Top 8 Hacker News posts, summarized
HN discussion
(541 points, 577 comments)
Anthropic has restricted its consumer product Claude to users aged 18 and older, requiring age confirmation during account setup. When systems detect signals indicating a user may be under 18, the account is disabled and the user is prompted to verify their age through Yoti, a third-party age verification provider. Yoti offers three verification methods: facial age estimation via selfie, ID document upload (passport, driver's license, national ID), or sharing a verified "over 18" attribute from the Yoti Digital ID app. Anthropic states it never receives or stores personal identification data; Yoti deletes all images and personal data immediately after verification and provides only a pass/fail result. Yoti is independently audited for SOC2 compliance.
Commenters expressed skepticism and frustration about the policy, with several questioning the rationale and noting the switch from Persona to Yoti without explanation. Some users threatened to abandon Claude if subjected to verification, while others reported the restriction has been active for months in regions like Germany, citing instances of minors being locked out during legitimate educational use. Speculation arose about legal motivations (minors cannot enter binding contracts) and whether this represents a step toward KYC implementation. Critics argued the move harms young creators and trivializes larger AI safety concerns, while one commenter proposed OS-level age flags as a less invasive alternative to ID verification. A few dismissed the policy as performative given widespread circumvention on other platforms.
HN discussion
(496 points, 564 comments)
An open letter signed by 25 Fields Medalists argues that the rapid advancement of LLMs in solving major mathematical problems reveals a fundamental misalignment between AI companies and the mathematical community. While AI systems can now produce solutions to longstanding open problems, mathematics as a discipline values conceptual understanding, insight, and the slow human process of developing, transmitting, and integrating ideas — not merely generating true/false answers. The authors warn that rushed AI solutions bypass proper writeups, attribution, and community vetting, severing the transmission chain that makes mathematical knowledge "alive" and usable. This mirrors a broader crisis in intellectual work where AI produces outputs without the understanding traditionally forged through years of training. The signatories call for urgent action by mathematicians, AI companies, and society to ensure AI enhances rather than destroys the genuine practice of mathematics and other creative professions.
Commenters largely validated the letter's diagnosis but questioned its remedy. Several noted the core tension: AI has destroyed the traditional yardstick (solving open problems) for measuring mathematical contribution, likely forcing the field toward either secrecy or more subjective credit assignment. Others reframed the misalignment as structural — between mathematicians and the funders (governments, corporations) who want practical results without the social process of mathematics. The analogy to open-source exploitation was raised: mathematicians' ideas now feed "Plagiarism as a Service" companies, with Fields Medals offering no protection. A recurring theme was the absence of a concrete ask — what specific mechanism should replace the current benchmark-driven approach? Some argued the problem is societal ("us problem"), not technical, while others warned of "mode collapse" in human thought across fields, where cognitive handoff to AI may become irreversible before its consequences are understood.
HN discussion
(233 points, 221 comments)
The article argues that while LLMs generate formally correct code, they produce "sloppy" code characterized by unnecessary abstractions, duplication, and exponential LOC growth, eroding human agency. The author, from Earendil, explores quantitative metrics for sloppiness: simple LOC delta, verbosity (AST-based heuristic for redundant lines), and erosion (fraction of code mass in functions with cyclomatic complexity >10). SlopCodeBench evaluates agents iteratively with context erasure between checkpoints, mimicking real usage; even state-of-the-art models achieve a 0% strict solve rate, showing bad decisions accumulate. AI-as-judge evaluations are unreliable (random or inconsistent), and human review doesn't scale. The author concludes that human intuition remains essential and invites collaboration on further metrics like coupling and cohesion.
Commenters broadly validate the problem but debate framing and solutions. Several note that "coding solved" is premature—correctness is a baseline, not quality, and maintainability, design, and understanding remain gaps. The iterative benchmark approach is praised for capturing real-world accumulation of slop. Skepticism appears around cyclomatic complexity as a meaningful metric and whether optimizing LOC delta would incentivize code-golfing. Some argue LLMs merely replicate existing human slop from training data, while others suggest the era of sloppy AI code is a temporary reward-function gap. A few emphasize that architecture and design require human mental models, and that velocity to understanding may outweigh velocity to solution.
HN discussion
(186 points, 89 comments)
The author ran a Google Ads campaign for their puzzle app Dayzle at CA$40/day with an install goal. After removing a target cost-per-install, Google immediately spent double the budget and reported 21 installs, but the author's analytics showed only 1 real install. Investigation revealed 20 installs came from obsolete app versions no longer on the Play Store, across 28 phone models in 19 states, each opening the app once for zero seconds and never returning. Over two weeks: 56 billed installs, 33 with bot patterns, 7 from untargeted countries, and 13 genuine users who completed 92 games. The bots watched the shortest video ads, then installed from saved APK files rather than the Play Store. Since Google counts a view-then-install as a conversion, the algorithm optimized toward the bot farm, creating a feedback loop that wasted the ad spend. The author switched the conversion goal to "won a puzzle" to increase farming difficulty and filed an invalid-traffic refund request.
Commenters broadly validate the author's experience, describing ad fraud as systemic and worsening over years. Multiple users report bots evolving from simple clicks to installing apps, running them, and mimicking onboarding flows, with distributed device farms making detection difficult. One advertiser maintains 4,000+ IP network exclusions to block data-center traffic. Incentives for bot farms are debated: inflating install metrics for VC fundraising, click-fraud ecosystems, or Google's own misaligned incentives (profiting from fraudulent conversions). Several commenters label Google/Meta ads "a con" and note the platform has little motivation to fix fraud. A striking example: a developer paid Google Ads for traffic, then had their AdMob account banned for the resulting invalid traffic. The consensus is that small budgets are especially vulnerable, and advertisers must calculate customer acquisition costs factoring in bot traffic rather than trusting platform-reported metrics.
HN discussion
(142 points, 117 comments)
The article catalogs a range of bodily oddities—variations in perception, anatomy, reflexes, and involuntary sensations—with prevalence estimates for each. It covers aphantasia (absent or faint mental imagery, affecting ~4% of people, with ~1% experiencing total absence) and its opposite, hyperphantasia; camptodactyly, a congenital finger flexion contracture (<1%); Linburg-Comstock syndrome, an extra forearm tendon linking thumb and index finger (~20%, though studies range 5–60%); the oculocardiac reflex, where eye pressure slows the heart (universal but variable strength); proctalgia fugax, brief severe rectal cramping (~7%, range 4–18%); and ASMR, a pleasant tingling triggered by auditory or tactile stimuli (estimated 20–70%). Each entry notes prevalence ranges and clinical significance, emphasizing that most are benign variations.
Commenters shared personal experiences matching listed oddities (e.g., falling sensations in feet, photic sneeze reflex, sleep paralysis, voluntary Eustachian tube opening, synesthesia) and suggested additions like exploding head syndrome, BPPV (canalith repositioning), Haidinger’s brush (polarization detection), and "piss shivers." Several questioned prevalence classifications (e.g., supertasting at 25% labeled rare, brain freeze estimates). A recurring theme was skepticism about the site’s sourcing and production: multiple users suspected LLM-generated content and AI-assisted design, criticized reliance on Wikipedia as a sole citation, and noted inaccurate images undermining trustworthiness. Others praised the site as delightful, educational, and worthy of school curricula, while one humorously proposed "autoheadslap" from Nvidia stock regret.
HN discussion
(157 points, 89 comments)
GrapheneOS has released version 13 of its Messages app, a complete rewrite using Jetpack Compose and Material 3 that replaces the legacy interface. The update rebuilds every screen with a two-pane conversation layout for large screens, new onboarding covering SMS privacy and permissions, and correct display cutout handling. Major new features include conversation pinning, snoozing notifications (1, 8, 24 hours, or indefinite), marking as unread, swipe-to-archive, quick actions from avatars, and immediate application of pinning, archiving, and deletion. Conversations gain rebuilt message bubbles, full-screen message details with copyable metadata, SMS segment/character counters, MMS subject editing, blocked-sender banners, and a redesigned recipient picker with alphabetical sections. Attachments and media see a rebuilt picker with photo/video capture, flash controls, and Android's embedded picker; redesigned audio recording with slide-to-cancel; attachment captions; and rewritten photo and vCard viewers. Sharing and forwarding use a new picker with search, recent conversations, alphabetical contacts, and multi-select. Settings are rewritten with a new Privacy section. Privacy and security improvements include opt-in YouTube link previews (disabled by default), shared-content validation rejecting file URIs and private app files, non-exported widget receivers, immutable pending intents, allocation limits on EXIF/MMS parsing, and numerous crash fixes. Message handling fixes cover immediate SMS import, correct failed-message notifications, inline replies not opening home screen, unread notifications surviving reboots, and race conditions between MMS downloads and notifications. Multi-user and work-profile support improves, accessibility gains screen-reader labels and announcements, and testing expands with CI integration. Platform requirements update to minSdk 36, targetSdk 37, with new dependencies including Compose BOM, Material 3 Adaptive, Navigation 3, Coil 3, and CameraX.
Commenters noted the rapid release following a recent announcement and requested screenshots of the redesigned interface. A key question asked about RCS support, while another user inquired whether the app is installable now or requires the next OS release. A regression report highlighted that long-pressing a link now selects the entire SMS instead of opening a link menu, breaking one-time code copying; the issue was tagged as a feature rather than a regression, raising concerns about testing thoroughness. Broader discussion questioned GrapheneOS's prioritization of a messaging app given its core OS security focus, with theories ranging from platform completeness to donor influence. Other threads criticized the call app's UX, expressed dislike for Material 3/Material You's increased whitespace, and noted regional irrelevance of SMS apps where WhatsApp/Signal dominate. Hardware compatibility desires included Fairphone support and frustration over Pixel-exclusive availability. Feature requests included encrypted SMS (like SMSecure) and a full system backup replacement for Seedvault. One user reported a past bug where contact names mismatched actual senders/recipients for months.
HN discussion
(140 points, 68 comments)
The author benchmarked RTK (Rust Token Killer), a tool that compresses terminal output for AI coding agents, across 1,740 attempts on Terminal-Bench 2.1 using Claude Code (Fable 5.0) and OpenCode (DeepSeek V4 Pro). With RTK, Fable costs fell 5% and DeepSeek costs rose 5%; pass rates dropped slightly for both. When weighting tasks equally, Fable was 1% more expensive and DeepSeek 17% more expensive with RTK. Nearly all of Fable's savings came from a single task (winning-avg-corewars), while DeepSeek showed the reverse pattern on that same task. The `rtk gain` metric—which reports raw output bytes removed divided by 4—proved misleading: in one task it credited 120.5M tokens saved per `head -1` call by comparing against the full file, though the command would never return it. RTK also introduced bugs (a `find` loop caused 339 errors and 9× cost) and only processes shell commands, missing ~50% of terminal interactions handled by `Read`/`Grep`/`Glob` tools. Terminal output constituted just 11% of Fable's and 40% of DeepSeek's input tokens, and caching makes repeated reads cheap (1/10 and 1/30 cost). DeepSeek's RTK attempts took 18% more turns on average, erasing per-turn savings. The author concludes RTK is a niche optimization, not a general cost-saver, and does not recommend it.
Commenters broadly validated the findings, characterizing RTK and similar "token-saving hacks" as snake oil. Multiple users noted that LLMs are trained on standard shell output; altering it confuses models and increases tool calls. The `rtk gain` metric was criticized as fundamentally broken—it ignores piping (e.g., `rtk cmd | tail -5` shows 5 lines but RTK claims full output savings) and persists state across sandbox boundaries. GodelNumbering demonstrated RTK producing wrong `grep` results (260 vs 966 lines) with 60× CPU overhead. Several commenters argued the "bitter lesson" applies: models already optimize via `head`/`tail` and improve at this over time. Alternative approaches cited include local code embeddings for semantic search (aeneas_ory's lumen), subagents for summarization, and project-specific skills over generic ones. JetBrains' independent SkillsBench benchmark was referenced as confirming no savings. Consensus: independent, expensive benchmarks are the only reliable signal, and most popular tools fail them.
HN discussion
(146 points, 48 comments)
The Global Glacier Extinction Explorer presents a study shifting focus from aggregate glacier mass loss to the fate of individual glaciers worldwide, aligned with the UN International Year of Glacier Preservation (2025). A glacier is classified as extinct when its area falls below 0.01 km² or its volume drops below 1% of its initial value. Projections derive from three global glacier models (GloGEM, OGGM, PyGEM) driven by multiple General Circulation Models across warming scenarios of +1.5°C, +2.0°C, +2.7°C, and +4.0°C, yielding median extinction years with 25th–75th percentile ranges. The analysis uses Randolph Glacier Inventory v6.0 outlines and Farinotti et al. (2019) ice thickness estimates, without accounting for glacier fragmentation during retreat. Results are visualized via an interactive web map built on open-source tools, with data publicly available on Zenodo.
Commenters identified potential data inaccuracies: one user found the mapped perimeter for Glacier Blanc (Massif des Écrins) incorrect by hundreds of meters, questioning whether the error is visual or propagates into predictions. Another noted Okjökull (Iceland) is already considered extinct but the tool projects survival until the 2070s, with the 1.5°C scenario showing later extinction (2079) than 2.7°C (2075). A user expressed confusion over glaciers appearing to survive longer under 4°C than 2.5°C warming. The maximum displayed category is "Survives through 2100." Positive feedback praised the map's visceral impact over traditional climate visualizations, with a feature request for interactive hover/click on the extinction timeline panel. One commenter shared a paper on elevation-dependent warming at 4,500–5,500 m — precisely where glaciers form — as a relevant mechanistic context. Others contributed personal observations of retreat in New Zealand and Alaska, while a few comments were off-topic or reported Safari compatibility issues.
Generated with hn-summaries