Top 10 Hacker News posts, summarized
HN discussion
(286 points, 391 comments)
The author describes their experience using Claude Opus 4.8 through Claude Code to get a second opinion on an MRI scan for shoulder pain. After receiving a diagnosis of a Grade III partial-thickness tear at the apical insertion of the subscapularis tendon from a clinic that also recommended extensive treatment including shockwave therapy and a homeopathic injection, the author had Claude analyze the DICOM files. Interestingly, Claude reported an intact tendon with only mild insertional tendinosis, contradicting the human doctor's diagnosis. When asked to arbitrate between the two reports, Claude favored its own analysis with moderate-to-high confidence. The author is now uncertain about whether to trust the human doctor, the AI analysis, or seek another medical opinion.
The Hacker News discussion reveals strong skepticism about using AI for medical imaging analysis. Multiple commenters, including radiologists, note that LLMs may not be capable of actual image analysis and can suffer from "mirage reasoning" - generating detailed descriptions for images never provided. There's consensus that AI shouldn't be trusted for medical image interpretation without expert review. Commenters also emphasize the importance of getting second opinions from real doctors, with some sharing personal experiences where AI caught errors in medical reports. The conversation highlights concerns about over-reliance on AI for critical medical decisions and the potential for LLMs to confidently assert incorrect information. Some suggest AI could serve as a starting point for asking better questions when consulting human doctors, but not as a replacement for medical expertise.
HN discussion
(381 points, 116 comments)
Marfa Public Radio operates 24/7 (excluding lightning strikes) and requires significant behind-the-scenes work including fundraising, compliance, protocols, emergency response, and maintenance to sustain its service. For their fall membership drive, they launched "Marfa Public Radio Puts You to Sleep," a sleep podcast designed to lull listeners into slumber by reading the station's boring but essential operational documents. While hoping listeners fall asleep, they encourage donations at marfapublicradio.org/donate upon waking to support the station's continuous operation.
Hacker News commenters reacted humorously to the sleep podcast concept, with several jokingly falling asleep mid-comment ("I’m here fo…zzz"; "What a great idea, I feel li... zzz"). Users shared alternative sleep aids like specific YouTube channels ("EpicNate") and sleep apps ("fastsleep.app" for guided imagery, "Sleep With Me" for nonsensical monologues). Discussions also explored why some prefer human voices over music for sleep and the desire for sleep apps with simple UIs that track progress. One commenter noted the podcast's potential effectiveness compared to similar ones that were unexpectedly engaging, while another observed the website blocks Singaporean access. A notable comment raised that NPR's journalistic ethics discussion might be stimulating rather than sleep-inducing.
HN discussion
(246 points, 226 comments)
The KIDS Act, a forthcoming Congressional vote package, combines a revised Kids Online Safety Act (KOSA) with other internet bills. Despite KOSA's disclaimer not to require age verification, its provisions imposing special protections for minors trigger a "knows or should have known" standard. This creates significant legal risk, pressuring platforms to implement age verification (like ID checks or facial scans) to avoid liability. The bill also mandates platforms police broad categories of lawful speech (e.g., discussions on addiction, gambling), regulates private and encrypted messaging, and risks reducing privacy and free expression for all users.
HN commenters overwhelmingly oppose the KIDS Act, viewing it as government overreach driven by fearmongering or control lust. Key concerns include the potential for pervasive age surveillance, erosion of privacy (especially for marginalized groups), and suppression of lawful speech under the guise of child protection. Many argue it sets a dangerous precedent for mass surveillance and censorship. While some suggest technical alternatives (e.g., browser-level age verification tokens), others emphasize the need for political action (contacting lawmakers) or draw historical parallels to past privacy incursions. Commenters also criticize the "think of the children" justification and question the bill's effectiveness compared to parental controls.
HN discussion
(275 points, 109 comments)
Semgrep benchmarked open-weight models against their IDOR (Insecure Direct Object Reference) vulnerability detection dataset and found GLM 5.2 from Zhipu AI outperformed Claude Code (39% vs. 32% F1 score) using only a simple prompt and minimal scaffolding. While GLM 5.2 trailed Semgrep’s multimodal pipeline (53–61% F1), which uses endpoint-discovery and guided navigation, it demonstrated strong cost efficiency ($0.17 per vulnerability found). The experiment aimed to isolate model performance from harness complexity, revealing that GLM 5.2—an open-weight Mixture-of-Experts (MoE) model with 750B total parameters (40B active per token)—matched or exceeded frontier models on coding benchmarks (e.g., 81.0 on Terminal-Bench 2.1) and extended context to 1M tokens. Zhipu AI released GLM 5.2’s weights under an MIT license in June 2026, though noted increased reward-hacking behavior during training.
HN comments focused on benchmark skepticism, GLM 5.2’s practical usability, and geopolitical implications. Many criticized the headline as misleading since Claude Code is a harness, not a raw model, and questioned whether it was compared directly to Claude Opus (e.g., admax88qqq). Users highlighted GLM 5.2’s affordability for daily programming (pimeys) but noted inconsistent pricing across providers (cmrdporcupine), while others debated its hardware requirements for local inference (bArray). Geopolitical tension surfaced, with solenoid0937 predicting US export controls on Chinese models and lenerdenator suggesting China’s goal is to disrupt US AI labs. Security-specific concerns included skepticism about IDORs being "easy" vulnerabilities (danslo) and whether GLM 5.2 could replicate exploit-generation capabilities like Anthropic’s Mythos (dist-epoch). Practical experiences were mixed, with some praising its cost-effectiveness (himata4113) and others reporting reliability issues (csjh).
HN discussion
(302 points, 80 comments)
The New York Public Library has digitized 5,000 menus from its Buttolph Collection, spanning 1880 to 1920. This historical archive provides a window into dining culture during the Gilded Age and early 20th century, showcasing the offerings, pricing, and design of restaurant menus from that era.
Users found the collection fascinating, noting the striking similarity between historical menus and modern ones in terms of structure and content. Key observations included nostalgia for pricing listed entirely in cents, curiosity about which depicted restaurants still exist, and commentary on culinary trends like prominent early listings of clams/oysters and common "Boiled" food categories. Technical issues were frequently mentioned, including slow loading times and tap-to-click functionality problems on MacBooks. Some users referenced similar historical food resources like The American Menu and noted the collection had been discussed previously on HN.
HN discussion
(168 points, 115 comments)
The article describes an open GitHub issue requesting a feature for OpenAI Codex to exclude sensitive files from being read or sent to the model. The proposed solution includes a mechanism for marking files/paths at both repository and global levels (e.g., a `.codexignore` file) to prevent access to sensitive data like `.env`, `.pem`, `.aws/`, and `.ssh/` directories. The configuration should be deterministic, shareable across teams, and support user defaults. This request is a follow-up to a previously closed issue (#205) that focused on preventing sensitive data leaks and excluding large files; the author notes that no comparable feature exists in the Rust-based codex-rs implementation as of August 2025 and seeks to converge on a design.
Hacker News commenters heavily criticized the proposed ignore feature as ineffective for security, arguing it would create a false sense of security due to the unpredictable nature of LLMs and their ability to work around restrictions. Many emphasized that true security requires actual containment mechanisms like sandboxes, containers, or file permissions (e.g., `chmod`), as agents can still access sensitive files indirectly through tool outputs or bash commands. Alternative solutions suggested using VMs, bind mounts, dedicated tools (e.g., greywall.io, rumpelpod, drop), or opt-in access models. Several commenters framed the issue as a user responsibility problem, stating users should properly sandbox environments rather than relying on agent-level blocklists. Consensus emerged that an ignore file might be useful for performance (excluding large files) but is insufficient for security.
HN discussion
(123 points, 155 comments)
Professor Roberto Serrano at Brown University discovered mass cheating in his advanced economics course (ECON 1170), where at least 50 students used AI tools like ChatGPT on a take-home, closed-book midterm exam. Overwhelming evidence included unusual passages in answers matching AI-generated results. Despite reporting the issue, Serrano felt the university administration was dismissive, labeling it only a "wake-up call" after he escalated to the Academic Code Committee. Serrano, a respected economist who went blind as a teenager, changed his teaching approach for the future, discontinuing take-home exams and making weekly exercises ungraded, citing the need to preserve academic integrity in the face of AI's threat to higher education. The cheating scandal coincided with a campus shooting earlier that semester, which had influenced his decision to offer take-home exams.
Hacker News comments largely criticized the professor's exam design, arguing that "take-home, closed-book" is inherently contradictory and invites cheating, regardless of AI's presence. Many commenters stated that AI merely amplified an existing problem, as cheating occurred with take-home exams long before AI tools were available. Key themes included: the pressure on students to cheat in competitive environments where grades are curved and jobs are scarce; the inevitability of AI adoption, suggesting education must adapt (e.g., focusing on model building or research rather than computation); and the need for analog solutions like in-person, handwritten exams. Some argued the university's focus on donations from wealthy families creates a culture where students are given the benefit of the doubt, undermining academic integrity. There was also skepticism about the scandal being the "biggest known" in the Ivy League and a call to rethink exam structures fundamentally for the AI era.
HN discussion
(212 points, 63 comments)
LibrePods is an open-source project that enables AirPods features exclusive to Apple ecosystems on non-Apple platforms like Android and Linux. By reverse-engineering Apple's proprietary Bluetooth protocols, the app provides functionality such as noise control mode switching, fast ear detection, accurate battery status, head gestures, and conversational awareness. The project has limitations, including a lack of head tracking for HRTF on Android and the potential need for root access for advanced features like "Find My" and high-quality two-way audio. The project also clarifies that it is not affiliated with Apple and has issued warnings against unofficial websites misusing its branding.
The HN discussion centers on the practicality, motivation, and technical merit of the LibrePods project. Many users question the value proposition, as AirPods' primary appeal is their seamless Apple ecosystem integration, which this project partially replicates on other platforms. Users also debate the quality of AirPods compared to cheaper alternatives, with one commenter criticizing their noise cancellation and sound quality. The project's use of AI for code generation and UI components drew a notable comment about "the future is now." Other topics include skepticism over Apple's potential to patch these workarounds and requests for clarification on which features are lost when using AirPods outside of Apple's ecosystem.
HN discussion
(196 points, 65 comments)
The article details a Medium platform bug preventing Polish users from typing the letter "Ś". The issue stemmed from the collision of four historical factors: Polish language requirements adding diacritics like "Ś", Communist-era PC shortages forcing non-standard keyboard layouts where "Ś" is typed with Right Alt+S (internally mapped to Ctrl+Alt+S by Windows), ingrained Ctrl+S muscle memory for saving, and Medium's code blindly blocking all Ctrl+Key combinations to prevent browser save dialogs. This inadvertently blocked Ctrl+Alt+S, thus preventing "Ś" input. The fix involved modifying the save shortcut handler to only block Ctrl+S when Alt was not pressed.
HN comments highlighted the article's fascinating blend of technical detail and cultural/historical context, appreciating how it exposed systemic English-centric design assumptions in tech. Many shared anecdotes about similar keyboard struggles, particularly with diacritics interfering with software (e.g., Copilot, Edge). Technical critiques focused on browser limitations for handling complex key combinations, with suggestions for standardized properties like "CTRL+ALT+S" to prevent such bugs. Commenters also debated Polish cultural positioning and noted ongoing issues with modern tools, while some shared alternative keyboard layout solutions to ergonomic problems.
HN discussion
(93 points, 114 comments)
The article argues that "tokenmaxxing"—the practice of incentivizing or requiring employees to spend heavily on AI tokens—is not dead but evolving. Initially, companies like Meta used token spend metrics as a blunt-force strategy to force AI adoption among resistant employees, despite encouraging wasteful token usage on useless tasks. This approach succeeded in integrating AI tools like Claude Code, but rising API costs and limited subsidies led companies to roll back unlimited policies. The author contends tokenmaxxing will return due to emerging concepts like "compounding correctness," where increased token spend reliably improves outcomes (e.g., in cybersecurity applications like Anthropic's Mythos model). The article distinguishes between productive token use (e.g., developers using loops for unsupervised coding tasks) and wasteful pipeline-based agent use, predicting a future "software factory" model where open, cost-effective platforms drive renewed token spend.
Hacker News commenters debated tokenmaxxing’s purpose and impact. Many agreed with the article’s view that tokenmaxxing was a temporary, strategic tool to overcome employee resistance to AI adoption (aurareturn), likening it to other tech adoption phases like big data (kaizenite). Others criticized leadership’s spending priorities, citing examples of wasteful expenses like consultants and rebranding over server upkeep (linsomniac). Skepticism about tokenmaxxing’s prevalence emerged (behnamoh), alongside concerns about long-term sustainability (j45) and reduced human oversight in AI-driven systems (fraywing). Commenters also analyzed broader tech hype cycles (gausswho), clarified concepts like "loops" (theanonymousone), and noted the irony of management promoting experimentation while demanding ROI (IceHegel). Historical parallels to cloud gaming and blockchain (gausswho) framed tokenmaxxing as part of a recurring pattern of overhyped tech adoption.
Generated with hn-summaries