HN Brief: 2026-06-13
Today’s HN was dominated by the fallout from the US government ordering Anthropic to cut off non-citizens from its latest models—and the crowd was in no mood to sympathize, given Anthropic’s years of hyping danger to push for regulation. A second throughline was the ongoing tug-of-war between open- and closed-source AI, with a manifesto for open models drawing fierce pushback that hardware moats are the real bottleneck, while a new trillion-parameter coding model sparked another round of “benchmarks don’t match reality” debates. A handful of threads also wrestled with the concrete costs of AI slop—maintainers drowning in LLM-generated PRs, and an autonomous agent finding 21 zero-days in FFmpeg, reopening old wounds about unsustainable security triage for essential infrastructure.
Click into “Statement on US government directive to suspend access to Fable 5 and Mythos 5” for the sharp community rebuttal of Anthropic’s regulatory capture dance; “CRISPR tech selectively shreds cancer cells” for a rare genuinely hopeful biotech advance tempered by delivery reality; “Twenty One Zero-Days in FFmpeg” to watch the community grapple with open-source security economics that break as fast as bugs get found; “I Am Not a Reverse Centaur” for the increasingly bitter divide between maintainers and AI-driven contributors, with no reconciliation in sight; and “Ryanair dark UX patterns summer 2026 refresher” for the classic low-fare trade-off calculus that never gets old, plus real horror stories that undercut the “just click carefully” defense.
Statement on US government directive to suspend access to Fable 5 and Mythos 5 [comments]
2035 points · 1495 comments · www.anthropic.com · 7h ago
Anthropic published a statement saying the US government ordered them to cut off all non-US citizens from their latest Fable 5 and Mythos 5 models, citing a narrow jailbreak Anthropic says is trivial and available in other models. The thread mostly turned into a sharp rebuttal of Anthropic’s own messaging — people pointed out that the company has spent years hyping up how dangerous and powerful their models are to push for regulation, and now that the government actually took them at their word, they’re crying foul. A lot of folks argued this is actually a marketing gift, turning the ban into “so advanced the government had to stop it,” and some even suspected the company orchestrated the whole thing for regulatory capture. The legal definition of “foreign national” got a deep dive, and a surprising number of people confirmed they got full refunds on their subscriptions already. The consensus in the comments is less sympathy for Anthropic and more “you asked for this, now lie in it.”
CRISPR tech selectively shreds cancer cells, including "undruggable" cancers [comments]
797 points · 189 comments · innovativegenomics.org · 16h ago
The article describes a new CRISPR-Cas12a2 system that detects a specific mutant RNA from the p53 tumor suppressor—found in nearly half of cancers—and shreds the cell’s chromatin, killing only cancerous cells while leaving healthy ones untouched. HN is cautiously optimistic: many recognize this as a genuine advance in the tooling, but the dominant note is that delivery remains a massive hurdle, and even if it works in vitro, getting the enzyme to every cancer cell in a patient is years or decades away. A long tangential thread broke out over whether society prioritizes adtech over cancer research, with people arguing the dollars actually favor biomedicine and that the real bottleneck is regulatory speed and long ROI timelines. There was also some heated pushback against comparing this to Theranos, with several commenters insisting that real genetic engineering is a maturing class of therapies, not a single wonder drug, and the timeline is measured in decades—so don't expect floodgates to open overnight.
Open source AI must win [comments]
791 points · 243 comments · opensourceaimustwin.com · 5h ago
The submission argued that AI must remain open source to prevent cognitive functions from becoming a rent-seeking subscription model controlled by a handful of companies. The thread mostly pushed back on that premise, drawing a Photoshop-versus-GIMP analogy: frontier labs will always produce better models because open-weight projects can’t command the capital for cutting-edge training, though many conceded that local models like Qwen are already competitive for most practical tasks and will only improve as hardware catches up. A vocal minority countered that the real moat isn’t models but hardware—trillion-dollar firms are cornering DRAM and fab capacity, squeezing consumer hardware out of existence and making even open-weight models unusable without a datacenter. Others dismissed that as too pessimistic, pointing to past deflationary tech curves and the fact that you can already run a 72B-parameter model on a Mac Studio, so the fight isn’t over yet.
Kimi K2.7-Code: open-source coding model with better token efficiency [comments]
424 points · 225 comments · platform.moonshot.ai · 21h ago
Kimi K2.7-Code is a 1-trillion-parameter MoE coding model claiming 30% fewer thinking tokens than its predecessor, and it posted competitive benchmark scores against GPT-5.5 and Claude Opus 4.8. The thread quickly turned into a debate about whether those benchmarks reflect real-world use, with many arguing that Claude is still noticeably better in practice — one person said they keep coming back to Claude to fix the outputs of Chinese models, while another pointed out that on less-gamed benchmarks like DeepSWE, the gap widens considerably. A strong consensus emerged that cost-per-token is misleading: the real metric is total tokens and human attention spent wrestling with the model, and the marginal quality difference can cost far more in debugging hours. Some also questioned whether Chinese labs can be trusted with US enterprise data, though several noted that many US providers host these open-weight models anyway. The whole conversation pivoted from the model release to a broader assessment of whether Anthropic's current lead is a real moat or just inertia, with frequent comparisons to iPod-era lock-in and plenty of personal war stories about switching and switching back.
"Don't You Just Upload It to ChatGPT?" [comments]
404 points · 325 comments · correresmidestino.com · 14h ago
A freelance translator pushes back after a gym acquaintance casually asks why she doesn't just upload documents to ChatGPT, arguing that AI is a tool, not a replacement, and that she still has to triple-check everything. The thread immediately split into two camps — many agreed with the author but insisted her stance is temporary, predicting that AI will surpass human translators and software engineers within 5-10 years, with some developers already preferring AI output over human coworkers and ditching manual coding entirely. Others pushed back hard, calling that outlook "misanthropic doom" and arguing that LLMs are fundamentally unreliable, hallucinate facts, and have shown zero improvement in trustworthiness despite years of hype. A separate debate emerged over whether the shrinking market for human translation is actually unfortunate: some argued that AI has raised the quality floor and pointed to human translators botching jobs (including injecting personal politics into game translations), while others countered that low-quality machine translations are now the norm and cited examples like Amazon's submit button reading "Walking" in another language. The conversation largely bypassed the article’s central point about localized, nuanced translation craft in favor of big-picture predictions about whether humans will become obsolete.
Electric motors with no rare earths [comments]
397 points · 100 comments · www.renaultgroup.com · 9h ago
Renault’s post explains their electrically excited synchronous motor (EESM), which replaces rare-earth permanent magnets with an electromagnet to dodge China’s stranglehold on supply. The thread quickly dug into the trade-offs: the rotor is less mechanically robust and heavier, so power density suffers compared to permanent-magnet motors, but the efficiency hit is small because the excitation power is a tiny fraction of total draw. Several people pointed out that BMW already uses a similar magnet-free design with almost twice the power and 800V architecture, though at a much higher price point—prompting the usual debate about whether BMW’s reliability offsets the performance edge. A deeper tangent erupted over supply-chain geopolitics, with one detailed comment arguing that EESMs are driving EU-India free trade agreements and that Indian joint ventures are now integral to European motor production, while others dismissed that as propaganda and said European re-industrialization is already happening without India. Meanwhile, a few commenters corrected misconceptions about AC induction motors and noted that Nissan and other automakers also use EESMs, though availability varies by region.
How to setup a local coding agent on macOS [comments]
353 points · 84 comments · ikyle.me · 14h ago
The linked article walks through setting up a local coding agent on an M1 Max Mac with llama.cpp, Gemma 4, and speculative decoding, boasting 72 tokens per second. The thread quickly branched into practical alternatives—several people pointed out you can skip the manual model downloads by just passing `-hf` to llama.cpp, and others recommended tools like ollama with opencode or oMLX for a simpler setup. A real split emerged around whether local models are worth the hardware cost and effort: some argued that even on an M5 Max with 128GB, local models are "toys" compared to cloud APIs, while others countered that for privacy, reliability, and learning, the trade-off is fine and that today's open models are shockingly good compared to last year's frontiers. Several experienced users pushed back on the article's benchmark methodology—generating only 128 tokens is "benchmarking the overture, not the opera," since MTP speedup depends on longer outputs and system prompts, and a few folks reported that MTP actually broke markup in their agent setups. The conversation also got into model personality differences (Gemma writes like a concise code blogger, Qwen is terse like Chinese open-source docs) and the importance of being able to swap both models and agent harnesses.
A Call to Action: Stop the FCC's KYC Regime [comments]
319 points · 216 comments · blog.lopp.net · 17h ago
Jameson Lopp's post attacks an FCC proposal that would force phone providers to collect government ID and other personal info before activating service, arguing it won't stop determined criminals but will hurt privacy for journalists, survivors, and protesters. The thread immediately split on whether KYC actually reduces spam: some pointed to Europe where it supposedly works (citing the UK and Netherlands), but others shot back with Italy's long-standing KYC requirement that hasn't stopped rampant scam calls—an argument that went unresolved because spam often originates outside the country. The real conversation shifted to why STIR/SHAKEN caller-ID authentication hasn't been enforced properly, with one deep dive explaining that spammers route calls through legacy TDM systems that strip the headers, and a proposal that carriers should simply refuse to accept calls from those systems. Another major vein was the tension between wanting anonymous communication and wanting traceability for harassment—several people landed on the idea that paying with a traceable credit card should be enough for law enforcement, without handing full PII to telcos that have terrible breach records. The consensus: KYC looks like a blunt, data-hungry fix that won't touch the spam root cause, while the real leverage point—shutting down spoofed and unauthenticated calls—remains politically untouched.
Palantir loses legal challenge against Swiss investigative magazine [comments]
308 points · 60 comments · www.ft.com · 11h ago
A Swiss commercial court threw out almost all of Palantir’s legal attempt to force investigative magazine Republik to publish rebuttals to articles about Switzerland repeatedly rejecting the company’s services. The thread quickly turned into a long digression on the *Lord of the Rings* name: multiple people pointed out that in the books the palantíri provide technically accurate intelligence that leads to disastrous strategic decisions, making it a perfect metaphor for a surveillance contractor whose CEO recently called Trump’s win a “landslide.” Others debated whether Europe will actually follow through on decoupling from US tech—skeptics said Europe talks big but fails to deliver, while a few noted Denmark and the Netherlands are already expressing unease. There was also the usual meta fight over using archive.ph to bypass the FT paywall, with accusations that the site alters content and uses visitors as a botnet, and a counter-argument that paywall dark patterns created the problem. The strongest consensus was that Palantir’s name is astonishingly on-the-nose, and not in a good way.
I Am Not a Reverse Centaur [comments]
262 points · 199 comments · blog.miguelgrinberg.com · 14h ago
Miguel Grinberg, a Flask maintainer, announced he’s done entertaining LLM-generated pull requests that amount to slop—he now requires pre-discussion in an issue and proof of human involvement before any PR will even be looked at. The thread immediately split: many maintainers cheered, saying the policy is overdue and that open source is drowning in low-effort AI junk, while others pushed back hard, arguing that the “must talk first” rule was already a barrier pre-LLM and often used to reject perfectly good contributions that just didn’t fit a maintainer’s personal vision. A big undercurrent was whether open source still matters when LLMs can generate code on the fly—some said quality libraries still win because they’re reliable, but others worried about discoverability, attribution, and the erosion of community norms. A lively subthread unfolded around the democratization angle: non-programmers are thrilled to finally build things, but plenty of commenters dismissed that pride as fake, drawing analogies to iPhone camera films and sparking accusations of elitism. The real split isn’t about tooling—it’s about whether the goal is collaborative craftsmanship or just getting something that works, and nobody seems to have a good way to reconcile the two.
Pirates, a naval warfare game inspired by Sid Meier's Pirates [comments]
249 points · 78 comments · piwodlaiwo.github.io · 14h ago
The linked article wasn't available to this summarizer; from the discussion, it's a browser-based naval combat game inspired by Sid Meier's Pirates. Most of the thread is players tearing into the game's AI and balance — the consensus is that it's trivially easy to win by simply circling or strafing the enemy, regardless of ship size, and the cannonballs are nearly invisible as tiny black dots on a dark background. Several people called for adding wind dynamics, realistic sailing physics, and ship-specific differences like gun caliber and hull strength to prevent a sloop from dominating a ship of the line. The developer popped up frequently, taking notes and promising to add wind, opponent selection, and multiplayer. A strong side conversation emerged about the original Sid Meier's Pirates (including its removal from the iOS App Store) and a handful of commenters used the thread to plug their own naval or retro-inspired games.
Ryanair dark UX patterns summer 2026 refresher [comments]
246 points · 188 comments · blog.osull.com · 20h ago
The article is a walkthrough of the nine-step minefield Ryanair makes you navigate during check-in to avoid paying for insurance, seat selection, priority boarding, and a dozen other upsells. The HN thread quickly split into a familiar fight: several commenters argued the low base fare—£50 from Bournemouth to Alicante—makes the dark patterns a worthwhile trade-off, with one doing napkin math that your time spent clicking through the traps would need to be worth $1,500/hour to make a more expensive airline rational. Others pushed back hard, pointing out that US budget airlines like Frontier offer similar prices without quite the same gauntlet, and that the math on ancillary revenue versus fare difference doesn't hold up. A deeper thread dug into Ryanair’s actual business model—fast turnarounds, exclusive airport deals, and single-plane fleets—suggesting the dark patterns are less about nickel-and-diming and more about shaping passenger behavior to enable those efficiencies. And of course, several commenters shared horror stories of being misled by the Ryanair app to the wrong gate, missing flights, and being told it was their fault, undercutting the “just click through carefully” defense with evidence that the patterns can cause real, expensive failures.
WASI 0.3 [comments]
243 points · 90 comments · bytecodealliance.org · 18h ago
The WASI 0.3 release makes async native to WebAssembly Components by putting the host in charge of a shared event loop, eliminating the clunky three-step dance from 0.2 and simplifying the ABI. A big chunk of the thread immediately pivoted to whether WebAssembly itself has ever truly broken through—lots of people pointed to invisible success stories like Figma, Cloudflare Workers, and Flight Simulator plugins, while others griped that WASM still can’t beat JavaScript on raw number-crunching benchmarks and that debugging non-trivial C/C++→WASM code is a nightmare. There was a sharp split over the process: one frustrated developer said the component model felt like “insider wizardry” developed in the shadows, but a maintainer countered with a barrage of public calendars, Zulip chats, and conference talks, asking what else they could do to make the work visible. Another commenter worried about scope creep, noting that the project’s name change from “Systems Interface” to “Standard Interfaces” signals an expanding scope that might hamper real-world adoption before a 1.0 is even reached.
The Future of Email [comments]
192 points · 203 comments · www.fastmail.com · 21h ago
Fastmail's blog post argues that as AI assistants start reading and acting on our email, robust authentication standards like SPF, DKIM, and DMARC become the critical infrastructure keeping the inbox trustworthy—moving beyond mere delivery verification to proving who actually sent a message. The HN thread mostly ignored the authentication details and instead took a sharp detour into a long-running grievance: why can't banks, healthcare providers, and government agencies just use regular email instead of forcing users into walled-garden "secure message centers"? A lot of people vented that those portals break their personal archives and are a pain to log into, though others pushed back that institutions need features like message recall and audit trails that plain email can't provide. A second major strand debated the idea of making email invite-only, where you pre-authorize senders like a social network, with some arguing that just shifts spam to the friend-request queue and others pointing out that masked email services (including Fastmail's own) already offer a practical alternative. There was also a brief, skeptical aside about whether this whole push is just another Big Three power grab, with Google and Yahoo already mandating DMARC for bulk senders.
Swift at Apple: Migrating the TrueType hinting interpreter [comments]
191 points · 84 comments · www.swift.org · 12h ago
Apple rewrote its TrueType hinting interpreter from C to Swift, gaining memory safety and a 13% performance boost while preserving pixel-perfect rendering. The thread quickly turned into a broader discussion of Apple’s “Rewrite in Swift” push across the OS, drawing comparisons to Microsoft’s similar font‑engine rewrite in Rust—and people debated whether Swift’s borrow checker and ABI mattered more than Rust’s stricter ownership model. Several folks dug into the MIT licensing choice (unusual for Apple), and there was some unease about the project’s heavy use of LLM‑assisted code, though the author insisted every line was human‑reviewed and exhaustively tested. A side argument erupted over macOS’s decision to render UI text unhinted on low‑DPI displays, with some calling it a deliberate push toward Retina and others pointing out that billions still use 1080p screens.
Slightly reducing the sloppiness of AI generated front end [comments]
189 points · 118 comments · envs.net · 17h ago
The author tried prompting an AI to generate front-end code in a “Qt style” and found that it dramatically reduced the “AI slop” look compared to other style prompts. Most of the thread pivoted hard into blaming the problem on “web slop” instead of AI itself — the abandonment of consistent, bordered widgets from desktop paradigms like Xerox PARC or Windows 98, with several people linking a Win98-style version that many called refreshingly usable and un-sloppy. Others pointed out that Qt is so heavily represented in the training data that “Qt app” is basically a named distribution in the model’s latent space, which explains why it coheres. A vocal minority disputed the whole premise, arguing that the examples all look ugly to them and that “AI slop” is just a subjective label for lazy design, while a separate camp shared workflows where they get great UIs by iterating with screenshots and style guidelines. The thread never really resolved whether slop is a real property or just a vibe, but everyone agreed that modern web UI conventions are a mess.
Twenty One Zero-Days in FFmpeg [comments]
184 points · 112 comments · depthfirst.com · 9h ago
An autonomous security agent trawled through FFmpeg's codebase and found 21 zero-day vulnerabilities, including one in the AV1 RTP depacketizer that lets a single 183-byte packet give an attacker control of the instruction pointer whenever FFmpeg is pointed at an untrusted RTSP stream. The thread immediately split between people who saw this as an LLM demo shot across the bow and those who shrugged — FFmpeg has been a fuzzer piñata for years, and its maintainers have a well-known hostility toward bug reports and a habit of making April Fools jokes about security. The deeper argument was about sandboxing: most people running FFmpeg in production do it without any isolation, and even browser sandboxes are only one chain away from compromise, while others countered that these exploits still need an ASLR leak and are typically combined with other bugs. A recurring bitter note was that FFmpeg is irreplaceable ("the only game in town") and that security reporters flood volunteer maintainers with LLM-generated slop, so the real problem isn't the zero-days but the unsustainable economics of open-source vulnerability triage.
There is a shadow hanging over this Fable thing [comments]
171 points · 133 comments · 12gramsofcarbon.com · 2h ago
The US government directed Anthropic to cut off access to its strongest models (Fable 5 and Mythos 5) for anyone who isn't a US citizen, leading Anthropic to just kill access entirely — and the article's author, an admitted AI doomer, worries this is less about safety and more about score-settling from an administration with ties to OpenAI's investors and a history of antagonizing Anthropic. HN latched onto the bigger precedent: if the US starts restricting the most capable LLMs to citizens only, China will almost certainly reciprocate, meaning within a couple of years the frontier models could be completely unavailable to the public, which would crater the datacenter buildout bet. A major split emerged between those who see this as a genuine regulatory Rubicon and those who think it's just Anthropic getting punished for earlier marketing stunts — with the latter camp noting Anthropic's own admission that GPT-5.5 can match Mythos's capabilities, undercutting their previous hype. A few commenters also dragged the discussion into a side tangent about whether generative AI is really transforming game design, arguing that the flood of AI-generated slop will make discovery impossible rather than producing gems. Through it all, the thread kept circling back to the same uncomfortable conclusion: if you can't run the model on your own hardware, assume it can be taken away at any moment.
A dumpster arrived behind my university's library [comments]
163 points · 159 comments · yalereview.org · 17h ago
The article is a first-person essay by an English professor watching thousands of deaccessioned books dumped into a dumpster during a library renovation, using Edith Wharton’s half-destroyed library and Derrida’s distinction between text and physical book to argue that discarding books kills the medium ideas travel through. The thread split hard: some dismissed this as routine collection management, noting the same library recently acquired a 444-year-old book and that weeding is necessary to make room for new acquisitions, while others pushed back hard that dumping books from a historical figure’s bequest or clearing shelves just for open lounge space is a new and objectionable trend. People who actually work in libraries argued most books have no value and that interlibrary loan makes keeping local copies unnecessary, but counter-arguments pointed out that coordinated last-copy systems like ReCAP exist precisely to prevent this kind of loss, and that the real problem is funding and space constraints in poorer districts. The conversation meandered into how e-readers with page-turning feel completely different from vertical scrolling on a phone or laptop, with several readers insisting they don’t F-pattern skim on their Kobos, and a nostalgic side thread about scavenging discarded library books as kids before that practice was banned by state property laws.
Show HN: Putt.day a daily mini golf game [comments]
150 points · 70 comments · putt.day · 9h ago
The linked article wasn't available to this summarizer; from the discussion, "Putt.day" is a daily browser-based mini golf game. The big controversy is the par of 6 – nearly everyone found it impossible on a straight playthrough, but the thread quickly filled with people sharing exploits: bouncing the ball off a wall at the tee to skip most of the course, reducing scores to 4, 3, or even 2 strokes. That led to a split between those enjoying the hidden shortcuts and those arguing that the physics are off (too much rolling resistance, camera angles that prevent full-power shots), with the developer showing up to acknowledge the feedback and push tweaks in real time. Some players just edited localStorage to cheat their score, while others genuinely appreciated the ghost-player feature and the "pooping-friendly" mobile design.
Looking Forward to Postgres 19: It's About Time [comments]
147 points · 40 comments · www.pgedge.com · 15h ago
Postgres 19 is adding native temporal table support based on the SQL:2011 standard, letting you ask "what did this data look like last Tuesday?" using range types and `WITHOUT OVERLAPS` constraints instead of clunky GiST extensions and manual row splitting. The HN crowd is mostly thrilled—many have been hand-rolling these patterns for years with exclusion constraints and stored procedures, and they see the new `FOR PORTION OF` syntax and temporal foreign keys as killing a whole class of footguns in domains like trading, tax rate corrections, and slowly changing dimensions. A small minority is uneasy about `UPDATE` operations now adding rows, which breaks a DBA assumption, but the consensus is that this is what tooling and ORMs need to make bitemporal design mainstream. The developer who worked on the feature chimed in, confirming that system-time (transaction time) still isn't native but is on their roadmap, and linked to talks about temporal relational algebra for the planner. A split emerges between those who think snapshotting databases like Dolt solve the same problem and others who argue that querying overlapping time ranges requires proper SQL algebra, not just versioned snapshots.
Adaptive PDFs [comments]
145 points · 69 comments · sgaud.com · 15h ago
The article introduces a technique that embeds hidden markdown structure into PDFs via the spec's replacement-text property, so text extractors return clean headings and tables while humans see the same visual layout. The HN thread quickly corrected the original misleading title ("A PDF that changes based on how its read"), since the document doesn't adapt to who's reading, only to how it's extracted — but the security implications got serious airtime: several people pointed out this is a perfect vector for injecting hidden instructions into LLM pipelines, extending an old problem of PDFs with scrambled copy-paste text. Some pushed back with "tagged PDFs already exist and are required for accessibility," though others noted most tools ignore them, and LaTeX users argued the toolkit can generate tags if you bother. The thread also derailed briefly on whether the article itself was AI-generated (author admitted to LLM polish for non-native English) and a separate bug where paragraphs were truncating mid-sentence, which the author fixed live.
European sunscreens are safer than American (2024) [comments]
136 points · 83 comments · www.ms.now · 18h ago
This one takes Alex Tabarrok's 2024 argument that the FDA’s drug-classification of sunscreen has locked out better UVA-blocking ingredients available in Europe, potentially worsening skin cancer rates. The discussion quickly pivots from the specific sunscreen chemistry into a broader brawl over regulation versus a free market, with one side enthusiastically using the article as evidence that FDA caution is deadly and the other pointing out that the free market also produced dangerous historical products like formaldehyde-laced milk. A key factual correction emerged: someone noted the FDA *just* approved a new sunscreen ingredient, bemotrizinol, for the first time in twenty years, which deflates the article’s premise that nothing has changed. Several Europeans also pushed back hard on the article's proposed 'peer approval' system—where the FDA would automatically accept EU approvals—arguing they trust the EMA far more than the FDA and see no reason to reciprocate, citing thalidomide as the canonical reason to be skeptical of American regulatory speed.
Ask HN: Why is there some sort of a scam website being advertised on HN?
133 points · 51 comments · news.ycombinator.com · 14h ago
The linked article wasn't available to this summarizer; from the discussion, it's an Ask HN pointing out that a scammy-looking ad for "Gauntlet AI" was sitting on the front page, and asking why. Hacker News immediately identified Gauntlet AI as the latest rebrand of BloomTech and Lambda School—the same YC-backed coding bootcamp that was fined and banned from lending for deceptive practices—and the thread is split between people disgusted that YC lets this company post privileged job ads and those who insist the program is legit because it's free and some participants got jobs. Several users dug up old posts showing the same company has been pulling this for over a year, and pointed out that Paul Graham once praised the founder's persistence in the face of "haters." The moderator Dang eventually stepped in, deleted the ad on the grounds that it technically wasn't a job posting, but the broader takeaway is that the community feels YC isn't doing enough to police its portfolio companies, especially when they keep rebranding to escape their track record.
Maxproof [comments]
132 points · 13 comments · arxiv.org · 20h ago
The MaxProof paper describes a system that uses a "defense-in-depth" generative verifier and population-level test-time scaling to push an LLM past the human gold-medal threshold on the 2025 IMO and 2026 USAMO. The HN thread quickly zeroed in on the fact that 46 human contestants also scored exactly 35/42 at the 2025 IMO, which was the modal score and the highest fraction of gold medalists since 1981 — making the AI's claimed score look less like a breakthrough and more like it got stuck in the same scoring traffic jam as the teenagers. Several commenters pointed out that the paper itself confirms the model solved 7/7 on the first five problems (which were unusually easy) and 0/7 on the famously hard problem 6, exactly matching the human pattern. There was pushback on the idea that partial credit is rare at the IMO (it's not), but the consensus was that the AI's 35/42 is less impressive when the human distribution clusters at the same number and the easy/hard split is the same. One person wondered whether the test-time scaling harness is more valuable than the model weights themselves, while another cracked that this proves the need for more formal verification — and that "Maxproof" is a terrible name.
Where Did Earth Get Its Oceans? Maybe It Made Them Itself [comments]
125 points · 62 comments · www.quantamagazine.org · 16h ago
The article lays out the evolving scientific debate over where Earth's oceans came from, moving from comets to asteroids to a newer theory that the planet may have cooked up its own water by reacting hydrogen with magma under high pressure. HN largely ran with the broader implications for life and civilization, sparking a lengthy argument about whether an intelligent aquatic species could ever develop technology—many pointed out that water makes fire, tool manipulation, and writing far harder, while others countered with orcas and octopuses as near-misses that just haven't had enough time or generational knowledge transfer. A separate thread zeroed in on the "sample size of one" problem, saying we can't draw strong conclusions from a single example of civilization. The article's own hedging—that the answer is probably a mix of all three sources—got folded into the discussion as a reminder that the question remains far from settled.
Tell HN: Meta is down
124 points · 94 comments · news.ycombinator.com · 18h ago
Meta is down—Facebook, Instagram, Messenger—and the submitter notes that their own status page at metastatus.com isn’t registering the outage, hinting at something more widespread. The thread immediately fixates on the fact that Downdetector itself went to a 404 under the load, sparking a chain of jokes about recursive down detectors and a brief meta-outage rabbit hole. A split emerges: some are gleefully hoping it stays down or suffers a catastrophic data loss (calling it a "modern miracle"), while others point fingers at AI-assisted code changes, BGP misconfigs, or the usual Friday-afternoon release curse. People with business accounts are getting roasted by clients because Meta’s APIs are dead but their status page shows nothing, and a flood of new accounts shows up because this HN post is the top Google result for "is meta down"—leading regulars to welcome them to "one of the few places worth lurking."
EV demand up 50% in France and Germany since Iran war [comments]
118 points · 76 comments · www.reuters.com · 13h ago
The linked article wasn't available to this summarizer; from the discussion, Renault’s CEO reported EV orders surged 50% in France and Germany since the start of the Iran war, and the thread turned that into a debate about whether war and Trump’s blockade policies are acting as a de facto—if brutal—carbon tax that’s finally pushing people off fossil fuels. Some commenters argued it’s wild to make a multi-year car purchase based on a short-term geopolitical shock, but others pushed back hard, saying the real driver is that European drivers have been teetering on the edge of switching for years—better batteries, cheaper models like the Renault 5, and city bans on ICE—and the war just tipped them over. A few people with domain knowledge noted that Renault’s EVs use rare-earth-free motors, and the conversation spun out into the US strategic petroleum reserve running dry in weeks, whether the US is actually self-sufficient in oil (it’s not, because refineries need heavy crude imports), and how much of the fuel price spike simply flows to oil company profits rather than functioning like a proper carbon dividend. The overall consensus: nobody actually thinks war is a good way to accelerate EV adoption, but plenty of people are treating it as a rational hedge and finally admitting that petroleum has always been the unreliable partner in the relationship.
Tesla Full Self Driving uses bicycle lane in official Denmark approval video [comments]
114 points · 50 comments · politiken.dk · 14h ago
The linked article wasn't available to this summarizer; from the discussion, Tesla’s own PR video for its Full Self-Driving approval in Denmark shows the car turning into a bicycle-only lane within seconds, ignoring clearly posted no-cars signs, painted bike symbols, and a yellow center marker. The thread immediately zeroed in on how Tesla curated and released that footage—it cuts just as the car enters the lane, and the wider context makes clear the maneuver is illegal in Denmark, not some California-style merge-before-turn rule. Several people dug into Tesla’s persistent failure to recognize basic traffic signs like “Do Not Enter” and “One Way,” arguing this isn’t a one-off but a systemic software blind spot that the company has chosen not to fix. Others pushed back with claims that the critics are smear artists or that drivers will just eat the ticket and Tesla will patch it later, but the dominant take is that approving a system that can’t read a bike-only sign is reckless, especially when lidar-equipped competitors like Waymo handle such edge cases far more reliably.
Keygen.music [comments]
110 points · 68 comments · keygen.music · 16h ago
This is a site collecting tracker music modules (MOD, XM, S3M, IT) from demoscene and cracking groups, organized by the groups that made them and playable through a built-in web player. The thread absolutely savaged it — the near-universal consensus is that the site is “vibecoded” LLM slop with broken navigation, a confusing drag-float player, no obvious play button, and general UX that disrespects the demoscene’s tradition of technical craft. Multiple people pointed to keygenmusic.tk, chiptune.app, and modarchive.org as superior alternatives that have been around for years. A long sidebar argument broke out about whether AI should generate lofi samples for tracker music, with proponents arguing it’s just another synthesis method and opponents countering that burning compute on a bass kick is wasteful when waveform generators already exist — but that debate was mostly a distraction from the main event, which was dunking on the site for being exactly the kind of thoughtless AI output the scene would hate.
Generated 2026-06-13 08:20 UTC
Generated by Sauron from Hacker News discussions and linked articles.