HN Brief: 2026-08-12

France to ban unsolicited telemarketing calls [comments]

1016 points · 481 comments · www.lemonde.fr · 23h ago

France is banning all unsolicited telemarketing calls starting next week, shifting from an opt-out system to an opt-in model where businesses need prior consent before calling anyone. The conversation quickly moved beyond the article itself, with people immediately pushing to extend the ban to door-to-door salespeople, street marketers, and religious proselytizers. A big split emerged between those who see these interactions as fundamentally similar to telemarketing intrusions and others who argue that religious speech and charitable solicitations are constitutionally or morally different. Several people with domain knowledge pointed out that the previous French system (Bloctel) was recently hacked and that scammers never obeyed the no-call list anyway, so the real question is whether this law will actually stop the fraudulent calls or just the legitimate ones—and some commenters noted that France already mandated telcos to block spoofed numbers, which may do more heavy lifting than the new consent requirement.

Stealing Reasoning Traces from Proprietary LLM APIs [comments]

606 points · 272 comments · stolen-thoughts.com · 18h ago

Researchers extracted hidden chain-of-thought reasoning from OpenAI, Anthropic, and Google's proprietary models by grabbing the encrypted reasoning blocks returned to users and replaying them into weaker sibling models (like Haiku or mini variants) that lack the same anti-distillation safeguards, causing the weaker model to decrypted and spit out the stronger model's reasoning in plaintext. HN immediately recognized this as an inevitable consequence of allowing users to switch models mid-conversation—the encrypted blocks have to be portable across sessions and models, so they're encrypted with a shared key rather than a per-session or per-user key, meaning this is a fundamental architectural choice, not a cryptographic flaw. Several people noted this is already patched according to the authors, and the fix will likely break model-switching mid-conversation or force per-user encryption keys, both of which degrade user experience. The thread also flagged that the recovered traces contain real secrets—API keys, passwords, access tokens—and that the models are thinking in aggressively token-efficient "grug speak," which was independently confirmed by other research. A smaller split emerged around whether this is actually a distillation technique for Eastern labs, with one camp arguing it's just stealing and another countering that the encrypted blocks are clearly intended to be protected.

England set to be one of the first countries to eliminate hepatitis C [comments]

530 points · 375 comments · www.bbc.com · 19h ago

England is on track to become one of the first countries to eliminate hepatitis C, having already met the target of treating 80% of known cases through initiatives like A&E blood tests and free home-testing kits, with antiviral tablets curing over 95% of cases in 8 to 12 weeks. The thread quickly zeroed in on the fact that the article specifies *England* rather than the UK, sparking a long, tangled debate about devolution—people explained that Scotland, Wales, and Northern Ireland each run their own NHS and public health bodies, so they pursue the same WHO elimination goal independently with different timelines and reporting. That discussion devolved into a surprisingly deep argument about whether internal UK borders actually matter for disease control, with some insisting it’s trivial to drive from Scotland to England while others pointed out that parts of Scotland are harder to reach than continental Europe. A separate tangent saw people comparing the UK’s public health wins against the US’s current struggles with measles and polio, with one side calling it a trade offer and the other side deadpanning that they’re plenty happy with their own politics. There was also a minor semantic scuffle over whether England qualifies as a "country" when the UK is also a country, with several participants genuinely educating each other on the recursive nature of sovereign states and constituent nations.

Compression is prediction [comments]

486 points · 203 comments · ngrok.com · 12h ago

The linked article from ngrok walks through the fundamentals of compression — transforms, models, and entropy coders — to make the case that LLMs and compressors are solving the same mathematical problem: prediction is compression, measured by bits per symbol. The thread mostly ran with the philosophical implications rather than the technical details, with a strong contingent arguing that this framing demolishes the "stochastic parrot" critique — if training is just finding a compact approximation of the process that generated the data, then novel ideas are exactly what you'd expect from a good compressor, the same way Newton's laws emerge from compressing orbital data. A smaller group pushed back hard, pointing out that the equivalence runs one way: LLMs are great compressors, but traditional compressors like gzip are terrible predictors, so the overlap is weaker than the article suggests. Several people linked Grant Sanderson's ongoing "Compression is Intelligence" video series and the two-decade-old Hutter Prize as prior art, noting that this isn't a new insight so much as a core information-theory fact that's become newly relevant for arguing about what LLMs can actually do.

OpenAI’s head of ethics leaves less than a year after joining [comments]

405 points · 401 comments · www.ft.com · 19h ago

The linked article wasn't available to this summarizer; from the discussion, it's about the departure of OpenAI’s head of ethics less than a year into the role. The thread immediately split into a much broader debate about whether companies—especially large tech corporations—can be ethical at all, with one side arguing that corporations are inherently amoral profit-maximizers and the other pushing back with examples of local businesses that maintain strong community ethics. The conversation swerved hard from OpenAI into a critique of Silicon Valley’s homelessness crisis, zoning laws, and the economics of rent, where a lengthy subthread argued that high local productivity directly drives high rents and that construction policy can’t easily break that cycle. Someone pointed out that this person was also Meta’s chief ethicist, which got a quick dismissal as “ethicist-for-pay,” while another comment thread devolved into comparing corporate moral failure to East Berlin’s internal passport system. The consensus among the most engaged commenters was that corporate ethics roles are essentially performative, with the real debate being whether any countervailing force—like internal procedures or individual morality—can survive being outcompeted by less scrupulous actors.

Mojo 1.0 [comments]

374 points · 177 comments · www.modular.com · 15h ago

Mojo has officially hit 1.0, positioning itself as a stable, production-ready systems language with Pythonic syntax, built on MLIR and aimed at high-performance compute, especially GPU kernels. The HN crowd is deeply split: a sizable chunk dismisses the release as uninspiring because the compiler is still closed-source (promised open-source by 2026), calling it a marketing stunt and pointing to Lattner’s Swift history as a warning that “stable” means little, while others argue the real value is in replacing CUDA and Triton for AI‑workload kernel writing, and the Qualcomm acquisition validates the approach. The “why learn a new language in the age of LLMs” debate flares up too, with the counter that Mojo’s niche—writing fast, memory‑safe GPU code that Python can’t handle—gets easier to adopt when LLMs can help translate, not harder. A quieter but sharp thread notes the lack of performance benchmarks against PyTorch + Triton or ThunderKittens, which feels like a glaring omission for a 1.0 claiming production readiness.

Go is an ideal language for AI-assisted software engineering [comments]

357 points · 408 comments · developers.googleblog.com · 15h ago

Google's blog argues that Go is uniquely suited for AI-assisted development because its design prioritizes readability and maintainability over cleverness—things that matter more when humans spend their time reviewing AI-generated code instead of writing it. The HN discussion largely agrees with the premise but immediately pivots into a heated Go-versus-Rust debate, with one camp insisting Rust's stricter compiler catches more AI slop at compile time, while others counter that Go's simplicity, fast compilation, and uniform tooling make it easier for both LLMs and human reviewers to work with. A significant split emerges over whether Rust's steep learning curve is a problem for AI-assisted workflows—some argue the borrow checker just means the LLM has to fix more errors, while others say that's exactly the point, since those errors get caught automatically rather than at runtime. A few commenters dismiss the whole framing entirely, suggesting the post is just Google marketing to train LLMs on Go, while others bring up Zig and TypeScript as alternatives with different trade-offs. One recurring theme is that nobody's actually reading AI-generated code anymore—several people describe workflows where they just have the LLM summarize its own output, which undercuts the article's core argument about readability being paramount.

Nvidia's Risky Business [comments]

331 points · 166 comments · stratechery.com · 21h ago

Ben Thompson’s latest Stratechery post draws a direct line from Jay Cooke’s railroad bonds triggering the Panic of 1873 to Nvidia’s current gamble on financing AI data centers through new debt structures with Apollo and BlackRock, arguing the hyperscalers are blowing through bond markets and Google is already tapping equity. The thread largely accepted Thompson’s historical framing as apt but split hard on the Nvidia-specific thesis: some argued Nvidia’s CUDA moat has survived twenty years of challenges and will carry it through robotics and beyond, while others countered that the real threat isn’t competitors’ hardware but that the market is pricing in *infinite growth in demand* when efficiency gains (Jevons paradox aside) could easily flatten the curve, making those bonds a huge liability. A strong undercurrent pushed back on the idea that Nvidia’s dominance is permanent, drawing parallels to Cisco in 2000 and Intel in 2010—dominant players that kept selling but saw their margins and stock prices crater when growth expectations recalibrated. One cogent comment noted that the first-order assumption (demand for compute is huge) is almost certainly correct; it’s the second-order assumption (demand must *grow every year*) where these investment theses tend to implode, and the article’s own warning about unmarked pension-fund risk echoes that same fragility.

London Underground begins scanning passengers' faces [comments]

310 points · 340 comments · www.btp.police.uk · 22h ago

The linked article wasn’t available to this summarizer; from the discussion, the British Transport Police are expanding a live facial recognition trial across London Underground stations, claiming it’s intelligence-led and targeted at wanted individuals. The thread immediately splits into two camps: one side argues this is straightforward authoritarian drift, pointing out the police have already used facial recognition at protests and will inevitably track people for minor offenses or political speech. The other side pushes back hard, saying the UK’s surveillance state is vastly exaggerated in popular imagination, that police actually control very few cameras and mostly have to beg businesses for footage, and that London is safer than any comparable US city partly because of robust transport policing — with one person arguing the paranoid American view of London is simply misplaced. A substantive sub-thread also breaks out over whether Portugal’s ban on private recording in public is preferable, only to be met with multiple people pointing out Portuguese driving is statistically dangerous and that the safety-vs-privacy tradeoff is real. The deeper argument is less about the tech itself and more about whether incompetence is a reliable defense against eventual abuse, and whether mass surveillance is a proportional response to terrorism given how few people actually die from it.

Apple Silicon and macOS VMs: Faster LLM Inference with llama.cpp [comments]

290 points · 43 comments · github.com · 17h ago

A team called Cua published a blog post showing they tricked Apple’s Virtualization.framework into letting a macOS guest VM use the full Metal GPU capability of the host Apple Silicon chip, which made llama.cpp inference run 7x to 16x faster than a stock VM, nearly matching bare-metal speeds. The thread spent a lot of time clarifying that this isn't a general llama.cpp speedup—it's specifically a fix for a VM problem where the paravirtual GPU reports a conservative Metal profile, causing apps to select slower kernel paths. Several commenters pressed on the mechanism, noting that it's a process-scoped compatibility layer that intercepts Metal capability queries rather than actual GPU passthrough, and that the technique relies on private Metal internals and won't work for hardened executables. A side debate broke out over whether Apple deliberately sandbags the VM GPU profile for compatibility across chips or simply hasn't bothered, with some pointing to Apple's hostile relationship with NVIDIA as context for why the ecosystem has to hack around the platform. Overall the tone was impressed by the speedup numbers but wary of the narrow scope—nobody thinks this helps you run llama.cpp faster outside a VM, and several people wished the title had been clearer about that from the start.

Grok Bot [comments]

266 points · 218 comments · x.ai · 14h ago

The linked article from x.ai pitches "Grok Bot" as an autonomous agent that logs into your apps, learns workflows by watching you, and works 24/7 across projects for $200/month. The thread is deeply skeptical, with the dominant angle being that nobody in their right mind would hand over credentials to a product from Musk’s camp—many argue the brand is toxic enough that enterprises outside the U.S. will refuse entirely. A long, heated subthread tears apart the claim that prompt injection is "largely solved," with people pointing out that a 2.6% failure rate is catastrophic for security and that the fundamental data/command intermingling issue remains an arms race attackers will win. Others zero in on the absurdity of trusting an agent to run unsupervised with access to email, bank accounts, and JIRA, though a few defenders argue you can give the bot its own limited accounts—countered by the reality that you’re still on the hook when it does something illegal or stupid. A separate vein of comments dismisses the whole release as yet another commoditized, undifferentiated "make a clone of Foo" wrapper, noting that Cursor’s brand value was the only thing keeping enterprise doors open and that Musk is torching it.

More than 10 firms pay up to $100k a month for access to Truth Social posts [comments]

256 points · 290 comments · www.bbc.com · 22h ago

The BBC reports that Trump Media is selling a $60,000–$100,000/month API feed to high-frequency trading firms, giving them a several-minute head start on the president's Truth Social posts before the public sees them. The thread immediately labeled it "insider trading as a service," with most people arguing this is a blatant, unprecedented monetization of executive office and that the business model is essentially selling privileged access to the commander-in-chief's real-time utterances. A vocal contingent used the news to launch into broader arguments about the current administration being the most corrupt in U.S. history and comparing the country's trajectory to the fall of Rome, sparking a deep debate about whether historical analogies to Roman decline hold water or are dramatically overblown. The conversation split between those seeing this as a clean, direct case of corruption and others who pushed back, arguing that the systemic rot has been there for decades and this is just the least subtle version of it.

Woman pulled over twice after Flock-linked software connected her to homicide [comments]

254 points · 230 comments · guessingheadlights.com · 14h ago

The article recounts a Wisconsin woman being pulled over at gunpoint twice in one week because a Flock license-plate camera flagged her car for a homicide alert that police had forgotten to delete. Hacker News was furious—not just at the technical failure, but at the original piece’s shrug-and-move-on tone, with several people tearing into the reporting for treating a preventable armed stop as a mere “hiccup.” The discussion quickly turned to the broader surveillance debate: some described local efforts to vandalize Flock cameras and speak at county commission meetings against expansion, while others argued that the real danger is the police culture of defaulting to overwhelming force, even when the alert could be stale or the driver might not be the suspect. A significant split emerged on whether crowdsourced home-owner surveillance networks would be a community-safety improvement or just an even more unaccountable version of Flock—with one side calling it a dystopian nightmare and the other side genuinely enthusiastic about building it.

Nvidia Nemotron 3.5 Lightning and NeMo Switchyard [comments]

220 points · 113 comments · blogs.nvidia.com · 12h ago

NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model built for high-volume specialized agentic tasks, alongside NeMo Switchyard, an open-source library for routing prompts to the most capable model across a mix of providers. The HN thread immediately zeroed in on a fundamental technical tension: the repo’s README marks Switchyard as "Experimental software. Not for production use," directly contradicting the press release’s enterprise-ready tone, and people dug into the code only to find no clear handling of prompt caching across model switches, which cuts to the heart of whether routing actually saves money in long-running agent sessions. A detailed technical sub-thread emerged explaining that switching models doesn’t ruin cache efficiency—you keep a warm copy of the KV cache per model and the real cost win comes from generating fewer expensive output tokens—but skepticism ran high that the router itself would need to be as smart as the frontier models it routes between, making the whole thing feel like snake oil to some. Others pointed to Meta’s new dense 30B model as performing better on benchmarks, and a smaller but sharp debate kicked off about whether smaller, efficient models are the real future versus chasing multi-trillion-parameter scale, with one person literally betting their company on that thesis.

WorldClaw Agentic 3D open-world generation at scale [comments]

206 points · 62 comments · tencent-hunyuan.github.io · 10h ago

The linked article wasn't available to this summarizer; from the discussion, this is a Tencent project that generates 3D open worlds using a pipeline of image models and LLMs, composing scenes by having an AI generate a 2D image, then extracting objects into 3D. The thread immediately split into two camps: people who think procedurally-generated worlds inherently lack the handcrafted soul of games like Skyrim or Cyberpunk, and people who argue that AI storytelling is advancing fast enough to close that gap—citing an award-winning AI-written story from Granta as evidence that even literary judges can be fooled. A practical debate broke out about the actual pipeline: there's no code released, the GitHub repo is empty, and multiple practitioners confirmed the core technique (image model → extract objects via SAM → 3D via tools like Tripo) is real but noted that game-ready assets need aggressive decimation to avoid unusably high vertex counts. One commenter made the sharp observation that LLMs are essentially "amazing and flexible constraint solvers" for procedural generation, and that this approach is less about new models than about integrating existing ones cleverly—but the examples shown are clearly cherry-picked, with buildings placed in water and other oddities that betray the lack of human polish.

What I learned by putting GitHub Copilot behind a MitM proxy [comments]

179 points · 24 comments · www.lighthousenewsletter.com · 21h ago

The article is a deep dive where someone intercepts GitHub Copilot's network traffic with a MitM proxy to see exactly what it sends and stores. HN immediately latched onto the finding that Copilot's "recent edits" context can pull in secrets from files like `.env` with no default exclusion rules, sparking a debate about how to actually solve that problem — suggestions ranged from secret injection proxies to eBPF-based monitoring that bypasses TLS entirely. A few people pushed back on the author's broader conclusion that "context is becoming the product," arguing that even without carefully curated context, high-end LLMs perform just as well, and that more context just adds failure points. There was also a minor factual correction that the OpenAI Codex client is actually open source, contrary to what the article implied.

New Bedford police officer accused of using Flock cameras to track ex-partner [comments]

176 points · 80 comments · newbedfordlight.org · 6h ago

A New Bedford police officer is accused of using Flock license plate readers to track and stalk an ex-girlfriend after she ended their relationship, with the officer allegedly running hundreds of searches and accessing camera networks outside the city. The thread immediately turned into a broader debate about surveillance abuse, with many pointing out that this exact scenario—cops using always-on tracking tech against romantic partners—was dismissed as a theoretical risk in previous Flock discussions. Several people noted that the officer was one of the top users of the system and that her search reasons (“traffic infraction”) looked like thin cover, raising questions about why a patrol officer has unfettered access to a nationwide location database without extra authorization or automated flags for suspicious search patterns. Some defended the tech’s crime-solving utility, but the counterargument was brutal: this isn't hypothetical anymore, it's happening, and the classic "nothing to hide" line gets demolished by the fact that being in public doesn't mean you consent to perpetual, aggregated tracking by a system that can be weaponized by anyone with a badge. A few commenters also flagged that Y Combinator is an investor in Flock, which led to a split between those accusing HN of flagging the story down and others arguing it's just not tech-newsy enough for the front page.

Company Offering '100% Human-Written, Never AI' Medical Research Is 100% AI [comments]

173 points · 37 comments · www.404media.co · 5h ago

An investigation by 404 Media found that Research Gold, a site promising "100% human-written, never AI" medical research services like systematic reviews and meta-analyses, is entirely a fraud. The PhD methodologists they list as their team are completely fabricated, with AI-generated profile photos and no online footprint, while real academics they scraped from LinkedIn had their identities used without permission. When the reporter called, an AI assistant named Sarah refused to admit it wasn't human and kept trying to sell the service, and email exchanges were also clearly AI-generated. The HN thread largely took the article as further evidence of the ongoing collapse of trust online, with people noting that entire company personas can now be automated from registration to customer service, making fraud cheap and easy. A split emerged on whether this is primarily a problem with the technology or the perverse financial incentives of anonymous website ownership, though there was broad agreement that the scam is a perfect example of the "inversion" where companies now pretend AI work is human because the output has gotten good enough to fool customers.

Manus will return to operating as an independent company [comments]

154 points · 77 comments · manus.im · 17h ago

Manus announced it's splitting from Meta and returning to independent operation, which means it has to delete user data generated after Meta's acquisition for certain accounts to satisfy regulators. The HN thread quickly zeroed in on which regulators were forcing the unwinding — the consensus was that China's government blocked the deal and is literally preventing Manus's founders from leaving the country, treating this as a technology-transfer and capital-flight control issue rather than antitrust. A big chunk of the discussion shifted to whether anyone actually uses Manus anymore, with a split between people who found it expensive and slow compared to Claude Cowork or ChatGPT Work, and defenders who say its browser-use and harness work still outshines competitors for non-coding tasks. Several people argued the product's whole premise got hollowed out by model improvements, since the "bitter lesson" is that frontier models no longer need the wrapper Manus provided.

llama.cpp [comments]

154 points · 71 comments · llama.app · 3h ago

The linked article is basically the official website for `llama.cpp`, the open-source project that lets you run large language models entirely on your own machine, positioning itself as a direct, local alternative to cloud APIs. The Hacker News thread largely treated this as a legit launch from the core `llama.cpp` team, though a huge chunk of the discussion immediately swerved into a debate about the `curl | bash` installer on the site, with many people arguing it's no sketchier than cloning a repo and building, while others insisted it lacks the cryptographic guarantees and audit trail of a proper package manager. A few users with specific Intel Arc and AMD GPU hardware reported real friction, with broken ROCm support on integrated chips and a failed multi-day compile being met with pushback from others who said compilation should take minutes. The thread also split into camps comparing `llama.cpp` directly against Ollama, with one side arguing Ollama is just a slower wrapper that was sketchy about attribution, while others conceded Ollama's ease-of-use was the main reason to pick it over `llama.cpp` until now.

Making holograms with a pen plotter [comments]

154 points · 16 comments · blog.jordan.matelsky.com · 13h ago

The article walks through using a pen plotter to scratch reflective ridges into CD jewel cases, creating simple holograms that show apparent depth when lit by a point source. The thread loved the "old Internet" spirit of the project, with several people immediately recognizing it as a reinvention of William Beaty's hand-drawn "abrasion holography" technique from the 1990s, just automated with a plotter. A few commenters pushed back on calling these "holograms" at all, arguing the term is thrown around too casually, though the author himself preemptively acknowledges the debate and invites a better name. Some practical tangents emerged: one person noted that Steve Mould has a great explainer video on the physics, another suggested augmenting the plotter with a piezoelectric scanner for finer line spacing, and a few folks shared their own experiences making similar scratch holograms with a compass, confirming the technique works but isn't exactly the full holographic effect.

Show HN: Git-knife – Edit commit messages, authors, and dates like a spreadsheet [comments]

151 points · 96 comments · github.com · 16h ago

The article introduces git-knife, a desktop GUI that lets you edit commit messages, author identities, and dates in a spreadsheet-like table by shelling out to `git commit-tree` without touching file contents. The HN crowd was largely intrigued by the tool's niche utility, with several people sharing real use cases: fixing a misconfigured email across a batch of commits, reconstructing historical timestamps for archival projects like Simon Willison's repurposing of Tim Berners-Lee's original browser, or re-attributing commits after a PR merge went wrong. A vocal chunk pushed back hard on the practical wisdom of making history rewriting this easy — one person summed it up as "making easy something you normally should not do" — though the author and others countered with the surgery-versus-knife analogy, arguing that dangerous tools are exactly what you want when you actually need precision work. The conversation also split into a lengthy meta-debate about whether the project was LLM-generated, with some complaining about AI-tainted documentation and code comments, while others defended using LLMs to build niche utilities they'd never bother to write from scratch.

CFTC declares market emergency, orders Kalshi to continue to operate in New York [comments]

143 points · 129 comments · www.cftc.gov · 7h ago

The CFTC just invoked emergency powers to order Kalshi to keep operating in New York after the state AG filed a lawsuit seeking to shut down what it calls gambling, with the feds asserting that state gaming laws can't preempt federal derivatives regulation. The Hacker News thread immediately splits on whether Kalshi is a legitimate futures exchange or a grift-subsidized betting platform, with a long, technical back-and-forth about the difference between prediction markets and insurance—one camp argues Kalshi creates risk out of thin air (you’re fine unless you bet), while the other points to real-world use cases like bars hedging against World Cup traffic losses. A strong contingent calls the CFTC’s move transparent corruption, noting Donald Trump Jr. is a strategic advisor with equity, and predicting the Supreme Court will side against the agency. The thread also raises a legal nuance: the NY AG may have only requested a state-specific TRO, not a national one, meaning the CFTC’s sweeping preemption order could be unsupported by the actual filing.

How we used to get jobs: A newspaper classifieds story [comments]

142 points · 106 comments · ironicsans.ghost.io · 13h ago

The article is a personal essay about a photographer who landed a dream job at Ralph Lauren in 2000 through a newspaper classified ad, written as he digs through his digital archives to confirm the memory and wonders if he's misremembering details like a boss saying "butt load of resumes." HN immediately zeroed in on whether the pre-internet job market was actually "simpler" or just differently broken, with a sharp split: some argued that newspaper classifieds and walking into offices meant less competition and more signal, while others pushed back hard that those methods heavily rewarded social capital and insider knowledge, making it brutal for anyone without the right parents or friends to tell them how to dress or where to look. A lot of people shared stories of physically showing up at companies uninvited, cutting off a broken hand cast to maintain a strong handshake, or finding that small businesses still hire that way today, while the thread also noted that the experience of apartment hunting in NYC via the Village Voice was its own kind of broker-gatekept hell. One commenter pointed out that the mangled HN headline ("How we used to get jobs" vs. something more engaging) itself became a minor meta-topic, and another dryly noted that HR departments still run newspaper ads for PERM labor certifications.

Nvidia Nemotron 3.5 Lightning [comments]

119 points · 29 comments · huggingface.co · 18h ago

Nvidia released a new open-weights language model, Nemotron 3.5 Lightning, a 30B-parameter Mixture-of-Experts model with only 3B active parameters that uses a hybrid Mamba-2 and Transformer architecture. HN was largely dismissive of the raw benchmark scores, noting it lags behind equivalents like Qwen 3.6, but the crucial defense was that this is one of the few truly open models where the training data, recipes, and pipeline are fully released alongside the weights. There was strong consensus that Nvidia has a unique and durable incentive to keep pushing open models because it sells the hardware that runs them, and several practitioners reported the model runs fast locally (100 tokens/sec on a Mac) and writes solid code, though one test showed it comically fails the "car wash" reasoning puzzle. Some commenters noted the quantized NVFP4 version shows a significant accuracy drop on SWE-bench compared to the BF16 release, and there was an interesting tangent about whether a Qwen 3.8-derived MoE would be more exciting than their upcoming dense model.

The US tried to stop cartel money-laundering; devastated mom-and-pop businesses [comments]

115 points · 187 comments · www.theguardian.com · 17h ago

The Guardian reports that a Trump-era FinCEN policy lowered the cash-reporting threshold for border-zone businesses from $10,000 to $200, ostensibly to fight cartel money laundering, but instead crushed the money-transfer and bill-pay services that mom-and-pop shops provide to unbanked immigrant communities. The HN thread outright rejected the article's framing, with a strong consensus that the policy’s real, intended goal was never cartels—it was to make life unlivable for immigrants by cutting off their financial lifelines and creating a climate of surveillance and fear. The top comments argued that the “side effect” exactly describes the purpose, pointing to the mismatch between going after $200 burrito-and-money-order transactions while Trump pardoned and protected big-money launderers, including his own financial ties flagged by Capital One for AML concerns. A tangential but heated debate broke out over whether Trump voters in immigrant communities are genuinely duped or are knowingly signing up for the cruelty, with several people invoking Poe’s Law after sarcasm about Trump’s own debanking was misread as a MAGA talking point.

OpenSSH 10.5/10.5p1 [comments]

111 points · 35 comments · www.openssh.org · 14h ago

The linked article is the official release notes for OpenSSH 10.5, a new version shipping with several security fixes and quality-of-life features like a `-Z` flag to list which SSH keys will be tried for authentication. The thread immediately zeroed in on one line in the release notes where the OpenSSH team explicitly says they "very much welcome" AI-assisted security bug reports, which kicked off a debate about whether the project actually accepts AI-generated code or just AI-aided bug discovery. Several people dug up OpenBSD project leader Theo de Raadt's recent messages rejecting AI contributions on copyright grounds, and a developer chimed in to clarify that the distinction is between "AI reports" (welcome) and "AI-written code" (not welcome, due to legal risk from potential copyright infringement in training data). The practical takeaway most people agreed on is that the team's strategy shift—moving to more frequent releases instead of batching fixes—makes sense given that adversaries are surely using the same tools to find unpatched vulnerabilities.

US hires over 2k video gamers as air traffic controllers [comments]

110 points · 95 comments · www.cbsnews.com · 8h ago

The CBS News piece reports that Transportation Secretary Sean Duffy claims the FAA has hired over 2,000 video gamers to train as air traffic controllers, hitting 94% of its hiring goal. The Hacker News thread immediately pushed back on the framing, arguing this isn't a novel "gamers to ATC" pipeline but rather a desperate recruitment drive that happens to draw from a demographic—young people under 31—who nearly all play video games anyway. Several people with domain knowledge pointed out the real bottleneck: the FAA's medical and aptitude screening, which historically weeds out over 90% of applicants, especially those with ADHD diagnoses, poor vision, or any psychiatric history. A strong split emerged between those who see this as a harmless marketing stunt and others who view it as dangerous spin meant to deflect from chronic understaffing that has contributed to near-misses and crashes. The thread also got sidetracked into a debate about whether sim racing or VATSIM (volunteer online ATC) actually transfer useful skills, with pilots chiming in that the radio work and situational awareness from flight sims saved them thousands in training costs.

CSS properties you should know for better text designs [comments]

108 points · 18 comments · master.dev · 14h ago

The article walks through five CSS properties—`background-clip`, `vertical-align`/`align-content`, `box-decoration-break`, `letter-spacing`, and `text-combine-upright`—that give developers more control over typographic styling beyond basic font choices and spacing.

The HN thread quickly diverged from the article itself into a debate about browser compatibility and when it's safe to use modern CSS. Several people pushed back hard against the article's recommendations, pointing out that `background-clip: text` has significant implementation bugs across browsers (especially Firefox) and that relying on `color: transparent` will invisibilize text entirely for anyone whose browser doesn't support the clip. There's a split on letter-spacing: one commenter argues it's "massively overused" and should never appear in body text, citing ligature-breaking and readability issues, while others counter with niche cases like Fraktur or traffic signage where it's appropriate. A separate thread corrects the article's naming—the CSS property is `box-decoration-break`, not `box-decoration-mode`. The conversation also turned into a broader resource-sharing session on where to learn modern CSS, with strong consensus that MDN and Baseline (30-month support windows) are the gold standard, and that targeting "last two major browser versions" is now meaningless since browsers ship major versions every few weeks.

The brain may be about to have its Ozempic moment [comments]

104 points · 154 comments · economist.com · 12h ago

The Economist article, behind its paywall as usual, is about the prospect of drugs that could safely reduce the need for sleep, paralleling how Ozempic revolutionized appetite control. Hacker News immediately dove into the social and biological implications, not the drug science. A deep thread split on whether sleep is a sacred, fundamental biological function—like death—that we shouldn’t mess with, or a boring chore that only seems essential because we haven’t figured out a better engineering solution. The strongest pushback came from someone arguing that sleep’s primary role isn’t waste-cleaning but likely a form of contrastive learning that the brain can’t shortcut, warning that evolution couldn’t optimize away unconsciousness despite immense pressure, so any replacement pill would backfire. Elsewhere, the conversation pivoted to the inevitable dystopia: even if the drugs work perfectly, employers would pressure everyone into taking them, making 16-hour workdays the new baseline, and anyone who refused would become unemployable—a repeat of the ADHD medication dynamic in schools.