not the corporations.

NASA’s Mars Sample Return mission is dead [comments]

377 points · 310 comments · www.science.org · 12h ago

The article reports that NASA’s Mars Sample Return mission has been cancelled, a program that had grown wildly over budget and faced technical hurdles. The HN discussion quickly pivoted to the geopolitical angle, with multiple people pointing out that China’s Tianwen-3 mission, planned for 2028, will likely return samples first—some framing this as a symptom of American decline, while others pushed back, arguing the US program was poorly designed from the start (relying on the Perseverance rover’s scattered tube drops) and that China’s simpler approach isn’t directly comparable. A significant strand of the comments veered into a broader history argument about the space race, with one person detailing how the USSR dominated early space firsts and another invoking NDT’s old speech that war, not vision, drove Apollo—while others countered that NASA’s spinoff tech list is overstated. A few side discussions spun off into cynical tangents about Chinese censorship, fleeing scientists, and even whether people should just stop reading the news entirely. The overall split was between those mourning the loss of US leadership and those saying the mission was a mess that deserved to die, with China’s trajectory looming over both camps.

Transformers Explained Visually [comments]

365 points · 59 comments · poloclub.github.io · 12h ago

The linked article is a deep-dive visual explainer of how Transformer models work, using GPT-2 as a live, in-browser example to walk through tokenization, attention, and MLP layers. HN ran with it as a springboard for both praise and pushback: some folks flagged that absolute positional encoding is outdated compared to modern RoPE, while others debated whether the "temperature as creativity" framing is misleading—one person noting that "safety" is the wrong word since low temperature makes text feel artificially stiff. A recurring tangent spun off into literal power transformers (the electrical kind), with people sharing DIY winding stories and a link to a homemade high-voltage turbine setup. There was also a genuinely useful analogy: an attention head’s multiplication of attention weights by values effectively acts like a dynamically-constructed dense layer, which one commenter said was rarely spelled out clearly. The thread split between "this is a great intro" and "this still assumes too much about MLPs," with a couple of folks linking older but similar visualizers and a Schmidhuber citation debate popping up for good measure.

Turn off and restrict access to Apple Intelligence features on Mac [comments]

293 points · 193 comments · support.apple.com · 14h ago

The linked article is Apple's official support guide for turning off and restricting Apple Intelligence features on macOS, detailing steps for disabling things like Siri AI and summarization tools. HN ran with the frustration angle, with several people complaining that even after disabling these features, the local AI models don't free up their disk space—one user cited reclaiming 13GB on iOS and still hitting "not enough space" errors during updates, while another noted Apple doesn't offer a full delete button, drawing a pointed comparison to the infamous free U2 album incident. The thread also devolved into broader gripes about Apple's "crapware maturity," with one camp arguing Apple is now echoing Microsoft's bloat tactics, though others pushed back, noting that macOS actually lets you skip iCloud and run without an account, unlike Windows' aggressive account requirements. Another tangent touched on how Apple's local "intelligence" is still dumb—one person shared a story of a Genmoji suggestion popping up mid-text about pizza, which some saw as proof the whole feature is both invasive and useless. A side conversation on Little Snitch and firewalls did surface a claim that Apple's OS can bypass user-space network filtering, but the thread didn't settle that one.

M5 Ultra Mac Studio Review [comments]

245 points · 246 comments · www.macstories.net · 18h ago

The M5 Ultra Mac Studio review on MacStories dives into how the new hardware finally makes local AI assistants—like Qwen3.8-Flash-Next—fast enough to replace cloud models for everyday agentic work, and the author spent four days testing it on a massive WWDC-notes project. HN ran with two big angles: first, the price-vs-performance argument got heated, with one camp pointing out that a 5090 PC costs half as much and delivers comparable or better raw throughput, while others shot back that the Mac’s unified memory lets you run much larger models than 32GB VRAM can touch, and the dense-vs-MoE architecture debate spilled into a side argument about whether Apple’s memory bandwidth advantage actually matters for sparse expert models. The other split was cultural—folks dragged in the usual Apple-versus-PC tribal war (including a weird tangent about Metal and gaming), but the sharpest pushback came from a technical crowd noting that the M5’s 1.2 TB/s bandwidth still caps token generation around 81 T/s without speculative decoding, and that a dense Qwen model isn’t the right comparison when MoE models are where the memory actually shines. One standout jab summed up the skepticism: a commenter who tried local models quipped that the result is "able to be WRONG, faster," which sparked a brief but sharp exchange about ADHD-supervision metaphors gone off the rails. The consensus seems to be that the M5 Ultra is genuinely impressive for its class, but whether it’s worth the price tag depends entirely on whether you’re optimizing for absolute speed or the freedom to run big local models without renting cloud GPUs.

Raspberry Pi blocks changing RAM chips [comments]

240 points · 188 comments · forums.raspberrypi.com · 19h ago

Here's the summary of the HN thread on Raspberry Pi blocking RAM chip swaps:

The story is that Raspberry Pi has started locking down their newer Compute Modules (specifically the CM5) so that users can't swap out the RAM chips, with the EEPROM checking for only the RAM it shipped with. The HN discussion quickly split into two camps: those arguing this is an anti-consumer move that violates the spirit of the Pi's hobbyist ethos, and those defending it as a necessary business decision because shady resellers were buying low-RAM units, hot-air reworking cheap or QA-reject RAM chips onto them, and reselling them as higher-RAM SKUs—which then bounces back as support and RMA nightmares for the foundation. Some of the more interesting tangents included whether this violates the EU's Right to Repair directive (with folks pointing out that part locking against unoriginal parts is prohibited there, though the chips are technically still functional), the legal argument that price discrimination and hardware locking aren't illegal, and someone suggesting the obvious alternative of just drilling a hole or etching the board to mark the RAM size—so the foundation could let users go wild and still easily identify fraud. There was also a side argument about whether Raspberry Pi is even still cheap enough to justify the "affordable hobbyist computer" framing, with one user

comparing an 8GB Pi 5 setup to a MacBook Neo on price and getting pushed back on the math, but the core tension remained: is this protecting the ecosystem from bad actors, or is it just another step toward locked-down, soldered-everything computing?

Jev-Leftpad [comments]

229 points · 87 comments · github.com · 23h ago

The linked article isn't available to this summarizer, but from the discussion, it's clear this is a satirical npm package that uses the Jev AI model to decide how many spaces to left-pad a string—a task `padStart()` does in one line. The HN crowd ran with the absurdity, riffing on the original 2016 left-pad incident that broke the npm ecosystem, with several commenters joking that this will inevitably worm its way into critical dependencies and cause outages, or suggesting even more convoluted approaches like recursive Jev calls or spawning multiple models for precision. There's a lively sidebar debate about whether Jev's confidence scores mean anything, with some pushing back on the hype, and a few folks genuinely puzzled about where such a model fits, though most just appreciate the deadpan humor and the callback to the Fizz Buzz in TensorFlow era. The thread splits between those delighted by the absurdity and a handful of skeptics who are already fatigued by the pace of AI-vibe-coded junk, but the consensus is that this is a masterclass in trolling the tech ecosystem.

macOS 27: Workaround to avoid downloading AI models and save storage [comments]

224 points · 112 comments · www.reddit.com · 17h ago

The linked article is a Reddit guide for macOS 27 users on how to free up storage by preventing Apple Intelligence’s on-device AI models from downloading, primarily by using Low Data Mode or creating a Siri language mismatch before deleting the model files from recovery mode. The HN discussion quickly moved past the technical fix and into a broader complaint about Apple’s forced AI rollout, with people arguing that the company is wrong to treat these multi-gigabyte models as mandatory when many users have 256GB drives and no interest in the feature; one side insists it’s about respecting user choice and disk space, while another points out that skipping the update isn’t viable long-term since even macOS 27 has stability bugs, and a few note that you can disable Siri entirely via Screen Time restrictions for an easier workaround, though there’s no consensus on whether Apple will ever make this opt-in given their pivot toward AI-driven stock performance. The comments also touch on tangential gripes like CarPlay forcing Siri on, the size of third-party model files, and a brief spat over whether Apple’s reported model sizes are accurate, with one user citing a 15.5GB figure on their own machine to counter a news article. Overall, the sentiment is split between “this is exactly why I’m holding off on upgrading” and “it’s not that big a deal, just manage your storage,” with a heavy dose of frustration that Apple’s AI push is eating into both disk space and user agency without offering a simple on/off switch.

US halts flights at busy East Coast airports, says fiber line cut [comments]

220 points · 129 comments · www.reuters.com · 13h ago

The FAA halted incoming flights across major East Coast airports for most of the day after a construction crew cut a primary fiber line in New Jersey, and the backup line also failed, which snarled travel for thousands and collided disastrously with the UN General Assembly arrivals. The thread quickly moved past the immediate news to hammer on the systemic failure: the fact that a life-critical system had no active monitoring to report the backup fiber was dead until the primary failed drew sharp, profanity-tinged ridicule from people who've worked in networking, with one recurring analogy being that nobody tests their backups until they need to restore from them. A separate, more optimistic vein emerged around a newly deployed ATC system, but that was met with deep skepticism—one person joked it was probably "vibe coded" given the contract-to-deployment timeline, though someone else pointed out it had to meet DO-278A standards, so it wasn't entirely reckless. The deeper argument split between those saying the real problem is a pathological lack of redundancy (with stories about how dual-path links still share conduits or get dug up by backhoes) and those blaming the FAA's bureaucratic blizzard for making competent people useless, with a notable aside wondering if a properly secured tertiary link over the general internet would honestly be worse than having nothing at all.

AI coding has made CI a bottleneck, so we reworked ours to keep up [comments]

219 points · 226 comments · linear.app · 12h ago

Here's a summary of the overnight HN thread:

The piece from Linear dives into how they overhauled their CI pipeline to handle the load from AI-generated code, cutting PR wait times from around 6 minutes to just over 5 despite test suites quadrupling this year—achieved through infrastructure upgrades, splitting slow test files, caching dependencies in base images, and batching short checks. The top comment thread sparked a debate on whether Bazel is worth the complexity for TypeScript projects, with one practitioner who led a multi-year Bazel conversion arguing the tool's setup costs have dropped recently but conceding the JS ecosystem doesn't see the same wins as compiled languages—while others pointed out Bazel's cache hit rates for typical PRs are solid but network I/O and fetching artifacts often become the real bottleneck, with some saying the classic "just run it locally" approach remains more practical for most teams. Elsewhere, a side argument broke out when someone admitted not knowing "CI" stood for Continuous Integration, leading to a heated exchange about whether tech writing should spell out acronyms on first use—with most people landing on "just use the term, but spell it out so new grads can google it"—and another deflection suggested tools like Linear are getting in the way of fully agentic dev workflows, though replies noted you still need a structured place to store history for AI agents to reference. The consensus is that Linear's optimizations are solid, incremental wins but hardly revolutionary, and the real takeaway is that AI agents are now writing half the tests, so the bottleneck is shifting from human patience to compute cost.

Python Workers are now generally available [comments]

219 points · 36 comments · blog.cloudflare.com · 18h ago

Cloudflare announced that Python Workers are now generally available, letting developers run Python natively with full access to the platform's bindings, popular frameworks like FastAPI and Django, and even database drivers via a socket bridge, marking a big step toward making Python a first-class citizen on the edge. The HN crowd mostly engaged with the technical trade-offs, particularly cold start performance—while one author admitted there’s still room for improvement, a Wasmer representative countered with benchmark claims that their own Python cold starts are 16x faster (60ms vs. ~1s), sparking a debate over whose numbers tell the real story. A maintainer of urllib3 added important nuance, praising Cloudflare’s upstream contributions but warning that such funding often supports one-off features rather than the long-term maintenance burden, using a CVE as an example of where browser-specific `fetch` semantics break expectations. The thread also veered into how Pyodide’s architecture ties you to a single Python version per workerd runtime, with a Cloudflare author clarifying that compat flags exist but note older versions lack JSPI support. A couple of tangents stood out: one user quipped that this is just a reinvention of PHP and Perl’s model, and another pointed out that Go still can’t run on Workers due to WASM limitations, with a suggestion that Elixir might be a better fit for clustering—though the funniest moment was the saga of the correct benchmark link, which kept getting misdirected before landing on a cold-start comparison, showing how eager folks are to compare real-world latency numbers.

Uber arbitration award over Emily Normandin-Parker’s death [comments]

156 points · 181 comments · consumerrights.wiki · 17h ago

The linked article covers a retired judge's arbitration award ordering Uber to pay $40 million to the parents of Emily Normandin-Parker, who died after a driver left her and another passenger at an unsafe freeway gore point while intoxicated. The arbitrator found Uber vicariously liable despite Proposition 22, ruling the law didn't immunize the company for the driver's actions as a common carrier, while also criticizing the driver's credibility and noting he prioritized his car over his passengers. The HN discussion quickly pivoted from the case's facts to a broad debate on forced arbitration, with one side arguing Uber's attempt to attach confidentiality and non-disparagement clauses to the settlement (with a $10 million penalty per breach) proves they're trying to suppress the precedent and negotiate around the award, while others defended arbitration as more efficient and cheaper than courts, noting the plaintiff could have rejected the ToS. A sharp split emerged on whether arbitration is inherently unfair—proponents of the current system argued it lets parties avoid costly discovery and class-action abuses, while critics countered that it's a rigged parallel legal system where the company often picks the arbiter and there's no meaningful recourse, with some suggesting the real fix is making arbitration optional but preserving the right to a public trial for certain claims. The thread also touched on whether consumers truly understand the contracts they sign, with one commenter arguing the "just don't use Uber" stance ignores the power imbalance in take-it-or-leave-it agreements, though others noted that arbitration's confidentiality means individual wins don't help other victims, reinforcing public distrust of the process.

Suspension of the de minimis administrative exemption for imports $800 or less [comments]

154 points · 124 comments · www.personalimportation.org · 11h ago

The linked article wasn't available to this summarizer; from the discussion, it’s about the Trump administration suspending the de minimis import exemption for goods under $800, which will take effect just before the midterms and could heavily disrupt cross-border purchasing, including Canadian pharmaceuticals. HN tore into the policy’s politics, with one camp arguing it’s a blatant tax hike on consumers that even right-leaning voters are souring on, while others pushed back that tariffs are legally delegated to the executive and that voters actually supported them at the ballot box. The thread also got into the weeds on implementation: a Canadian seller explained how shippers like USPS eat the duties up front, making it a logistical mess, and someone warned this could kill affordable drug imports for Americans who rely on them, though a few commenters noted it’s not a ban, just a fee. The overall mood was split between schadenfreude over the political timing and genuine alarm about the practical fallout for border enforcement and personal imports.

Amazon blocks Meta’s new Muse AI agent from shopping on amazon.com [comments]

146 points · 152 comments · www.forbes.com · 15h ago

The linked article covers Amazon blocking Meta’s new Muse AI shopping agent from its platform, citing violations of its Conditions of Use after a federal appeals court overturned an earlier injunction against a similar agent. HN ran with the angle that this is purely about control and ad revenue, not security, with several people noting Amazon is happy to direct shoppers to its own Alexa for Shopping agent and that blocking third-party agents kills the ads and promoted listings that actually drive Amazon’s profits—one commenter pointed out that if an agent bypasses the UI, Amazon loses its entire advertising business. The discussion split between those arguing that agents let users finally cut through dark patterns and "lightning deals," versus skeptics who think Amazon will just escalate to blocking residential IPs or forcing attestation, with one person noting that local agent setups on personal connections are nearly impossible to block cleanly. There was also a tangent mocking the "Rufus" rename as a desperate KPI grab, and a pointed observation that Amazon’s real moat is the account and the data tax it extracts, so the EU should step in and mandate agent access rather than let every platform wall off its own garden.

Frontier AI on Your Own Hardware [comments]

146 points · 76 comments · timdettmers.com · 13h ago

The linked article—a CMU professor’s manifesto about how open-source AI agents will spark an academic renaissance—quickly ran into two walls of skepticism. One crowd pounced on the prose itself, calling it riddled with AI-slop tells (the "wee" typo, a nonsensical "je ne sais pas," and a meandering structure that reads like marketing copy), with one person flat-out accusing it of being LLM-generated. The other, more substantive pushback came from industry folks disputing the core claim: while one commenter said the experienced dev market in London is "hot," another countered that recent CS grads already have the #2 highest unemployment rate, arguing the “everyone’s afraid of losing their job” framing is a real trend, not a fear to dismiss. The deepest argument wasn’t about whether agents can code, but whether they’ll eliminate the junior rung on the career ladder, leaving a future with zero entry-level jobs for anyone who isn’t already a senior—a role that can’t be filled by a “junior with a PM and a prompt.”

Ask HN: Is it impossible to disable Siri on macOS 27?

145 points · 75 comments · news.ycombinator.com · 19h ago

The linked article wasn't available to this summarizer; from the discussion, the post is about a user who disabled Siri and Apple Intelligence through every available setting on macOS 27, yet still finds the "Siri AI.app" process running—leading to frustration and comparisons to Microsoft's Recall. The thread quickly splits between those defending Apple's approach and those who see it as proof of deep integration that can't be fully shut off. A key technical point emerges: the "Siri" process may actually be Spotlight rebranded, since killing it makes the Spotlight UI flash, and several users note that even with Siri toggled off, the process persists under a different name depending on UI state. Some suggest workarounds like enterprise configuration profiles or mismatched language settings to disable Apple Intelligence entirely, while others joke about "Siri Classic" mode or admit they've given up and now run alternatives like GrapheneOS to escape Big Tech's tentacles entirely. A recurring side argument debates whether Apple's hardware quality justifies the lack of user control, with one person quipping that you wouldn't buy a car with a welded-shut hood just because it drives well.

Good people refuse to do bad things [comments]

138 points · 144 comments · carette.xyz · 16h ago

The linked piece covers Jacob Coxon, an Anthropic engineer who quit over AI extinction risks, and uses his story to argue that "good people refuse to do bad things" while criticizing tech workers for rationalizing complicity. On HN, the discussion split sharply: some defended the article's call for personal responsibility, sharing anecdotes of refusing unethical requests and arguing that financial security or coordination via unions are necessary preconditions for walking away, while others pushed back hard on the moral framing—noting that "good people" are easily misled by self-image and that the banality of evil means most wrongdoing comes from ordinary people, not villains. A significant faction dismissed the resignation as performative or misguided, with one skeptic arguing Coxon's stated concerns lack technical coherence and comparing him to a grifter, while another countered that the real problem is structural: a system that aligns individuals toward shareholder value, making personal virtue nearly impossible without collective action. The article’s final provocation—that staying for the salary is a cowardly transaction—drew both agreement and a pointed rebuttal that moralizing ignores the real-world constraints of mortgages, families, and the fact that people often only recognize they're in a "torment nexus" after the fact.

Don't Use AI to Write [comments]

138 points · 82 comments · paulbakker.io · 23h ago

The linked article argues that while AI tools are great for coding and research, using them for writing is a mistake because it skips the crucial thinking process that writing forces—though the author does approve of using AI post-draft for feedback and proofreading. HN ran with this and immediately pounced on the apparent contradiction, with one top comment mocking the logic: "you can use AI to structure your argument...that sounds a lot like using AI to write," while others defended it by sharing workflows where they write the core idea themselves and only use LLMs to untangle complex thoughts or translate into German, likening it to asking a passerby for advice. The thread also veered into a historical tangent comparing AI to the advent of the word processor and typewriters, where old-timers pushed back on the romanticized past. The sharpest split came from purists who argued that any AI seeding, even if rewritten, produces "generic derivative slop" and dulls your voice, while pragmatists countered that editing a model's output is still part of the thinking process—though the most pointed exchange in the thread had one side insisting using AI to even structure arguments is outsourcing the hardest part of the work.

Meta bans ads for Virginia Woolf play in Spain [comments]

136 points · 148 comments · www.theguardian.com · 17h ago

Meta’s automated ad system blocked a Barcelona theatre from promoting its stage adaptation of Virginia Woolf’s *A Room of One’s Own* on Instagram and Facebook, citing “social topics” like feminism and civil rights as prohibited content. The theatre pushed back publicly, noting the irony of censoring a 1928 feminist classic, and Meta suggested removing “key words” to avoid triggering its filters. The thread split sharply: some blamed Meta’s overbroad enforcement, while others pointed to the EU’s Transparency and Targeting of Political Advertising (TTPA) regulation, arguing Meta is maliciously over-complying to avoid fines—though several commenters rebutted that the law’s definition of political advertising is narrow and clearly doesn’t cover a play, making Meta’s interpretation a deliberate stretch. A side debate flared over whether the EU’s regulatory environment or Meta’s profit motives are the real culprit, with some defending European tech talent and others insisting US salaries and network effects are why alternatives never gained traction.

Apple Copland D11E4 Booting in the Browser [comments]

131 points · 37 comments · www.pagetable.com · 13h ago

The linked article showcases Apple’s abandoned Copland OS running in-browser via an emulated DingusPPC, giving users a hands-on taste of the ill-fated operating system. HN dove straight into nostalgia, with many reminiscing about classic Mac UI and swapping emulation resources like Infinite Mac, while others geeked out over the technical feat—though some quickly pointed out that the OS is just as crashy and unstable as it was back then, with one user noting it was rightfully axed. The discussion took a sharp turn into Apple lore, covering the Star Trek project (the Intel port) and its surprising legacy—parts of its code allegedly survived in QuickTime for Windows and Carbon—sparking debate over whether the cancellation was the right call, especially since some argued Ellen Hancock’s push to skip NeXT could have changed everything. A few commenters also geeked out over specific quirks, like the scrollbar not re-rendering text until release, and the general consensus that while Copland was a noble failure, its UI and responsiveness still feel memorable today. Meanwhile, the usual suspects couldn’t resist joking about the crash-heavy nature of the demo, and one person even got it to hard-lock on a QuickDraw 3D app, proving some things never change.

The Advisory Group on Mathematics and Artificial Intelligence [comments]

126 points · 60 comments · terrytao.wordpress.com · 12h ago

The linked article is a guest post on Terry Tao’s blog announcing the formation of an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the IAS, to advise AI companies—starting with a concrete ask from OpenAI about releasing over 100 purported machine-generated proofs. The thread splits sharply: many defend the math community’s calm, deliberate response as a model of adaptation, while others call it “forced optimism” or even sour grapes, arguing mathematicians missed the writing on the wall and are now damage-controlling by cozying up to labs. A recurring pushback is that the group’s independence is cosmetic—OpenAI approached them, they’re advising without power, and the whole arrangement reads to some as PR laundering for a company touting “solved” math problems without stating what they are. There’s also a tech-vs-math culture war simmering: software engineers are portrayed as having adapted faster (albeit with more machine-god hype), and at least one self-identified Nobel laureate suggests the real story is that OpenAI is stalling because it doesn’t actually have the claimed results. The loudest skepticism isn’t about whether AI can do math, but about whether this advisory group changes the underlying power imbalance—or just gives a veneer of human oversight to announcements that will keep coming anyway.

Why does mathmain need an encrypted loader? [comments]

126 points · 36 comments · safedep.io · 13h ago

The post details a supply-chain attack where an npm package called mathmain, a copy of mathjs, hides an encrypted remote-access implant that activates only when a specific matrix input produces the right password, with a decrypted payload communicating over a chat service and blockchain. HN dug into the mechanics, with some commenters speculating the trigger targets specific numerical-analysis workloads in vulnerable applications, while others flagged that the write-up itself reads oddly—likely LLM-generated—and that the real analysis came from JFrog, not the original post. A side debate flared up about dependency trust and CommonJS vs. ESM, with one commenter jokingly vowing to hand-write everything and another pointing out dynamic imports are still greppable, though not as easily as static ones. A tinfoil-hat theory even suggests the malware and the blog post are both AI-authored clickbait, given the strange emphasis on negations and the still-live npm package with no warning.

HERMES radio enables voice and data communication over vast distances [comments]

122 points · 56 comments · spectrum.ieee.org · 15h ago

Here’s what the HN thread on HERMES was about: the article covers Rhizomatica’s open-source digital shortwave radio system, HERMES, which uses the ionosphere to send files, GPS coordinates, and voice at a few kilobits per second for remote and emergency comms. HN ran with the regulatory can of worms first, with several people pointing out that encryption is illegal on US amateur bands (unless you publish the keys) and that "broadcasting" one-way is prohibited—though a few noted enforcement is complaint-driven and rare in practice, with the FCC focusing on persistent pirates. Some pushed back on the tech itself, noting 3 kbit/s is fine for email but nowhere near internet replacement, and one went off comparing it to LoRa or Winlink, while another suggested pairing it with local voice-agent TTS/STT for a phone interface. A side tangent dug into whether hams actually "broadcast" during CQ calls (they don’t, since it’s two-way), and someone linked a Data Slayer YouTube video that apparently covers the same hardware. There’s no real consensus—it splits between impressed by the resilience in the Global South and skeptical of the speed and legal friction in the US.