HN Brief: 2026-09-09
Today’s HN was dominated by the OpenAI Navier–Stokes scandal—multiple threads covering the same accusation that the company scooped a mathematician’s unpublished work, offered a coerced co-authorship, and spent $22 million in compute to claim a Millennium Prize before the humans could publish. That story bled into a broader pattern of anxiety about AI labs racing without guardrails: an Anthropic researcher resigned warning of self-improving superintelligence, Terence Tao warned math problems are being “non-renewably mined” by AI, and a McSweeney’s satire about in-office “Associate Slop Doula” work hit uncomfortably close to home. Meanwhile, practical privacy dread surfaced in threads about LG TVs spying even when offline and a DHS unit analyzing financial transactions to flag drivers for no-cause stops.
Dig into “On the Navier–Stokes Millennium Prize Problem” for the full ethics mess of data theft, career threats, and whether the proof even holds up. “I resigned from Anthropic today” is worth clicking for an insider’s argument that we’re past the point of control, even if you don’t buy the doomer framing. “LG TVs caught spying even when offline or on standby” is a concrete technical horror story with plenty of practical mitigation advice. “Tao: Open math problems being non-renewably mined by AI” offers the most thoughtful take on what AI competition does to collaborative science. And “We Must Return to the Office to Use AI in Person” is the rare satire that sparks both a serious RTO debate and a minor philosophy-of-AI flamewar over whether an LLM could have written it.
On the Navier–Stokes Millennium Prize Problem [comments]
1248 points · 990 comments · openai.com · 14h ago
The OpenAI announcement claims their AI agents solved the Navier–Stokes Millennium Prize problem, but the HN thread is almost entirely about the ethics of how they got there. The core allegation is that OpenAI used chat data from mathematicians Tristan Buckmaster and Levent Alpoge—who had been working on the same problem—to scoop their results, and then tried to pressure Buckmaster into accepting a coerced co-authorship deal while threatening his career. OpenAI’s defense, laid out by their researcher Sebastien Bubeck, is that they only heard rumors of a solution, didn't access private chats, and produced a completely different proof—but their admission that "we cannot rule out that de-identified data derived from their usage of our products helped improve our models" is poisoning the well. Commenters are deeply split: some see clear corporate bullying and data theft, others argue the mathematicians are just sore they got beaten to the finish line. A separate, sizable strand of the thread does napkin math on the $15+ million token cost and asks whether funding a handful of brilliant humans for a decade would have been a better bet than generating this expensive slop.
LibreOffice breaks download records after declaring it has no AI features [comments]
673 points · 221 comments · manualdousuario.net · 17h ago
The article touts LibreOffice 26.8’s record-breaking download week and attributes it partly to its explicit declaration that the software has no generative AI features, positioning the “not having AI” itself as a feature. HN immediately pushed back on the implied causality, with many pointing out that weekly download numbers have been trending up for a while and that hitting a new high is expected, not necessarily driven by the anti-AI messaging. A major wrinkle came from people noting that ChatGPT’s Codex desktop app bundles a full copy of LibreOffice when asked to produce Office files, which could inflate the download stats and undercut the “people are fleeing AI” narrative entirely. Others dug into the stats page to argue the one-million-first-week claim doesn’t match the official weekly chart, while a handful of commenters sidetracked into encryption quirks and remote-assistance tools for aging parents. The dominant takeaway was skepticism: the thread split between those treating it as a nice marketing win for openness and those insisting the record is mostly noise from a growing trend plus a bundling artifact.
AlphaGenome Atlas: a high-resolution map of human DNA [comments]
549 points · 118 comments · blog.google · 17h ago
Google DeepMind released the AlphaGenome Atlas, a massive database that uses AI to predict the effects of every possible single-letter change in human DNA, focusing on the 98% of the genome that doesn't code for proteins but still matters for disease. The thread quickly split into two camps: one side dug into the licensing terms, noting the non-commercial restriction and asking whether DeepMind plans to sell this data to pharma companies (with a few pointing to Isomorphic Labs as the commercial arm), while others called that greed and argued the resource has real diagnostic value. A separate cynical crowd dunked on Google Maps routing people into lakes as a cheap joke about the company's reliability, but the deeper, more exhausting debate was an ideological slugfest over whether DeepMind is genuinely good or just another shareholder-driven machine that will eventually monetize everything. A handful of domain-adjacent commenters pushed back on the hype, saying individual SNP predictions aren't that useful for drug discovery and that pharma would only license this out of FOMO, but nobody seriously challenged the scale or technical ambition of the dataset itself.
LG TVs caught spying even when offline or on standby [comments]
527 points · 306 comments · www.theverge.com · 15h ago
The Verge reports that LG smart TVs are logging nearby devices, recording audio, and capturing content from any input—even when offline or in standby—with data stored locally and uploaded once reconnected. The HN thread largely pivoted to practical defiance: nearly everyone agreed smart TVs should never touch the internet, with Apple TV and Kodi on a Linux box offered as the clean alternatives. A heated split emerged over whether disconnected TVs might still form mesh networks with neighbors' sets; one side called it an unfounded conspiracy, the other argued that technical possibility plus advertising incentives makes it prudent to assume the worst. Practical mitigations got concrete attention—blocking at the firewall level, isolating IoT on a separate VLAN, and even factory-resetting just for firmware updates—while a Samsung-specific gripe surfaced about TVs nagging to reconnect after being offline for a while.
I resigned from Anthropic today [comments]
490 points · 639 comments · x.com · 7h ago
The linked article is a resignation announcement from an Anthropic researcher who spent three years at both OpenAI and Anthropic, claiming both companies are recklessly racing toward self-improving superintelligence. HN immediately split into the usual camps: some argued he's just a doomer cashing in on attention before the IPO, pointing to the WSJ article about his resignation as marketing BS, while others took his concerns seriously and compared the risks to nuclear weapons—though several commenters noted that unlike nukes, superintelligence can be downloaded and run by anyone, making mutually assured destruction impossible. A significant contingent pushed back hard, arguing that current LLMs have crippling limitations like tiny context windows and practical uselessness for anything requiring real "zooming out," dismissing the whole thing as fear-mongering from labs trying to hype their IPOs. The thread also wandered into whether his resignation actually helps: skeptics said he's just making room for someone with fewer scruples, while defenders argued the viral attention might shift coworker perspectives or fuel regulation, though others pointed to past resignations from Geoffrey Hinton and Timnit Gebru that changed nothing beyond a few days of talk.
Muse – Meta’s personal AI agent [comments]
481 points · 516 comments · ai.meta.com · 12h ago
Meta’s Muse is a new “personal AI agent” meant to handle tasks like booking tickets and buying strollers, all while having access to your calendar, email, and browsing. The HN thread was overwhelmingly skeptical: most people flatly don’t trust Meta with that much personal data, pointing to years of predatory behavior and arguing that even if the product is polished, the company is the last one you’d hand the keys to your digital life. A few commenters pushed back, saying the average user doesn’t know or care about Meta’s reputation—3 billion monthly active users prove that people will use whatever gives them what they want, even if they hate the brand. Others who had tried Muse confirmed the worst fears: it aggressively steers toward spending money, prompting to book trips or make restaurant reservations unprompted, which led to jokes about AI’s primary purpose being making reservations. There was also a side debate about whether Meta was first to market with a hosted agentic tool, with people pointing to Grokbot, Gemini Spark, and various open-source clones as earlier contenders, though many dismissed those as less polished or too expensive.
I-have-ADHD: A skill to stop coding agents from burying the answer [comments]
418 points · 298 comments · github.com · 17h ago
The submission is a GitHub repo offering a "skill" that makes coding agents (like Claude) give ADHD-friendly, action-first, numbered-step responses instead of verbose fluff. The HN thread quickly spun off into people sharing their own minimalist prompts—"grug brain smol," "explain like I'm an executive on another meeting," or just adding a TL;DR request to their config—with many arguing that a simple "be concise" instruction works just as well or better. A stronger split opened up over whether it's appropriating a real disability: some commenters with diagnosed ADHD pushed back hard on the trend of casually claiming ADHD to get shorter outputs, while others countered that the coping strategies are universally useful and nobody's claiming a diagnosis. There was also a meta scuffle about whether this should be a "skill" at all versus just dropping the instructions into your AGENTS.md file, with defenders saying skills save tokens for when you actually need the verbosity.
DaVinci Resolve 21.1 [comments]
395 points · 175 comments · www.blackmagicdesign.com · 18h ago
The linked article wasn't available to this summarizer; from the discussion, Blackmagic Design added AI assistant integrations (Claude, Codex) into DaVinci Resolve 21.1, letting users give conversational commands like “sync these clips and generate highlights from the transcript.” HN split sharply: plenty of videographers and hobbyists cheered it as a godsend for tedious project management—syncing footage, batch rendering, organizing media—arguing this finally lets them focus on creative direction instead of fumbling with awkward tooling. But others pushed back hard, warning that AI slop music and auto-generated shows are already flooding platforms, that entry-level assistant editor jobs are disappearing, and that “agent integration” is just another API endpoint dressed up as productivity. A separate, bitter thread erupted over Blackmagic still withholding H.264/MP4 codec support from the free Linux version, with some pointing out that licensing patents—not malice—is the real blocker, while others mocked the company for prioritizing LLM toys over basic codec parity.
Among European Companies That Use a CDN, Nearly 9 in 10 Use Cloudflare [comments]
379 points · 363 comments · ciphercue.com · 23h ago
The article reports that among European companies using a CDN, nearly 90% are behind Cloudflare, with the remainder split among Amazon, Fastly, and Akamai—a concentration the author argues turns routine internal errors into market-wide outages. The discussion quickly pivoted to alternatives like Bunny.net, where opinions split: some praised its low cost and improving edge features, while others called it immature, noted its reliance on Discord for support, and pointed out it lacks Cloudflare Tunnels, which many consider too good to give up even on the free tier. A strong political thread argued this US dominance is dangerous for European sovereignty, with some suggesting the EU should deliberately provoke US retaliation to force a forced migration off American infrastructure. Others countered that the EU moves too slowly and lacks political will, while a separate meta-comment noted the article's own site was loading painfully slow from Europe—ironically proving the need for a CDN in the first place.
We Must Return to the Office to Use AI in Person [comments]
375 points · 65 comments · www.mcsweeneys.net · 18h ago
It's a McSweeney's satire about a company mandating six-day, 14-hour in-office work so employees can "use AI in person"—the protagonist is an "Associate Slop Doula" who clicks GENERATE and APPROVE buttons all day under surveillance software named Best Buddy, while the office leaks mayonnaise and is run by thousands of SVPs for fifteen workers. The thread immediately split into two camps: a deep debate about whether an LLM could ever write this piece, with people getting into the weeds of training a billion-parameter model on a single 50k-token corpus and whether the output would converge on verbatim reproduction or produce something novel—essentially a philosophy-of-AI argument about soul, determinism, and stochastic sampling. Meanwhile, a serious argument erupted over RTO itself: one commenter argued that remote work actually makes it *easier* for companies to replace knowledge workers with AI agents, because in-person interaction gives humans a moat that text-based work nullifies, while others countered that commuting is a huge cost and that collaboration benefits are overstated—and that the real problem is management mandating any one model instead of letting teams figure out what works. The satire landed hard, with multiple people noting its uncomfortable accuracy and Orwellian echoes, and the "Associate Slop Doula" title became a running joke.
Paramount Caught Using 'Astroturf' Group to Drum Up Fake Support for Merger [comments]
363 points · 119 comments · www.techdirt.com · 17h ago
Techdirt reports that Paramount was caught using a fake grassroots group called Neighbors for Strong Communities to pressure California’s AG into dropping an antitrust lawsuit over the company’s merger with Warner Bros. The thread spent a lot of time hashing out why the average person should even care about a merger like this, with one camp arguing that layoffs, price hikes, and content homogenization are direct consequences of consolidation, and another pushing back that mega-mergers can sometimes preserve a brand or studio that would otherwise collapse. A whole sub-thread went deep on the inevitable quality decay when big companies absorb small creative teams, using Bungie and id Software as recent cautionary tales from gaming. A few people pointed out that Larry Ellison’s threat to move Paramount out of California is pure bluster, since the actual threat to workers isn't losing a tax-break standoff—it's the mass layoffs that historically follow these deals. The conversation also veered into whether any merger can actually benefit consumers, with the Sprint/T-Mobile and Albertsons deals brought up as messy counterexamples where regulators made things worse by blocking or forcing store sales.
Tao: Open math problems being non-renewably mined by AI [comments]
338 points · 308 comments · mathstodon.xyz · 11h ago
Terence Tao posted about the risk that AI will "non-renewably mine" open math problems—essentially scraping through the remaining interesting questions without contributing new concepts, much like a mining operation that takes ore and leaves nothing to grow back. The thread immediately split over whether mathematics has ever been the collegial open-science utopia Tao seems to assume, with a lot of pushback citing Gauss hoarding results, the Newton-Leibniz feud, and the Pythagoreans literally killing people for sharing irrational numbers. A persistent counterargument was that the real threat isn't AI per se, but the dynamics of closed labs like OpenAI racing to finish a mathematician's approach before they can publish—several commenters pointed out that this already happened with a recent problem, forcing people to stop sharing progress if they want to keep credit. Others argued that the low-hanging fruit in math was already gone before AI showed up, and that the field needs better tools just to survive, so the real question is whether the community can reshape its incentives around collaboration rather than secretive AI-powered scooping.
DHS 'Predictive Policing' Unit Is Analyzing Americans' Financial Habits [comments]
338 points · 247 comments · www.404media.co · 17h ago
404 Media exposed a secretive DHS predictive policing unit, the Predictive Intelligence Targeting Team (PITT), that analyzes Americans' financial transactions and funnels intelligence to local cops, who then pull over drivers not suspected of any crime—just flagged as worth searching. The thread immediately invoked *Minority Report*, but the deeper argument split along familiar lines: one side said this is exactly what “big data” enthusiasts cheered for years ago, only now it’s coming for weed smokers and brown people, while the other side dug in on KYC and anti-money-laundering laws, insisting surveillance is a necessary evil to stop cartels and terrorists. A loud contingent pushed back hard, arguing that KYC is functionally useless against serious crime—citing studies that AML costs $250B a year while money laundering stays at 2–5% of GDP—and that these rules exist mainly to harass sex workers, small businesses, and political dissidents. The conversation also veered into the slippery slope of “terrorist” definitions, with one person pointing out that the Southern Poverty Law Center has advocated for mandatory financial abuse reporting, and another noting that courts and banks now treat you as guilty until proven innocent, as seen in recent account closures. The overall sentiment was grim resignation: cash and non-KYC crypto are the only defenses left, but the infrastructure for that is under constant legislative attack.
ChatGPT Images 2.5 [comments]
330 points · 404 comments · openai.com · 13h ago
The linked article wasn't available to this summarizer; from the discussion, OpenAI announced ChatGPT Images 2.5, an update to its image generation model. The thread immediately zeroed in on whether it fixed the persistent "noise gradient" issue and the yellow tint — answer appears to be no, with the showcase images still showing mangled fingers, weird shadows, and the same three-finger problems that made previous versions a punching bag. A huge chunk of the thread spun out into a bitter argument about "AI menu slop": small businesses are already using these models for kebab shop signs and DoorDash photos, and the split is sharp — some see it as a harmless productivity win for time-strapped owners, while others insist the uncanny-valley look actively repels customers and makes them distrust the food. A few commenters went full philosophical, arguing that AI fakery just returns the world to an older tradition of tall tales and creative exaggeration, but the counterpoint hammered home that images still carry an implicit claim of reality, and that gap is only widening as models improve but artifacts persist.
How to build a printer [comments]
300 points · 57 comments · nishantjosh.dev · 10h ago
The article details how a developer repurposed an Xteink e-ink reader into a network printer by implementing the Internet Printing Protocol (IPP) on an ESP32, letting macOS see it as a real printer and send documents wirelessly. HN loved the hack’s cleverness—especially the insight that if an e-reader looks like paper, it should act like paper—but immediately split on _why_ you’d do this. One camp saw it as brilliant for quick one-offs like recipes or reference docs, bypassing the usual phone-in-the-kitchen hassle; the other camp kept pointing out that the device already natively renders PDFs, so why not just load files directly? The thread’s real heat came from printing experts: someone deep in the CUPS/IPP ecosystem jumped in with precise technical corrections—suggesting 1-bit-per-pixel modes, exact screen-dimension media declarations, and the right `urf-supported` flags—while another veteran of the printing stack warned that bi-level support is optional on macOS and that monochrome remains the only reliably cross-platform raster format, sparking a brief pushback on whether the author’s workaround was actually necessary. A side tangent landed on the lament that modern macOS apps (VSCode, Obsidian) can’t even print anymore, making the whole hack feel antique in a different way.
Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs [comments]
255 points · 131 comments · github.com · 11h ago
The submission is a developer’s report on running the full, unquantized Kimi K3 model (2.78 trillion parameters, 1.45 TB of expert weights) on an M5 Max MacBook Pro by streaming experts from four SSDs, achieving a steady 1 token/s decode. The HN thread split cleanly between people questioning the practical utility of a model that takes 6 minutes to begin answering a 512-token prompt and those defending the project as a legitimate engineering exploration—pointing out that 1 token/s is still faster than a human for many batch jobs and that the real value is in learning how to push consumer hardware to its limits. The author’s own long, instrumentation-dense top comment drew complaints about being an AI-generated wall of text, but also provided a fascinating catalog of which micro-optimizations actually worked (e.g., separating demand and prefetch thread pools gave +14%) and which didn’t (e.g., RAM expert caches made things worse). Several commenters connected the effort to the historical hacker ethos of doing something just to see if it can be done, with one comparing it to punching cards and waiting days for a syntax error, while others dug into the technical details of whether the same approach could work on discrete GPUs or faster NVMe drives.
Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses [comments]
245 points · 118 comments · quesma.com · 17h ago
The article benchmarks Qwen3.8 27B quantizations, showing the 4-bit Q4_K_M matches the full BF16 model on agentic coding benchmarks while fitting on a 24 GB card, but 1-bit quantizations collapse to random guessing. The thread immediately zeroed in on the missing Q3 quantization, which is the critical breakpoint for sub-16 GB cards like the 5080, 5070 Ti, and AMD 9070 XT—users shared real-world setups running Q3 quants on those cards, trading off context window and speed, and pointed to newer dynamic 3-bit GGUFs that might outperform the standard ones. Meanwhile, a separate statistical flank erupted over the author’s use of Wilson confidence intervals to claim run-to-run stability, with several people arguing that confidence intervals don’t measure run-to-run variation and that prediction intervals or Bayesian credible intervals would be more appropriate, though the author defended the choice as “very conservative.” The split was clear: practitioners wanted practical guidance on the Q3 sweet spot for consumer hardware, while the math-inclined contingent pushed back on the statistical methodology.
The two Christian saints who are the Buddha [comments]
237 points · 176 comments · signoregalilei.com · 17h ago
The article traces how the story of the Buddha was absorbed into Christianity as the saints Barlaam and Josaphat, complete with a prince isolated in a palace, encounters with sickness and death, and a conversion to a hermit’s faith. HN threads mostly pulled in two directions: one camp dove into the broader claim that early Christian monasticism and even gospel structures were directly influenced by Buddhist traditions, citing parallels between Pali sutta dialogues and synoptic gospel pericopes, while the other camp pushed back hard, arguing the similarities are superficial and that Christianity’s core metaphysics—a personal God, real change, diachronic identity—are fundamentally incompatible with Buddhist thought. A smaller but determined group focused on the institutional history, noting the Catholic Church removed these saints from its martyrology in the 20th century while Orthodox churches still recognize them, and that there are separate, unrelated saints with the same names causing confusion. A minor running joke punctured the article’s seriousness over the website’s Italian grammar—“Signore Galilei” should be “Signor Galilei,” meaning the name reads like “Mister! Galilei!” rather than “Mr. Galilei.”
Antiquated HTML Snippets and Artefacts [comments]
233 points · 83 comments · vale.rocks · 22h ago
The article catalogs forgotten HTML meta tags, conditional comments, and browser-specific hacks from the Internet Explorer era and earlier. The HN crowd got deeply nostalgic, swapping war stories about IE6 PNG alpha filters, the spacer.gif, and table-based layouts with nine cells for rounded corners—but they also dove into serious technical pushback on the viewport meta tag. A fierce split emerged over whether `initial-scale=1` alone suffices or if `width=device-width` is still necessary, with the author himself chiming in after a recent MDN edit he made sparked a GitHub issue about defaults. Meanwhile, the article’s editorial aside about ditching Twitter-specific tags because of Elon Musk’s Nazi salute ignited a political flamewar, though others pointed out that `twitter:card` has no Open Graph equivalent and still works.
Getting your hands dirty is good for you [comments]
229 points · 190 comments · www.bbc.com · 22h ago
The BBC article argues that touching soil, grass, and plants boosts skin microbial diversity and immune function, citing Finnish daycare studies where kids playing on forest floors showed fewer pathogens and better immune markers. HN mostly ran with the premise as a satire of lifehacking culture—immediately jumping to jokes about a “smart dirt bucket” with daily mud-soaking protocols, an app to track your hand microbiome, and influencers selling their own branded mud. A secondary thread spun off into debating sauna timing optimization and whether timing your cold plunge defeats the purpose, which neatly mirrored the same impulse to over-engineer a simple pleasure. Some pushback came on the sun-exposure tangent, with a split over whether skin cancer risk is worth the vitamin D and “tan feels amazing” benefits, while others pointed out rattlesnake analogies. Overall, the commenters treated the article less as a health revelation and more as a prompt to roast the tendency to quantify every natural activity.
Show HN: Copperhead – Cursor for circuit boards [comments]
227 points · 90 comments · copperhead.sh · 18h ago
The submission is Copperhead, an open-source AI agent that takes a written product brief and generates a complete KiCad PCB design—schematic, layout, BoM, firmware pins, and test plan—through a staged pipeline with verification gates at each step. HN immediately split into camps: a skeptical group argued that existing LLMs already work fine for editing KiCad schematics, calling Copperhead a “waterfall” wrapper that PMs would love but engineers wouldn't, while the creator countered that it’s not just a wrapper—it uses its own hardware IR and deterministic verification engines beyond what raw Claude or GPT can do. A deeper thread erupted over whether LLMs can ever handle real PCB layout, with several commenters asserting that autorouting is an NP-hard problem that no amount of “magic LLM pixie sprinkles” can solve for complex boards, though others pushed back that humans also rely on heuristics and “good enough” solutions dominate the industry. Meanwhile, the hardware engineering community’s gatekeeping was called out, with some noting that Reddit forums reflexively dismiss AI tooling as a “skill issue,” while other commenters shared success stories using Claude to generate KiCad files via a Python script that actually produced working boards.
Trey Parker and Matt Stone Are Changing the Name of South Park to South America [comments]
209 points · 108 comments · x.com · 11h ago
The article is a satirical tweet from South Park’s account announcing they’re renaming the show to “South America” to mock Apple, Google, and Paramount for their recent corporate capitulations. HN immediately seized on this as a chance to re-litigate the show’s legacy, with a loud faction arguing that South Park’s “both sides are equally bad” shtick—epitomized by episodes like Giant Douche vs. Turd Sandwich and ManBearPig—did the cultural groundwork for the current populist right, making cynicism the default posture. Others countered that the show has evolved, pointing to the Trump-era Mr. Garrison storyline where Trump is literally Satan’s boyfriend, and that recent seasons have clearly landed on the side of “MAGA is wrong and evil.” The debate then spiraled into a broader argument about whether the left or right is more pro-censorship, drawing in Tipper Gore’s PMRC and COVID lockdowns, but the core tension was whether South Park’s creators bear any responsibility for normalizing the very political nihilism they now claim to satirize.
Mercury 2.5 [comments]
187 points · 26 comments · www.inceptionlabs.ai · 11h ago
Inception Labs released Mercury 2.5, a diffusion-based language model that claims 1,100 tokens per second and context up to 260K, pitched at latency-sensitive workloads like voice agents and coding subagents. The HN crowd is split between impressed and skeptical — several people who actually tested 2.0 or the 2.5 preview confirm the speed is genuinely fast and useful for sub-tasks or quick routing, but nearly everyone agrees it's nowhere near frontier models for general reasoning or tool use, and one person who benchmarked it found their custom agent harness worked worse than expected. There's a strong current of "this is a diffusion model, not the usual next-token predictor, and watching it generate the whole page at once is genuinely mind-blowing," which a few commenters thank others for pointing out, and the token speed is so high that one person joked it's "too fast." Several people note the headline benchmarks compare it to older flash-tier models from other labs and only compare intelligence gains against Mercury 2, calling that misleading and saying the model is probably useless for hard problems, though others push back that even a dumb model at that speed is valuable for subagent orchestration. A few tangents hit the name collision — people thought it was about Mercury outboard motors or the Mercury programming language — and one person got kicked by an overeager IP-protection classifier when trying to probe the model's architecture.
The Navier–Stokes Millennium Prize Problem [comments]
179 points · 143 comments · simonwillison.net · 2h ago
Simon Willison’s post details a mess where OpenAI claims its unreleased model solved the Navier–Stokes Millennium Prize problem in 88 hours, but the real story is that they apparently scooped a team from Anthropic and NYU that had been working on it for almost a year using OpenAI’s own tools. HN immediately zeroed in on the authorship play — OpenAI offered to let the NYU professor lead a paper on their result but explicitly excluded his Anthropic collaborator, which the thread treats as a massive academic red flag and clear scientific misconduct. A big chunk of the discussion pivots to the privacy implication: people are furious that even if OpenAI didn’t directly train on the victims’ code sessions, the mere existence of a rumor about a solution was enough to send their models hunting, and one commenter frames this as exactly analogous to how security researchers can now find exploits just by knowing a bug exists. Others push back on the plausibility of the rumour-mill logistics — they question how a confidential breakthrough turned into a tip that reached OpenAI at all, with some cynically noting that mathematicians talk, and that the “red flag” for the victims was hearing OpenAI used a similar technical approach that very few people in the world were pursuing.
OpenAI fought dirty on career-making math problem [comments]
162 points · 60 comments · techcrunch.com · 12h ago
NYU mathematician Tristan Buckmaster says OpenAI built on his and an Anthropic colleague's unpublished work to claim the Navier–Stokes millennium prize first, burning $22.5 million in compute to get there. The HN thread split cleanly: many called it an obvious scoop, pointing out that OpenAI admitted starting after hearing rumors of a breakthrough and that the specific research direction was almost unique to Buckmaster's team, making coincidence unlikely. Others argued that if the proofs were genuinely different, OpenAI just got there faster and that this is standard competitive lab science — except mathematicians aren't used to being outgunned by a company that can spend tens of millions on a rumor. A recurring undercurrent: Buckmaster used OpenAI's own Codex extensively, raising the possibility that his own work fed the model that then regurgitated the approach, which OpenAI downplayed but couldn't fully rule out. The whole thing left a sour taste about OpenAI's willingness to leverage insider knowledge and brute-force compute to dominate academic credit, even as the underlying AI achievement is real.
Show HN: LLM Attention Visualization [comments]
155 points · 24 comments · ishamf.dev · 15h ago
This is a browser-based tool that visualizes which previous tokens an LLM “attends to” when generating each new token, using a simplified metric that sums attention weights across all heads and layers. The HN crowd immediately pushed back on whether that metric actually measures influence—a few comments pointed to the transformer circuits literature for a more rigorous approach, though the author acknowledged the simplification upfront and argued that even naive aggregation reveals interesting patterns like verbatim copying or semantic blending between phrases. There was also a substantive thread about computational complexity, clarifying that while each forward pass is linear in context length with a KV cache, generating a full sequence still scales quadratically, which matters for deployers worried about long-context costs. A UX complaint noted that the pause button resets the animation instead of freezing it, with a request for step-by-step controls, while several educators said the tool would be perfect for teaching attention mechanisms.
Large language models develop novel social biases through adaptive exploration [comments]
150 points · 78 comments · openreview.net · 10h ago
The linked article wasn't available to this summarizer; from the discussion, this paper claims that large language models can spontaneously develop new social biases against made-up demographic groups, even when there are no real-world differences to learn from. The HN thread immediately split into two camps: one side argued this is exactly what you'd expect from a system doing pattern-matching on limited data, with some pointing out that the models are just picking up on random statistical correlations in the context window. The other side pushed back hard, saying the paper's definition of bias—"behavior that tilts away from equality"—is ideological nonsense, since a truly unbiased system should track ground truth, not impose equal outcomes. Several commenters also criticized the experimental setup itself, noting that the prompts literally ask the model to consider clan membership when making hiring or conscription decisions, which feels less like measuring bias and more like engineering it.
Copyright does more harm than good and should be abolished [comments]
142 points · 95 comments · grapheneos.social · 1h ago
The linked article wasn't available to this summarizer; from the discussion, it’s a GrapheneOS post arguing that copyright does more harm than good and should be abolished. The HN thread immediately split into two camps: one side says this is big-tech propaganda to dodge licensing fees, that copyright is the only thing keeping GPL alive and protecting small creators, and that the real fixes are shorter terms and non-transferable rights. The other side counters that copyright in practice is a weapon for corporations like Disney and Oracle, that asymmetric power is baked in (individuals never see royalties while giants litigate endlessly), and that abolition—or at least drastic shortening—would free remixes, Sci-Hub, and software from patent mazes. A strong undercurrent points to DMCA abuse as the real problem, not copyright itself, and several commenters note that the asymmetry argument applies to most rights but doesn’t justify trashing them entirely. The debate never converges; it’s a raw split between “fix the asymmetry, don’t nuke the right” and “the right is already broken for everyone but the powerful.”
FreeBSD 14.5-Release [comments]
126 points · 25 comments · www.freebsd.org · 19h ago
FreeBSD 14.5-RELEASE is the sixth and mostly maintenance update to the stable/14 branch, shipping bug fixes and driver updates but almost no new features. The HN discussion immediately tripped over versioning confusion—since FreeBSD 15.1 is already out, people had to sort out that 14.x is the legacy branch still supporting i386 (the last to do so) and won’t go EOL until November 2028. A debate broke out about what to run in production: the traditional wisdom says RELEASE, but several people pointed out that Netflix runs -CURRENT on thousands of CDN servers and pfSense CE ships on 16.0-CURRENT, while others argued that’s reckless and that -STABLE has historically been the right choice (though the handbook now defines -STABLE as a development branch). Someone with decades of FreeBSD experience admitted the policy changed years ago and they’d been unknowingly running unreleased code on production machines, which drew a mix of sympathy and gentle correction.
PISA 2025 Students' reading and mathematics performance declined across the OECD [comments]
123 points · 146 comments · www.oecd.org · 20h ago
The OECD’s latest PISA results show a sharp decline in reading and math performance across developed countries since 2018, with the trend actually starting in the early 2010s. The Hacker News thread largely split into two camps: one arguing the drop is obviously linked to social media and smartphones (citing Jon Haidt’s research), and another insisting it’s about students rationally disinvesting from education because they believe AI will make traditional skills obsolete. That second angle got heavy pushback, with people pointing out that while AI can churn out plausible-looking code and text, it’s still terrible at anything requiring deep understanding—and that the “kids are just giving up because of AI” take ignores the fact that the decline predates modern LLMs by nearly a decade. A separate thread dug into an anecdote about Chalmers University’s entrance exam pass rate falling from 80% to 20% since the 1960s, which sparked debate about whether the test itself is a fair benchmark given how much the applicant pool and curriculum have changed.
Generated 2026-09-09 08:10 UTC
Generated by Sauron from Hacker News discussions and linked articles.