0:00 / 1:35ai HackerRank’s AI resume scorer looks like a luck filter
A developer tested HackerRank’s open-source hiring-agent resume scorer and found the same resume could receive wildly different scores: 90, then 74, and across 100 runs a range from 66 to 99 using the default local Gemma 3 4B model. The tool parses a PDF, calls an LLM multiple times to extract resume sections, checks GitHub context, then grades candidates with heavy weight on open source and personal projects. The piece argues that this turns hiring into a luck filter: technical-skill checklist items were mostly stable, while project judgments varied, and experience scoring maxed out even for an old resume with one internship.
Discussion: Negative — HN is broadly alarmed and skeptical. The dominant reaction is that LLM-based resume scoring is noisy, legally and ethically risky, and likely to amplify already-broken hiring filters; a few commenters add nuance about applicant volume and model size, but even those tend to frame this as a crude triage mechanism rather than a fair evaluation system. (LLM nondeterminism and reproducibility, AI resume screening as arbitrary filtering, Overweighting GitHub, open source, and side projects)
0:00 / 0:39ai Qwen 3.6 27B makes local coding feel practical
The post argues that Qwen 3.6 27B is a new sweet spot for local development: slower than the Qwen 3.6 35B A3B mixture-of-experts variant, but, in the author’s view, higher quality for coding and general tasks. The author shows it running through llama.cpp with an 8-bit GGUF quantization and multi-token prediction, reporting about 30 tokens per second on a 128GB M5 Max MacBook and noting both Qwen 3.6 variants fit within 48GB of Apple Silicon shared memory. The broader point is that local open-weight models are becoming useful enough for offline work, private code, sensitive data, and setups that cannot be taken away by a model provider.
Discussion: Mixed — HN is interested in Qwen 3.6 and many commenters report good local results, but the thread is dominated by practical skepticism: heat, fan noise, high-end hardware costs, and whether local inference beats cheap cloud APIs. Several commenters push back on the article’s MacBook-centered framing, arguing for dedicated desktops, used Nvidia GPUs, Mac minis or studios, DGX Spark-style machines, or simply OpenRouter/frontier-model credits. (Qwen 3.6 27B seen as a strong local coding model, MacBook heat, fan noise, and ergonomics are a major concern, Hardware economics versus cloud API pricing)
0:00 / 0:33biotech Claude challenged an MRI diagnosis — and HN urged a real second opinion
A writer fed a 266 MB DICOM export of a shoulder MRI into Claude Code running Opus 4.8 after a clinic diagnosed a Grade III partial-thickness subscapularis tendon tear and began recommending interventions. Claude’s first pass said the tendon was intact, and a later comparison report favored mild tendinosis with no discrete tear, leaving the author unsure whether to trust the clinic, the AI, or neither. The story matters because it captures a near-future patient behavior: using frontier models to interrogate medical reports, while exposing how risky that is when the task is specialist image interpretation.
Discussion: Mixed — The discussion is interested in AI as a way to ask better medical questions, but strongly skeptical of using general-purpose LLMs to interpret MRI images. Radiologists and other clinicians in the thread repeatedly warn that Claude and ChatGPT are not reliable for radiology, while some commenters share anecdotes where AI helped them challenge or understand medical care. The dominant practical advice is: use AI for research and clarification, but get a human second or third opinion for imaging and treatment decisions. (LLMs are not trusted for medical image interpretation, AI can help patients ask better follow-up questions, Doctors and clinics may still over-treat or misdiagnose)
0:00 / 0:43science Max Planck’s retracted papers look like a publishing-system mistake
Ars Technica reports that historians Yves Gingras and Mahdi Khelfaoui investigated why Naturwissenschaften, now The Science of Nature, had removed two 1940s papers by Max Planck and replaced them with blank pages noting an “article violation.” The journal’s current editor-in-chief, Suzanne Scarlata, told Science she had not known about the retractions and suspected an algorithmic mistake that should be fixed. The case matters because it shows how modern publishing infrastructure can attach serious-looking retraction labels to historic scientific work without an obvious scientific-integrity issue.
Discussion: Mixed — Commenters largely agree this does not reflect on Planck’s science, but they are irritated by modern copyright and duplicate-publication rules being applied to 1940s work. A secondary thread criticizes clickbait-style headlines and notes this story had already circulated on HN. (Retractions seen as administrative or copyright-related, not scientific misconduct, Skepticism toward the idea of self-plagiarism, Concern that automated publisher systems can mislabel historical papers)
0:00 / 1:28software OpenRA playtest adds random maps and Dune 2000 upgrades
OpenRA has published a new playtest with random map generators for Red Alert, Tiberian Dawn, and Dune 2000, usable in both skirmish and multiplayer. Dune 2000 gets visual upgrades, Starport bulk purchasing, balance changes, and campaign difficulty tweaks, while the standalone Tiberian Dawn HD mod is now feature-complete with support for C&C Remastered Collection assets. The release also improves the map editor, adds timed autosaves, new missions, bot expansion behavior, and other fixes, keeping a long-running open-source RTS revival active and playable on modern systems.
Discussion: Positive — The HN mood is strongly nostalgic and appreciative, with many commenters praising OpenRA as a high-quality modernization of classic Command & Conquer-era RTS games. The main reservations are around AI balance, pathfinding, and especially long load times for very large saved games, while side threads celebrate LAN-era multiplayer, modding, RA2, and open-source engine remakes more broadly. (Nostalgia for Red Alert, Command & Conquer, and LAN multiplayer, Praise for OpenRA’s balance and quality-of-life improvements, Debate over AI difficulty, cheating, and micro-management)
0:00 / 0:49software A graphical shell rides over SSH
The author proposes a native graphical shell for SSH: servers expose app UIs as small HTTP services, usually on Unix domain sockets, with encryption and access handled through SSH rather than public ports. They have built Outer Loop as a dedicated SSH-aware browser and are releasing an open-source Outer Shell with APIs for app discovery and integration. The pitch is a remote-first graphical environment for servers, edge devices, ML boxes, and robots—somewhere between web admin panels and a GUI shell, but not just a traditional remote desktop.
Discussion: Mixed — HN was intrigued by the usability goal but heavily skeptical of the novelty and safety of the design. Many commenters compared it to existing approaches like Cockpit, X11 forwarding, SSH port forwarding, SOCKS proxies, Caddy, RDP, VNC, and Windows Admin Center, while a smaller group welcomed experiments that move remote work beyond character-grid terminals. (Prior art: Cockpit, X11 forwarding, RDP/VNC, Windows Admin Center, Security concerns around browser-like access to Unix sockets, Usability gap between SSH/TUIs and graphical remote workflows)
0:00 / 1:31security Open-weight GLM 5.2 surprises Semgrep on AI bug hunting
Semgrep says GLM 5.2, an open-weight model from Zhipu AI, scored 39% F1 on its IDOR vulnerability-detection benchmark, beating Claude Code in its comparison while costing roughly $0.17 per real vulnerability found. Semgrep’s own multimodal pipeline still led the test at 53–61% F1, likely helped by a purpose-built harness that enumerates endpoints and guides the model through relevant code. The result matters because it suggests open-weight models may now be viable for some security-analysis workloads, but the benchmark is narrow and the harness differences make broad model rankings risky.
Discussion: Mixed — HN was broadly impressed by GLM 5.2’s price-to-performance and several commenters reported good coding results in real projects. But the thread was also skeptical of the headline and methodology, especially the narrow IDOR task, the different harnesses being compared, and the practicality of running a 753B-parameter model locally. (Strong enthusiasm for cheaper open-weight coding models, Debate over model quality versus harness quality, Skepticism about apples-to-pears benchmark comparisons)
0:00 / 0:36hardware LibrePods brings Apple-only AirPods features to Linux and Android
LibrePods is an open-source project that implements the proprietary protocol AirPods use to talk to Apple devices, bringing features like noise-control switching, ear detection, accurate battery status, renaming, conversational awareness, and some accessibility settings to Linux and Android. Some features are implemented, others are planned or unknown, and several Android capabilities may require VendorID spoofing, Xposed, or root. The project matters because it narrows the gap between basic Bluetooth compatibility and the richer AirPods experience normally reserved for Apple’s ecosystem.
Discussion: Mixed — The thread is broadly impressed by the reverse-engineering work and the practical value for Android and Linux users, but skeptical about Apple’s ecosystem lock-in and whether firmware updates could break it. Several commenters clarify that AirPods already function as ordinary Bluetooth earbuds off Apple devices; LibrePods is about the extra Apple-integrated controls and settings. There are also cautionary notes about setup side effects, including one report of hearing-aid settings being removed. (Respect for the reverse-engineering effort, Frustration with Apple ecosystem lock-in, Clarification that baseline Bluetooth audio already works)
0:00 / 0:25hardware Memory prices, from core memory to the AI squeeze
Stanford DAM’s David Shim published an interactive memory and storage price tracker, extending the spirit of John C. McCallum’s classic dataset with downloadable raw data. It charts DRAM, NAND flash, and HBM costs, using McCallum’s historical data, Amazon/Keepa retail pricing for newer DRAM and SSD points, Epoch AI accelerator cost estimates, and sparse analyst estimates for HBM. The big caveat is that the headline $/GB figures are nominal cheapest listed retail prices, not inflation-adjusted contract prices or confirmed sales, while the HBM data is modeled rather than market-quoted.
Discussion: Mixed — HN liked having a revived, downloadable historical memory-price dataset, but the discussion quickly turned into caveats about methodology and frustration over current RAM prices. Commenters debated log scales, inflation adjustment, whether dollars per gigabyte is meaningful for older eras, and how AI demand and industry cyclicality are changing the market. (appreciation for preserving and extending the McCallum memory-price dataset, concern that cheapest-retail $/GB can understate real current prices, debate over inflation adjustment and log-scale interpretation)
0:00 / 0:52hardware Rocket Lab makes an eight-billion-dollar bet on owning the whole space stack
Rocket Lab and Iridium announced a definitive agreement for Rocket Lab to acquire Iridium in a cash-and-stock deal valued at $54 per Iridium share, implying about $8.0 billion in enterprise value. The deal would combine Rocket Lab’s launch and spacecraft manufacturing business with Iridium’s global L-band satellite network, spectrum, 2.55 million active subscribers, and recurring communications revenue. The companies say the transaction is expected to close in mid-2027, subject to Iridium shareholder approval and regulatory approvals, with Rocket Lab arranging a $3.6 billion bridge loan commitment for the cash component.
Discussion: Mixed — HN was split between seeing the deal as a smart strategic hedge for Rocket Lab and worrying that cheaper, vertically integrated space infrastructure will accelerate satellite crowding and orbital debris. The most engaged thread veered into space-junk, Kessler-syndrome, and orbital-commons regulation, while a smaller set focused directly on the business logic of using Iridium’s constellation and replacement needs to create demand for Rocket Lab launches and satellite manufacturing. (Vertical integration from launch and satellite manufacturing into operating communications services, Iridium’s spectrum, subscribers, and recurring cash flow as strategic assets, Concern about LEO congestion, orbital debris, satellite visibility, and atmospheric effects of reentry)
0:00 / 1:20policy Age checks as real-name infrastructure
The linked post argues that internet age-verification laws are not just about keeping minors out of restricted spaces, but about building systems that tie online accounts to real-world identities. Its core claim is that once identity checks become common, governments or law enforcement could more easily attribute speech to individuals at scale. That matters because age-gating proposals are spreading across US states, Europe, and Australia, and the technical design choices could reshape anonymity, platform liability, and everyday access to the web.
Discussion: Mixed — The HN thread is strongly wary of age-verification mandates, with many commenters treating them as a privacy and speech-control infrastructure rather than a child-safety tool. A minority pushes back that online age restrictions have real policy goals, that slippery-slope arguments can be overused, or that privacy-preserving cryptographic approaches could make age checks less dangerous. (Fear that age checks will become real-identity attribution for online speech, Concern about government-gated internet access and device attestation, Debate over whether zero-knowledge or cryptographic ID systems could solve the privacy problem)
0:00 / 1:30policy Bogus Copyright Claim Knocks a Pollen Investigation Off Google
Gergely Orosz says Google removed his 2022 Pragmatic Engineer article about the collapse of events startup Pollen from search results after a copyright complaint. The complaint allegedly claimed his Pollen investigation copied a 1998 New York Post article, despite the two not sharing a sentence, and was filed under a seemingly fake identity from Bouvet Island, an uninhabited territory. Orosz has appealed and argues the episode shows how copyright takedown systems can be abused to suppress unfavorable reporting.
Discussion: Negative — HN is strongly critical of the takedown system and of Google’s handling of apparently bogus copyright complaints. Commenters largely see this as reputation management weaponizing DMCA-like processes, while debating whether identity verification, court orders, attorney signatures, notarization, or real penalties for false claims would be practical fixes. (Abuse of DMCA and DMCA-like takedown processes, Google’s role as a private moderator with weak accountability, Lack of enforcement for false claims filed under penalty of perjury)
0:00 / 1:30policy EU ‘Chat Control’ fight flares up again behind closed doors
Former MEP and civil liberties advocate Patrick Breyer says EU institutions are moving on two fronts to revive or advance Chat Control measures: a push around the temporary Chat Control 1.0 regime and a permanent CSAR trilogue that could include scanning mandates and age verification. Breyer argues the moves could enable mass scanning of private messages, warrantless detection orders, and an effective end to anonymous online communication in Europe. The fightchatcontrol.eu campaign has been relaunched to urge citizens to contact lawmakers and government representatives before the meetings.
Discussion: Negative — HN reaction is overwhelmingly hostile to the reported negotiations, with commenters framing Chat Control as a threat to privacy, end-to-end encryption, anonymity, and democratic legitimacy. A few comments push for more precise analysis of EU legislative mechanics or distinguish privacy from anonymity, but the dominant mood is anger, fatigue, and distrust of EU institutions and lobbying pressure. (Fear of mass surveillance and weakening end-to-end encryption, Frustration at repeated attempts to revive rejected proposals, Distrust of closed-door EU trilogue and Council processes)
0:00 / 0:22policy EFF warns the KIDS Act could make age checks the default online
EFF says Congress is preparing to vote on the KIDS Act, a bundled package that includes a revised KOSA plus other internet bills, under an expedited process. The group argues that even where the text says age verification is not required, standards like “knows or should have known” a user is under 17 will push platforms to verify everyone’s age to avoid liability. EFF also warns the bill could pressure platforms to over-moderate lawful speech and create new risks for private or encrypted messaging.
Discussion: Negative — HN reaction is strongly hostile to the bill as described, with many commenters framing child safety as a pretext for identity checks, surveillance, and broader internet control. A minority argues that social media harms to children are real and policymakers are responding to parental concern, but even some of those commenters criticize the implementation as overbroad or flawed. (Age verification as de facto surveillance, Privacy risks from ID collection and third-party verifiers, Skepticism of the child-safety rationale)
0:00 / 0:30policy Mullvad Faces a Political Funding Backlash
A Mastodon post alleged that Daniel Berntsson, one of Mullvad’s two founders, owners, and CEOs, is the main financier of Sweden’s Örebro Party. Mullvad co-founder Fredrik Strömberg joined the HN thread to clarify that Mullvad has two equal owners and said Berntsson’s private donation is not part of Mullvad’s values or mission. The story matters because Mullvad’s brand is built on trust, privacy, and political principle, and users are now debating whether paying for the service indirectly supports politics they reject.
Discussion: Mixed — The thread is polarized and emotionally charged. Many Mullvad customers say they will cancel because owner profits may fund politics they oppose, while others argue the company’s privacy work is unusually important and should be judged separately from a co-owner’s private donations. (Ethical consumer choices around founder politics, Mullvad's trust and privacy reputation, Disagreement over how to classify Örebropartiet ideologically)
The Supreme Court ruled 6-3 in Chatrie v. US that geofence warrants seeking smartphone location data are Fourth Amendment searches and require constitutional privacy protections. Justice Elena Kagan’s majority opinion rejected the idea that users meaningfully waive privacy merely by carrying phones or enabling Google Location History, while leaving lower courts to decide whether the specific search here was reasonable and properly particularized. The decision matters because geofence warrants can identify large numbers of people near a place and time, not just known suspects, making this a major digital-era privacy ruling.
Discussion: Positive — HN is largely supportive of the ruling as a needed privacy check on dragnet location searches, though the thread is not celebratory across the board because commenters worry warrants can still be broad, rubber-stamped, or saved by the good-faith exception. The mood is privacy-forward and skeptical of law enforcement access to third-party location data, with some debate over how far the decision reaches. (Broad support for treating phone location histories as Fourth Amendment-protected searches, Concern that innocent bystanders near a crime can be swept into investigations, Debate over the good-faith exception and whether illegally or improperly gathered evidence is still usable)
0:00 / 0:29policy Memory giants face a new U.S. price-fixing suit
Samsung Electronics, SK Hynix, and Micron have been sued in California federal court by 14 consumers and three small businesses alleging they coordinated DRAM supply and pricing from 2022, contributing to roughly 700% price increases over four years. The plaintiffs say the companies used the transition to high-bandwidth memory and the discontinuation of DDR3 and DDR4 as a pretext to constrain supply; they are seeking class-action status that could cover buyers of products containing DRAM. The case matters because these three companies dominate DRAM, prior DRAM collusion cases led to fines and prison sentences, and a successful U.S. antitrust class action could mean treble damages.
Discussion: Mixed — HN is angry about soaring memory prices and quick to recall the early-2000s DRAM price-fixing scandal, but many commenters are skeptical that this complaint proves collusion. The main debate is whether the shift from DDR3 and DDR4 toward DDR5 and HBM is normal capacity allocation under AI-driven demand, or coordinated supply restriction by an oligopoly. (Prior DRAM price-fixing convictions make users suspicious of Samsung and SK Hynix, Several commenters note a similar 2022 case failed because plaintiffs could not show an agreement, Debate over whether ending DDR3 and DDR4 production is anti-competitive or just market transition)
0:00 / 0:25policy European ISPs want copyright blockers to pay for collateral damage
EuroISPA has told the European Commission that piracy-blocking regimes are becoming disproportionate and are causing collateral damage to legitimate internet services. The group cites incidents including Italy’s Piracy Shield blocking affecting more than 7,700 domains, a Portuguese hosting provider losing email connectivity with Italian customers for 16 days, and Spain’s LaLiga orders allegedly disrupting access to banking apps, developer tools, and payment platforms. EuroISPA wants rightsholders held liable for overblocking damage, with clear compensation mechanisms, arguing this can be supported under existing EU enforcement law rather than new legislation.
Discussion: Mixed — HN is broadly supportive of EuroISPA’s call for accountability, with commenters seeing overblocking liability as overdue. The mood is strongly distrustful of rightsholders and automated or fast-track takedown systems, but the discussion also branches into harder questions about where legitimate content removal ends and censorship begins. (Support for compensating victims of overblocking, Distrust of copyright enforcement and rightsholder power, Concern about shared-IP, DNS, VPN, and platform-level blocking)
0:00 / 0:24policy Instagram Photos in Meta Glasses Ads Spark a Privacy Backlash
A viral X post claims Instagram is incorporating users’ photos into ads for Meta Glasses, framing it as the arrival of ultra-personalized advertising. The HN discussion quickly connected the claim to Meta’s long-standing terms allowing user content and identity signals to appear in sponsored contexts. The concern is not just whether users technically agreed, but whether friends, bystanders, and people dependent on Instagram have any meaningful choice.
Discussion: Negative — HN commenters are overwhelmingly critical, treating the reported ad use as another example of Meta pushing user consent and platform lock-in too far. Some note this is not new—Facebook had similar sponsored-content terms as far back as 2013—but that history makes the reaction more resigned than reassured. (Broad platform terms of service and consent, Use of personal likenesses in advertising, Bystanders and non-users appearing in uploaded photos)
0:00 / 0:33policy Tidal lets AI music in, but turns off the money
Tidal’s new AI policy says the service will accept AI-generated music, but will hold it to content-integrity rules and will not monetize AI-generated tracks. The policy targets music that exploits an individual’s or group’s music, name, or likeness, deceives listeners, or lowers the quality of the service. The bigger significance is that Tidal is trying to change the incentives around generative music: allow it to exist, but remove the payout that can fuel spam and impersonation.
Discussion: Mixed — HN is broadly supportive of Tidal demonetizing AI-generated music and requiring clearer treatment of it, but the thread is full of caveats. Commenters worry about vague definitions, hard detection problems, artist impersonation, and whether users will get a true opt-out rather than just labels. (approval for demonetizing AI-generated tracks, concern over artist spoofing and likeness misuse, requests for AI-music opt-out or human-verified platforms)
0:00 / 0:25general Public radio’s bedtime podcast reads the boring paperwork
Marfa Public Radio launched “Marfa Public Radio Puts You to Sleep,” a sleep podcast for its fall membership drive that reads the station’s dull but essential documents, including compliance, protocols, emergency response, maintenance, and NPR ethics material. The pitch is both a joke and a fundraiser: listeners are meant to fall asleep, then wake up and donate to keep the station running 24/7.
Discussion: Positive — HN mostly loved the gimmick and used it as a jumping-off point to trade favorite sleep aids, from fictional baseball broadcasts to BBC Radio 4, the Shipping Forecast, chess videos, and meandering podcasts. A few readers reported access issues or said topics like journalistic ethics might be too interesting or politically charged to sleep through, but the overall mood was amused and appreciative. (Playful appreciation for intentionally boring audio, Recommendations for sleep podcasts, radio, sports, and white noise, Interest in Marfa as a quirky arts town)
0:00 / 0:24general Five thousand old menus show how dining changed — and didn’t
The Pudding built a curated story and visualization around roughly 5,000 menus from the New York Public Library’s Buttolph Collection, focused on the late 19th and early 20th centuries. The appeal is that menus are small historical artifacts: they show what restaurants served, how dishes were categorized, what ingredients were fashionable, and how dining culture looked on paper. Readers found the project both visually delightful and surprisingly informative, especially where old menus resemble modern ones while featuring now-unusual items and categories.
Discussion: Positive — HN readers were strongly charmed by the collection and treated it as a window into restaurant history, food culture, typography, and pricing. The thread is mostly appreciative and curiosity-driven, with some side tangents into beer coasters, dim sum plates, QR-code menus, and related food-history resources. (Food history and forgotten dishes, Menu design and printing aesthetics, Continuity between old and modern restaurants)
0:00 / 0:39general Zanagrams wins HN over with a disappearing-letter word puzzle
Zanagrams is a Show HN web word puzzle where players trace connected letters to find words, and solved letters or links disappear when no longer needed. The launch drew a big response, and the creator quickly added a playable tutorial, an optional timer, word-list fixes, and a Hacker News-themed puzzle. It matters because the thread shows a small, polished game getting rapid product feedback in public—and many commenters seemed ready to make it part of a daily puzzle habit.
Discussion: Positive — The discussion is strongly positive: commenters found the game polished, engaging, and replayable, with multiple people asking for more puzzles. The main criticism is not the core idea but usability and puzzle-design edges: clearer instructions, hints, a give-up/reveal option, scoring for bonus words, and frustration around plurals or obscure vocabulary. (High praise for the minimal UI, animations, and satisfying progression as letters disappear, Requests for hints, reveal/give-up, better sharing, timers, and bonus-word scoring, Some confusion about controls, rules, and whether word order can make a puzzle unwinnable)