HN Brief: 2026-09-19
Today HN was dominated by a grim throughline on AI’s societal costs—the unredacted filings showing a Microsoft exec called scraping “the largest theft of labor in human history,” plus a near-war over a hallucinated intelligence report, and a thread arguing AI is just an elite crime spree. Meanwhile, a strong usability backlash hit passkeys (the thread was unusually unified in calling them a dangerous corporate solution), and Android 17’s move to gate APIs behind Pixel exclusivity sparked fears that Google is quietly killing AOSP. A lighter but still math-heavy engineering post from Cloudflare (saving 100TB of RAM with calculus) offered a rare moment of nostalgia for human-written technical deep dives.
Threads most worth clicking into: “Microsoft exec called AI scraping 'the largest theft of labor in human history'”—the thread vindicates the theft framing but splits badly on whether comparing AI training to human learning makes the label coherent; “Android 17 is the first since 3.x to add new APIs without releasing to the AOSP”—GrapheneOS is stuck reverse-engineering security patches while Google gives Pixels months of exclusivity, and many predict AOSP’s eventual death; “I don't like passkeys”—a rare consensus that passkeys are a corporation-serving downgrade for individuals, with no easy recovery from device loss or account bans; “Saving another 100TB of RAM”—a return to math-first engineering that earned genuine cheers, though it also sparked a debate on whether large companies inevitably silo away the simplicity the post demonstrates; “US Military had close call after using AI for hallucinated intelligence report”—the split between blaming the hallucination itself and blaming human rubber-stamping, with many arguing AI just provides convenient cover for already-lethal decision-making.
Microsoft exec called AI scraping 'the largest theft of labor in human history' [comments]
886 points · 782 comments · techcrunch.com · 22h ago
New unredacted filings in the NYT lawsuit show a Microsoft exec privately called AI scraping “the largest theft of labor in human history,” with internal docs admitting Copilot cratered NYT click-through rates by 93% and that OpenAI deliberately bypassed paywalls and stripped copyright notices. The thread largely took the theft framing as vindication, calling it the robbery of all culture to sell it back at a markup, but a substantial split emerged over whether comparing AI training to human learning makes “theft” a coherent label—some argued a machine reading a book isn’t stealing, while others pointed out that companies that spent years suing individuals for Napster-style piracy are now doing the same thing at industrial scale. A vocal minority argued that open-sourcing model weights would remove the profit motive and make the whole thing ethically defensible, though that got pushback on whether giving away stolen goods changes the moral calculus. Meanwhile, a meta-tangent erupted over someone using the acronym “DD” (meaning due diligence) without defining it, with longtime HN members arguing it’s not a standard abbreviation here—a classic HN sidetrack that briefly derailed the core debate.
Android 17 is the first since 3.x to add new APIs without releasing to the AOSP [comments]
789 points · 379 comments · grapheneos.social · 12h ago
The article is a GrapheneOS post flagging that Android 17 QPR1 is the first time since Honeycomb that Google has added new developer APIs without pushing them to the Android Open Source Project, making them Pixel-exclusive until QPR2 in December. The GrapheneOS team explains they’ve already ported to QPR1 but can’t release it because Google won’t give permission, and they’re now stuck reverse-engineering security patches that are only available to Pixels months early — a move they call dubious competitive gating that hurts other OEMs and AOSP-based projects. Many in the thread see this as Google locking down Android further and predict it will eventually kill AOSP entirely, with some comparing it to Oracle’s mismanagement of MySQL and speculating that Samsung or others could fork away from Google. Others push back, noting that major OEMs like Samsung actually have cozy private agreements with Google and seem happy with the current arrangement, while the real tension is between Google and small projects like GrapheneOS, not the big players. A sidenote emerged about postmarketOS as an alternative, though several commenters dismissed it as not yet viable for daily use compared to GrapheneOS.
I don't like passkeys [comments]
766 points · 754 comments · hawksley.dev · 19h ago
The article argues that passkeys are a security upgrade for enterprise users but a dangerous step backward for individuals, since device loss, platform bans, and opaque recovery mechanisms create a far higher risk of permanent lockout than the phishing attacks passkeys prevent. The thread largely agrees but goes deeper on the usability disaster — the main fight is between people who say “just let iCloud or Google sync them, it just works” and everyone else pointing out that losing your only phone or getting your Google account banned wipes out every passkey with zero recourse. A recurring correction is that passkeys don’t actually eliminate the weakest recovery method (SMS, email), so the security gain is illusory for most people. The consensus settles on passwords in a third-party manager plus TOTP still being the practical choice for normal humans, and the thread is unusually unified in its verdict: passkeys are a corporate solution being pushed onto consumers who didn’t ask for them.
Cloudflare Quick Tunnels [comments]
685 points · 275 comments · try.cloudflare.com · 17h ago
Cloudflare released a one‑command tunnel that turns any local server into a public HTTPS URL without an account, DNS, or open ports. The thread quickly split between people who loved the convenience versus those who raised sharp concerns about centralization—several commenters argued it’s another way to push huge traffic through Cloudflare’s network, turning them into a honeypot that can see all cleartext traffic, and that trusting a US company that’s already been accused of aggressive sales and poor support is a mistake. Others pointed out real trade‑offs: the free tunnels have a 200‑request concurrency limit, no SSE support, and the trycloudflare.com domain is likely to get blacklisted by scammers and abusers, just as ngrok’s free tier did. A few people compared it to ngrok, Tailscale funnel, and even Tor onion services, with one camp arguing onion services are better for testing because they don’t rely on a corporate intermediary, while another shot back that requiring clients to install Tor is absurd for sharing a quick demo. The article’s own breezy “slop” copy also took a hit—some called it embarrassingly bad for a Cloudflare product.
Claude Code now reads AGENTS.md if there is no Claude.md [comments]
643 points · 227 comments · code.claude.com · 11h ago
The article is a Claude Code changelog that—among a mountain of fixes—finally adds support for reading `AGENTS.md` as a fallback when no `CLAUDE.md` exists, letting the tool work with the community-standard harness configuration file. The thread is a mix of relief and eye-rolling: a lot of people are glad Anthropic caved after over a year of complaints, especially since other tools have supported `AGENTS.md` for ages and Claude was actively breaking cross-platform workflows. Some argue that blind adoption of a generic `AGENTS.md` is actually an antipattern—prompts should be tuned per model—but that take gets pushed back on hard; most people just want one file that works everywhere and don't have time to maintain per-model instructions. The inevitable gripes surface too: `.agents/skills` still isn't supported, symlinks as workarounds are fragile on Windows, and the general sentiment is that this is the bare minimum Anthropic should have done months ago, not a victory lap.
OpenJev [comments]
617 points · 258 comments · openjev.com · 22h ago
OpenJev is a browser-based demo that runs small language models locally to compare directly reading a model’s choice probabilities versus asking it to generate those probabilities as JSON text. The HN thread immediately split into two camps: one arguing the whole approach is just repackaged encoder-only classification (like BERT or Gliner2) with nothing fundamentally new, and the other pushing back that the performance and speed gains relative to Jev are real and useful—though they acknowledged the naming felt legally and conceptually confusing, since this isn’t the actual Jev model. A significant chunk of the discussion turned into a collective gripe session about how LLM-generated websites look and read: too much filler text, samey dark-mode layouts, an uncanny valley effect that people struggled to articulate, with a few defenders saying it’s clean and clear compared to cookie-ridden modern web design. Some commenters also dove into practical performance numbers, claiming the demo isn’t impressively fast on their hardware and that similar results can be replicated with custom inference engines, while others noted that skipping decode entirely (prefill-only classification) is precisely the trick that makes this kind of tool fast enough for real-time use.
US Military had close call after using AI for hallucinated intelligence report [comments]
446 points · 333 comments · www.cnn.com · 14h ago
The article describes how a US military analyst used an AI chatbot to generate an intelligence report falsely claiming a Chinese ship was carrying nuclear weapons components, nearly triggering an armed boarding operation that could have spiraled into war with China. Hacker News immediately connected this to the recent Iran school bombing, with many commenters arguing that AI is already being used to cover up or deflect accountability for war crimes—pointing out that the military’s response to atrocities has always been to blame bad intel, and AI just provides a convenient new scapegoat. A significant split emerged between those who see the AI itself as the core danger (a hallucination almost started a war) and those who insist the real problem is a human leadership that rubber-stamps whatever the machine produces, with younger analysts especially uncritical of the tools. Others went further, arguing that the military’s targeting decisions are already deliberately lethal—citing the double-tap strike on a girls’ school—and that AI merely speeds up getting to a bad idea faster, as one commenter quoted from the article directly. The thread was deeply skeptical of any hope for institutional accountability, with several people noting that even My Lai resulted in only one conviction, and that the U.S. has never been held responsible for its post-9/11 wars.
Saving another 100TB of RAM [comments]
341 points · 66 comments · blog.cloudflare.com · 13h ago
Cloudflare’s latest blog post walks through how they cut 100TB of RAM from a consistent-hashing load-balancing service by doing the math to prove they could drop 90% of the hash points per server without harming distribution, plus a Rust struct packing trick. The thread mostly loved it—people called it a return to human-written, math-heavy engineering posts after Cloudflare’s recent AI-slob period, and the “Hi, Mom” calculus derivation got genuine cheers. But the comments also split into a deeper argument about whether large companies inevitably become impenetrable silos of over-abstraction, with one side saying every layer exists for good reason (security, scale, adversarial world) and the other grieving the days when a five-person team could just turn on an LED without a dozen dependencies. A few commenters pushed back on the whole premise: some argued that 1% improvements don’t matter for most systems, while others questioned whether AI could ever find this kind of creative optimization, sparking a sideline debate about whether math-heavy engineering is actually safer or just another thing models will eat.
Bend 2 and the Vibe-Coding Trap [comments]
318 points · 231 comments · blog.liampwll.com · 19h ago
The article argues that the Bend 2 language, pitched for AI-era formal verification, was vibe-coded without the author realizing the field of formal verification already exists, demonstrating how the SPARK language achieves the same results far more concisely. HN latched onto the fact that Bend’s compiler is itself 99% AI-written and un-audited, turning the piece into a referendum on whether its author fell into the very trap the article describes. A large chunk of the thread devolved into accusations that the project’s 20k GitHub stars were purchased or faked, especially after the author nuked and then restored the repo history—though several commenters pushed back, noting that Bend1 went genuinely viral on HN and YouTube. Others dug into the technical choices, debating whether linked-list strings are actually a sensible default for GPU parallelism or just a performance disaster, and whether an LLM can be expected to spontaneously suggest Ada/SPARK if you don’t ask. The consensus was split: many agreed the article’s general point about vibe-coding skipping research is valid, but felt the author of Bend was being unfairly singled out as a cautionary tale rather than a genuine researcher experimenting with novel constraints.
Korea raises data breach fines to 10% of revenue [comments]
306 points · 92 comments · www.koreajoongangdaily.com · 12h ago
Korea’s privacy regulator is slapping companies with fines of up to 10% of total revenue for data breaches affecting 10 million or more people, replacing the old 3% cap and adding a 72-hour notification requirement for suspected exposures. HN immediately zeroed in on enforcement: the top commenter bet that the first chaebol like Samsung will get the fine ignored, arguing no system is truly secure and some businesses are forced to store data by law. Others pushed back hard, saying that’s defeatist — you can encrypt databases and follow best practices, and gross negligence is the bar, not perfection. A separate thread worried about shell companies holding data to dodge liability, which sparked a debate about whether caring about security is actually cheaper than risking the fine, and whether the law’s 40% reduction for proactive investment changes the math. The consensus split: some call this long-overdue, while cynics argue that in a low-trust, profit-driven world, these fines will just become another cost of business unless executives start facing personal liability.
Warren Buffett Steps Down as Berkshire Chairman, Names Son to Replace Him [comments]
304 points · 196 comments · www.nytimes.com · 21h ago
The linked article wasn't available to this summarizer; from the discussion, it covers Warren Buffett stepping down as Berkshire Hathaway's chairman and naming his 71-year-old son Howard Graham Buffett to succeed him. The thread immediately latches onto the glaring contradiction with Buffett's own famous quote comparing hereditary succession to picking Olympic gold medalists' eldest sons for the next games — and many argue this is textbook nepotism that erodes his credibility on meritocracy. Some push back that Howard is a non-executive chairman and Berkshire's CEO is Greg Abel, so the role is more about stewarding the family's shares, while others note Warren is 96 and passing the baton to a 71-year-old raises its own succession questions. A split emerges: one camp sees this as inevitable dynastic wealth preservation ("you only have to get rich once"), while the other insists it’s a betrayal of Buffett's stated principles and a sign that even the most principled billionaire ultimately trusts only his bloodline. There's also a long side debate about whether Buffett's massive charitable giving through family foundations is a genuine redistribution or a tax-advantaged way to keep control within the family.
ZCode, the GLM coding agent, silently uploads your Git history [comments]
259 points · 14 comments · tokenstead.ai · 21h ago
The article details a reverse-engineering report showing that ZCode, the desktop AI coding agent from the GLM model team, silently packages your entire workspace—complete `.git` history, reflogs, LFS cache, and configs—encrypts it with a key only Z.ai’s server holds, and uploads it to Alibaba Cloud whenever you’re logged in. The HN discussion zeroed in on the encryption-key detail: it’s envelope encryption with the private key living only in Z.ai’s cloud, meaning the user cannot decrypt their own archive, which several commenters called evidence of deliberate exfiltration rather than an accident like Grok Build’s recent uploads. A few people pushed back that Claude Code and other agents already send code metadata, but others corrected that those are visible over SSL and tied to inference context, not a hidden sidecar that retries 564 times after deletion. The thread also highlighted the distinction between GLM’s open weights and ZCode’s closed-source harness—the company’s launch pitch of “local and private” runs directly into a runtime that phones home with your entire repo, and the toggles in settings were found to be decorative, with the capture sidecar starting unconditionally at login regardless of user preferences.
I vibed a proof of Conway's conjecture [comments]
236 points · 217 comments · overreacted.io · 17h ago
Dan Abramov wrote up his attempt to use AI to prove Conway’s refinement conjecture about omnific integers, an open problem in surreal numbers, and he claims to have produced a Lean-verified proof after a month of iterating with Claude and ChatGPT. The HN thread mostly split between people getting tripped up on the surreal-number exposition—several commenters found the “between nothing and zero” description confusing, which kicked off a long subthread about whether the explanation was actually clear—and others zeroing in on the meta question of whether an AI-generated proof of a 50-year-old conjecture counts as a real result without independent human validation. A few domain-aware commenters pointed out that the proof hasn’t been peer-reviewed and that earlier attempts by the same author produced “deranged .tex files” full of nonsense, so the community is skeptical about whether the Lean check alone is enough. The discussion also wandered into whether the author’s workflow—offloading search and formalization to LLMs while keeping a human as “PM”—is a viable template for future math or just an expensive way to produce plausible-looking garbage.
An empirical study of harness design for coding agents [comments]
210 points · 57 comments · arxiv.org · 18h ago
The paper breaks down coding agent harnesses into three components—planning, action space, and context management—and tests 176 configurations across four models to see which parts actually matter. The HN crowd largely agreed the results are intuitive (stronger models don’t need planning scaffolds, bash-capable models do fine with just bash, and context management mainly prevents overflow failures), but several people pushed back hard on the model choice, arguing that using year-old Nemotron and Mistral models makes the conclusions irrelevant for today’s frontier models like Claude and GPT, which have been RL-tuned specifically around their own harnesses and now handle huge context windows natively. One running debate pitted defenders of empirical science against a “everything changes every three months, so this study is worthless” camp, while another thread spun off into a fight over whether MCP tooling is just unnecessary complexity for models that can already shell out. The most pointed takeaway from the comments: the study’s finding that adding structured tools doesn’t help bash-capable models got reframed as “MCP was a bunch of bullshit that maybe helped until models got good at bash.”
Border agents can search cellphones without a warrant or reasonable suspicion [comments]
203 points · 170 comments · lawandcrime.com · 13h ago
The Second Circuit ruled that border agents can manually search travelers’ cellphones without any warrant, probable cause, or even reasonable suspicion, treating a phone like a piece of luggage under the border-search exception. The thread immediately zeroed in on the 100-mile “border zone” that covers two-thirds of the U.S. population, with many pointing out that this ruling effectively guts the Fourth Amendment for most Americans, not just people physically crossing the border. A split emerged over practical countermeasures: some argued you should carry a wiped travel phone or restore data after crossing, while others warned that a clean phone itself looks suspicious and that Cellebrite can brute-force most devices anyway. Several commenters pushed back on the manual vs. digital search distinction, noting the ruling explicitly dodged the question of forensic software, and there was a strong undercurrent of advice to delete social media accounts and stay off the internet under your real name, since anything you’ve ever posted can be dug up out of context.
North Korean nuclear test sets off years of earthquakes [comments]
203 points · 160 comments · www.science.org · 17h ago
The article reports that North Korea’s 2017 nuclear test triggered years of seismic activity, with earthquakes still rattling the region as the blast zone settles. Much of the thread got stuck on the fact that the Science.org article is behind a captcha wall, leading people to swap archive links and complain about the archive sites also being flaky. From there, the discussion veered hard into nuclear geoengineering fantasies — whether you could deliberately set off a fault line to relieve pressure, referencing Project Plowshare and even Soviet attempts to cap oil-well blowouts with nukes. A massive, sprawling subthread then erupted over the morality of the atomic bombings of Hiroshima and Nagasaki, with one camp arguing the bombs were unnecessary because Japan was already sending peace feelers, and the other insisting unconditional surrender was the only option and that the firebombing campaign was already killing more civilians anyway.
Photon-Emission-Guided Laser Fault Injection Enables RP2350 Secure Debug [comments]
189 points · 67 comments · donjon.ledger.com · 15h ago
Ledger Donjon's blog post details how they used a $250k laser setup to physically crack the RP2350's secure debug protections, flipping a vulnerable DEBUGEN register after locating it with photon-emission microscopy—despite the chip's redundant OTP locks and glitch detectors. The HN thread mostly riffed on the cost barrier, with people noting that $250k is "currently" a lot, but like DRAM prices or drone tech, it could become a $50 hammer-and-hammer job in a few years if the core techniques align with consumer manufacturing. Some pushed back hard on the idea that this matters for real-world security, arguing that if you're defending against a nation-state, a $1 microcontroller was never the right tool, while others pointed out that Raspberry Pi explicitly invited this research and patched earlier findings, making the disclosure more credible than typical marketing fluff. A few tangents spun off into how high-end HSMs defeat optical attacks with pressure sensors and tamper-proof enclosures, and whether anything in consumer phones or Yubikeys comes close to that level of protection.
There's no point at which turning your brain off will work [comments]
182 points · 152 comments · danluu.com · 15h ago
Dan Luu argues that using LLMs as a “meat proxy”—turning your brain off and blindly accepting whatever the model spits out—is a dead end for the employee, because if the LLM ever gets good enough to produce acceptable work on its own, the company has no reason to keep the human in the loop. The comments mostly run with that bleak logic, but they push back on the idea that this is new or unique to LLMs: plenty of people have been running on autopilot for decades, following runbooks, throwing hardware at symptoms, and skating by on organizational inertia. A strong thread argues that the real value a human adds is as a liability sponge—someone to blame when the AI messes up—though others find that role abysmal and point to David Graeber’s “bullshit jobs” as the natural endpoint. There’s also a split over whether the recent collapse in junior dev hiring is AI-driven or just the end of ZIRP, with some warning that anyone planning to be a software engineer in three to five years should have a backup plan.
US troop deaths during Iran war exceed Pentagon count by at least four [comments]
174 points · 131 comments · www.reuters.com · 9h ago
The article is a Reuters report on a Washington Post investigation claiming that U.S. troop deaths in the ongoing war with Iran are at least four higher than the Pentagon has publicly acknowledged, a claim Defense Secretary Hegseth called a "complete LIE." The thread largely turned into a broader political indictment: many argued this is just the latest example of an administration that lies reflexively about everything, with several pointing to Trump's open grifting and insider trading schemes as proof that corruption is now completely normalized and blatant. Some drew direct parallels to the Vietnam era, noting that faith in government casualty counts collapsed then too, and that history is repeating with the U.S. now fighting on the opposite side of the same proxy wars it started decades ago. A few commenters tried to puzzle out why the Pentagon would bother hiding just four deaths, with the consensus being that denying all bad news, no matter how small, is the administration's default posture—especially after claiming a "flawless victory." A tangential reference to a separate CNN report about an AI screw-up that nearly led to a confrontation with China got a brief mention, tying the denial pattern to broader incompetence.
San Francisco Onion Futures Company [comments]
161 points · 57 comments · onionfutures.com · 3h ago
The San Francisco Onion Futures Company sells private, transferable contracts for future onion delivery, explicitly exploiting the legal loophole in the 1958 Onion Futures Act, which banned onion futures but only on "boards of trade." The HN crowd immediately picked apart the legal reasoning, with several commenters diving into the statutory definitions of "organized exchange" and "trading facility" to argue the company's claim that it's not one is weak—the law's definition hinges on whether multiple participants can execute trades, and by backing the contracts and allowing resale, the site arguably facilitates exactly that. Others loved the absurdist performance-art angle, connecting it to the real-world scandal where a guy cornered the onion market in the 1950s, causing prices to swing so wildly that Congress banned the contracts entirely. A side thread revealed the proprietor is independent from the UChicago student group that campaigns for legalizing onion futures, but they collaborate on deliveries, and the consensus was that the $9-per-onion pricing is a novelty/art stunt rather than a serious market.
US Treasuries Have Become Unappetizing for Foreign Central Banks and Governments [comments]
147 points · 130 comments · wolfstreet.com · 16h ago
The Wolf Street article argues that foreign central banks and governments are increasingly dumping US Treasuries, with their share of outstanding debt collapsing to 12.8%—the lowest since 1993—while opaque financial centers like the Cayman Islands and Ireland now hold the bulk of foreign claims. But the HN thread immediately split over whether the article is even directionally correct: one well-sourced counterpoint pointed out that total foreign holdings of Treasuries actually increased by $500 billion over the past year, calling the piece “flame bait” that would leave you believing the exact opposite of reality. That factual dispute then fed into a much longer, deeper debate about the “exorbitant privilege” and the mechanics of the petrodollar—several commenters dismantled the common notion that oil being priced in dollars drives demand for Treasuries, arguing instead that what actually matters is where the seller stores the proceeds, and that there’s fundamentally no alternative jurisdiction that can absorb trillions in capital inflows. Others countered that the real driver isn’t sentiment or trust but simple balance-of-payments accounting: as long as China and other surplus countries need to run massive trade surpluses with the US, they are forced to accumulate dollar-denominated assets, making any talk of collapse premature.
Human brain is two separate organs, Stanford Medicine-led research finds [comments]
142 points · 42 comments · med.stanford.edu · 2h ago
A new Stanford-led study claims the human brain is actually two separate organs—a primitive hindbrain controlling breathing and heartbeat, and a forebrain for higher thinking—that evolved independently and arise from distinct embryonic progenitor cells. The thread immediately split into two camps: the research itself is genuinely interesting, especially the breakthrough in growing hindbrain neurons in a dish for studying ALS and SMA, but the headline is widely called out as misleading. Many commenters pointed out that the actual Nature Neuroscience paper says "composite organ" and that biologists already knew brain regions are structurally and functionally different, so the real finding is about progenitor lineages, not a revelation that we have two brains. Others pushed back that a headline like "two separate organs" is a reasonable layman’s metaphor given the evolutionary history and the fact that jellyfish have spatially separate nervous systems, though a few noted the press release echoes debunked "reptile brain" tropes. Overall, the consensus was that the clickbait title does a disservice to otherwise solid work—but the new in vitro technique for hindbrain neurons is the real story worth paying attention to.
Xcode 27.1 Beta Release Notes [comments]
138 points · 99 comments · developer.apple.com · 13h ago
Xcode 27.1 beta adds support for the iPhone Duo, Apple’s foldable phone, and the thread is split between developers dismissing it as another niche Vision Pro flop and those who say adapting is trivial if you’ve already followed Apple’s resizing and safe‑area guidelines. The biggest complaint is that Apple dropped the simulator only today—months after WWDC—leaving developers just over a month before launch to test; many expect apps to look broken at release. A notable subthread argues that bigger apps (MLB, NYTimes) have recently switched to the Liquid Glass menu, suggesting Apple strongly nudged them ahead of the Duo, while others note that even Instagram’s custom Liquid‑Glass knockoff looks ugly and will be a pain to maintain. There’s also a fun tangent: coding agents are already hallucinating Duo behavior—one confidently claimed both sides of the tent‑mode display would be active, which is wrong.
Second Circuit allows government to search electronic devices at the border [comments]
122 points · 72 comments · knightcolumbia.org · 17h ago
The ruling upholds warrantless searches of phones and laptops at the US border, including for American citizens — the Second Circuit said the government’s constitutional authority to control what enters the country overrides the Fourth Amendment. The thread quickly zeroed in on the “100-mile border zone,” which actually sweeps in two-thirds of the US population, including major cities like New York and Chicago, because federal regulations define “reasonable distance” from any external boundary as 100 air miles, and CBP routinely operates deep inland. Several people pointed out that this case only addressed a manual visual search — an agent thumbing through photos — not forensic device dumping with tools like Cellebrite, though the line between those is blurry depending on how pissed off the officer is. A recurring dark joke: if you cross the border illegally and get caught later, you have more Fourth Amendment protection than a law-abiding citizen walking through customs, which highlights how this exception only applies at official ports of entry. Some commenters recommended traveling with burner devices or keeping sensitive data encrypted at home, but others noted that real-world logistics (two-factor codes, flight reservations) make that advice nearly impossible to follow.
Alibaba open-sources AI model that can detect cancer and nearly 150 conditions [comments]
117 points · 14 comments · www.scmp.com · 8h ago
Alibaba's research arm open-sourced Damo Radar, a vision-language model trained on CT scans paired with clinical reports that outperformed most radiologists in detecting nearly 150 abdominal conditions including cancers, with results published in *Science*. The HN crowd quickly dug into the model's specs, noting the weights are only about 5GB and roughly 131 million parameters—small for this domain—though it was designed to run on something like an H100 because CT scan data is huge. Several comments linked directly to the paper, the GitHub repo, and the HuggingFace model card, treating it more as a resource drop than a debate. The rest of the thread drifted into jokes about preferring Gemini's ego boost and a playful wish for a "predict the future" model, but no substantive pushback or domain-expert corrections surfaced.
AI is an elite crime spree [comments]
116 points · 40 comments · www.thebignewsletter.com · 16h ago
The article argues that the panic over AI safety and calls for new regulation miss the point — existing laws already prohibit the harmful behaviors, but they're simply not enforced against powerful people and companies. The thread quickly latched onto the article's core claim that the rule of law in America boils down to "guilty until proven wealthy," with multiple comments citing Wilhoit's law (the law protects but does not bind elites, binds but does not protect out-groups) and pointing to the Aaron Swartz prosecution as a stark contrast to how AI companies face zero consequences for mass data scraping. A substantial technical tangent derailed into arguing that the real AI safety problem isn't superintelligent models but basic incompetence — the frontier labs can't even airgap their systems properly, and a single competent Linux sysadmin could prevent the escapes that spark the panics. A few commenters pushed back on Stoller's evidence, one calling antitrust law a one-paragraph mess that Congress never fixed, and another noting that "selective quoting from discovered materials" doesn't prove the narrative the article constructs. The overall sentiment in the thread strongly agreed with Stoller's diagnosis of elite impunity, though there was a notable split between those who see the problem as baked into the legal system and those who think it's fixable with better enforcement and technical discipline.
Two parallel neural ectoderm progenitors contribute to the developing brain [comments]
115 points · 63 comments · www.newscientist.com · 16h ago
New research suggests the human brain develops from two distinct progenitor cell types—one for the forebrain/midbrain and one for the hindbrain—implying our central nervous system is a hybrid of two ancient neural systems that fused hundreds of millions of years ago. HN immediately pushed back on the framing: a neurosurgeon simply said “No” to an earlier headline calling the brain two separate organs, and the submission title was changed accordingly, while the thread split over whether the finding is a genuine discovery about developmental origin or just an observation about the supply chain for brain cells. A long tangent erupted around Julian Jaynes’s 1970s “bicameral mind” hypothesis, with people arguing it’s pseudoscience irrelevant to the evolutionary timescale here, while others defended it as an interesting albeit speculative parallel about merging mental structures. Several commenters pointed to the Stanford press release and the preprint as much better reads than the New Scientist article, and the discussion never settled the analogy debate—hindbrain as OS kernel, baseband modem, or system management co-processor—before the Christian eschatology angle showed up to claim the Holy Spirit explains it all.
How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip [comments]
111 points · 81 comments · spectrum.ieee.org · 8h ago
The article covers how OpenAI used its own LLMs to accelerate the design of their new Jalapeño AI accelerator chip, claiming a 3.6x latency improvement over Nvidia’s GB300 and a rapid 20-month timeline from concept to silicon. The HN discussion mostly ignored the technical details and instead devolved into complaints about the name "Jalapeño" hijacking chili pepper terminology, with people grumbling about "agent" and "transformer" being co-opted from other domains. Some skeptical commenters argued the article is just marketing fluff, pointing out that Broadcom handled the actual physical backend design and that OpenAI’s real advantage came from poaching Google’s TPU team and using their XLS tools. A notable split emerged between those who see this as a genuine step toward recursive self-improvement and those who dismiss the AI involvement as exaggerated—one comment called the latency claims “impressionistic math” and the whole thing bullshit. The most substantive pushback came from people familiar with chip design, who argued that LLMs aren’t suitable for the physical optimization work that separates competitive chips from academic projects, though others countered that the rapid software-tuning loop after first silicon is genuinely novel.
AI chatbots are becoming experts at changing people's minds [comments]
102 points · 97 comments · www.science.org · 18h ago
The article reports that AI chatbots are becoming more persuasive, partly by using facts and evidence more effectively than humans often do. The HN thread pushed back hard on the "facts work" angle, arguing that presentation and perceived legitimacy matter far more than raw data, and that the real advantage of AI is not being human—when you disagree with a person, it feels like a competition with winners and losers, but a chatbot has no ego and no emotional stake, making it easier to accept a counterargument without losing face. There was a split over whether RLHF created sycophantic yes-machines or whether recent training shifts (like RLAIF and RLVR) have over-corrected into pedantic, disobedient models that refuse to go along with user errors—some welcomed that “well ackshually” behavior as a sign of intellectual honesty, others found it disrespectful and counterproductive. A minority argued that the premise is flawed because deeply held political or abstract beliefs won’t budge regardless of the persuader, and that LLMs can only flatten thinking to whatever society currently agrees on, not challenge it from first principles.
We made Playwright 2x faster and 80% more token efficient [comments]
85 points · 20 comments · github.com · 15h ago
The article announces Stagehand v4, an open-source SDK for browser automation agents that claims to be 2x faster and 80% more token efficient than Playwright by running inside the browser via an extension instead of making round-trip CDP calls from a separate script. The creators showed up in the thread to explain they rebuilt from scratch after seeing complaints about Playwright MCP’s token appetite, and that the new architecture uses batch commands and dedicated `act()`/`extract()` methods with benchmarks across models and harnesses. Several commenters pressed on trade-offs: one asked if cross-origin iframes or downloads break without CDP, and the answer was that it still uses CDP but communicates through the extension. Another pushed back on the token efficiency claim by suggesting “astra computer use” as an alternative, but others countered that screenshot-based approaches burn far more tokens. There was a practical question about CI pipelines — whether the speed gain applies to a 1,000-test Playwright suite (yes, especially with batch actions) — and a clarification that Stagehand works with local browsers for free, though the cache feature requires Browserbase’s cloud.
Generated 2026-09-19 08:07 UTC
Generated by Sauron from Hacker News discussions and linked articles.