Aug 19Thursday, August 20, 2026 · all daysAug 21 · today »
1.Don't Paste the AI, please(dontpastetheai.com)
996 points by pjerem 1 day ago | 547 comments | permalink
tl;dr: When someone asks you a question, don't just paste raw ChatGPT output back at them—they have the same tools and could've done that themselves. They want your context, judgment, and take, so use AI as a drafting partner but edit it down, quote only what's useful, or just admit you have no strong opinion. A short reply from you beats a long one from a model.
HN Discussion:
  • Reinforces the article by sharing similar policies or experiences against pasting raw AI output
  • AI-pasted context is actually better than the vague underspecified messages colleagues used to send
  • Shares concrete horror stories of coworkers pasting unedited AI content creating work for others
  • Criticizes the article itself as AI-written with poor metaphors, undermining its credibility
  • ~The article's polite framing is too weak given how much people love offloading thinking to AI
2.OpenRouter is joining Stripe(openrouter.ai)
944 points by rvz 2 days ago | 479 comments | permalink
tl;dr: OpenRouter, a multi-model AI gateway processing 10+ trillion tokens daily across 400+ models, is being acquired by Stripe. The company says it will continue operating independently with the same product, mission, and roadmap, maintaining its neutrality across model providers. The deal is expected to close in the coming weeks, pending customary closing conditions.
HN Discussion:
  • Long-time users praise OpenRouter's product and hope Stripe will be a good steward
  • Corporate consolidation via acquisition is bad for consumers
  • Strategic fit is strong: Stripe can build accounting/billing infrastructure for metered AI work
  • Concerns about provider integrity and model authenticity remain unresolved
  • Criticism of the 'Open' branding when the product isn't actually open
3.Turns are Better than Radians (2022)(computerenhance.com)
346 points by mayoff 2 days ago | 211 comments | permalink
tl;dr: Instead of debating pi vs tau, programmers should ditch radians entirely and use "turns," where a full circle is [0,1]. Most code multiplies by tau to call sin/cos, and those trig implementations immediately divide it back out—so using turns eliminates a pointless conversion, yields exact representations of common angles (0.25, 0.5, 0.75), and simplifies code. Some libraries already offer this via half-turn variants like CUDA's sincospi.
HN Discussion:
  • Radians are mathematically special because of Euler's formula and calculus derivatives
  • Turns break clean derivatives/integrals of trig functions, making calculus messier
  • Storing angles as turns enables exact quarter-turns and clean integer representations
  • ~Choice of angle unit is application-dependent; trig is used beyond geometry (signals, etc.)
  • Nothing prevents writing your own turn-based math functions today, so debate is overblown
4.Go 1.27(go.dev)
738 points by database64128 2 days ago | 261 comments | permalink
tl;dr: Go 1.27 adds three language updates: generic methods, struct literal keys that can reference embedded/nested fields, and generalized function type inference across composite literals, conversions, and channel sends. The standard library gains `encoding/json/v2` (which now backs v1 for faster unmarshaling), post-quantum ML-DSA signatures, native UUID support, and experimental SIMD packages. Runtime improvements include size-specialized allocation cutting small-object costs up to 30%, plus a generally available goroutine leak profiler.
HN Discussion:
  • Enthusiasm for the new UUID package and anticipation of ecosystem migration away from google/uuid
  • Appreciation for the post-quantum crypto additions and the team's proactive stance
  • Generic methods and improved type inference solve real ergonomic pain points
  • ~Struct literal field selector changes are welcome but could introduce subtle bugs with overlapping embedded fields
  • SIMD packages are exciting and work well with LLM-assisted code transpilation
5.Google has stopped pushing Git tags for some Android source code(grapheneos.social)
790 points by Animux 2 days ago | 308 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Provides technical context clarifying which repositories and releases Google stopped pushing
  • Tags are a distraction; the real issue is Google restricting access to source code needed by projects like GrapheneOS
  • Google's process may violate GPL by making source access inconvenient rather than a simple URL
  • ~Calling it a GPL violation is overblown; Android was never truly community open-source
  • Broader condemnation of Google's control over Android, calling for alternatives and regulation
6.A joke domain purchase turned in geopolitical warfare(sprocketfox.io)
1006 points by kareiva 2 days ago | 168 comments | permalink
tl;dr: What started as a joke domain redirect in 2018 evolved into SondeHub, the de facto global tracking service for weather radiosondes, drawing requests from militaries, the FAA, NTSB, and insurance claimants. The author details how their reverse-prediction feature inadvertently mapped artillery sites and how, in late 2024, they discovered Russian forces were hammering their API to guide long-range strike balloons into Ukraine — forcing awkward negotiations with AWS to keep the service running while limiting misuse. Also featured: a chemtrail-believing radiosonde vendor and a correspondent whose title is "Cheese Fortune Teller."
HN Discussion:
  • Appreciation for the article's human-written quality and the non-adversarial outcome
  • Sharing personal anecdotes about weather balloon/radiosonde experiences
  • Operators of similar public infrastructure relate to receiving strange official requests
  • Skepticism toward the radiosonde vendor's evasive email response
  • Technical questions about implementation details like the data proxying pivot and wind predictions
7.Unsloth Dynamic 3.0 GGUFs(unsloth.ai)
317 points by jonesy827 2 days ago | 118 comments | permalink
tl;dr: Unsloth released Dynamic v3.0 GGUF quantizations for Qwen3-VL-27B, claiming >10% better top-1 accuracy at the same file size versus other providers, using improved imatrix calibration, smarter layer selection, and pure post-training quantization (no QAT/QAD). New features include sub-1-bit quants (UD-IQ1_S at 6.2GB retaining ~72% top-1 accuracy) and a Divergence-300 @32 benchmark for evaluating multi-token trajectory fidelity against BF16. The quants work with llama.cpp and other standard inference engines.
HN Discussion:
  • Praises Unsloth's quants but requests speed/performance improvements for local inference
  • ~Requests clearer versioning of GGUF files to distinguish Dynamic 3.0 from prior releases
  • ~Skeptical of KL-divergence metrics; wants real-world coding benchmarks to validate quality claims
  • Questions the decision to remove MTP support in the new quants
  • Excited about sub-1-bit quants and asks for additional formats (MLX, NVFP4) and comparisons
8.Unlocking a locked/deactivated e-waste Cricut Maker(sprocketfox.io)
252 points by 1e1a 2 days ago | 67 comments | permalink
tl;dr: Author salvaged a deactivated Cricut Maker from e-waste and bypassed its lockout by using an RP2040 as a USB proxy between the cutter and computer, rewriting the serial number in transit (USB CDC packets had no checksums or crypto). After replacing perished rollers and installing the RP2040 interceptor, the machine registered fine in a Cricut account — even with fabricated serials — highlighting a potential attack vector where strangers could hijack or deactivate others' serial numbers.
HN Discussion:
  • ~Cricut's ecosystem lock-in is bad; hack still leaves user vulnerable to future disabling
  • Recommends open-source alternatives like Silhouette with Inkscape plugin for cutting
  • Cricut software is terrible and buyers should avoid the product entirely
  • Praises the technical cleverness of the RP2040 USB MITM approach
  • Laments the waste of these machines being discarded cheaply
9.Feature Request: Support AGENTS.md(github.com)
358 points by fg137 2 days ago | 216 comments | permalink
tl;dr: Coding agents like Codex, Amp, and Cursor are converging on AGENTS.md as a standardized Markdown file for describing codebases to AI tools. The request is for Claude Code to support this shared format instead of relying on its proprietary CLAUDE.md, which creates friction when collaborating with developers using other agents.
HN Discussion:
  • Anthropic's refusal to adopt AGENTS.md is petty and anti-competitive, hurting their position
  • Users are working around the issue with symlinks or JS injection to force AGENTS.md support
  • ~The concept should be broadened beyond AI into a general project metadata standard like PARSEME.md
  • These agent-specific files are unnecessary; better to isolate AI work in separate branches/worktrees
  • Suspicious behavior around the GitHub issue closure suggests Anthropic is handling this poorly
10.Casio F-B100W-1A(casio.com)
452 points by __fst__ 2 days ago | 375 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Casio is missing opportunities on nostalgia products and related hardware
  • The watch's price-to-feature ratio is poor compared to alternatives like Fitbits or smartbands
  • ~Third-party open source tools and mods make the watch more interesting and usable
  • The watch design wastes display space on non-time elements, hurting readability
  • Requiring a proprietary app and Casio Account for Bluetooth is a major drawback
11.Geolocating a random island using geometry and CUDA programming(yassa9.github.io)
517 points by yassa9 2 days ago | 86 comments | permalink
tl;dr: To geolocate a resort island from a single photo, the author extracted a geometric fingerprint (triangle of three landmasses with angles and distance ratios) and brute-force searched OpenStreetMap's global coastline dataset, using CUDA to test 80.7M candidate triples in 204ms on an RTX 3050. Successive filters—open-water rectangle test, coral cay shape/compactness, Sentinel-2 NDVI vegetation check, and Copernicus DEM elevation—narrowed 141K polygons down to 26 candidates, ultimately identifying Oan Resort in Micronesia (7.36°N, 151.76°E) with the camera facing NW.
HN Discussion:
  • Praise for the writeup's quality, clarity, and human-authored style
  • Connects the technique to established terrain contour matching used in missiles and drones
  • Notes similar approaches used by JPL for Mars landing localization
  • Highlights OpenStreetMap's value for OSINT and geolocation work
  • ~Suggests additional shortcuts like sun position or visual geoguessing could simplify the problem
12.fx :Tiny, open, native coding agent.(fx.sh)
310 points by handfuloflight 3 days ago | 134 comments | permalink
tl;dr: fx is an experimental open-source (Apache-2.0) coding agent CLI written in Zig, weighing ~6MB with 10µs cold start and single-digit MB memory footprint. It's model-agnostic, works with local or cloud inference, and compiles to WebAssembly with pluggable networking. Unlike heavy terminal IDEs, it favors a Unix shell-like interface and is designed to be embeddable and extensible via skills, plugins, and MCPs.
HN Discussion:
  • The feature list and minimalist design justify the project's existence and differentiation
  • Only the Zig implementation is novel; the harness and portability claims are oversold
  • The 6MB binary size is surprisingly large for what should be a simple Zig program
  • ~Terminology is inconsistent - conflating 'agent' with 'agent harness' is confusing
  • Local inference claims are undermined by apparent Vercel account requirement
13.PostgreSQL for Everything(raphaelbauer.com)
428 points by karlmush 2 days ago | 257 comments | permalink
tl;dr: PostgreSQL can replace many specialized systems in your stack: full-text search (Solr/Elastic), document storage (MongoDB), message queues (Kafka/RabbitMQ via SELECT FOR UPDATE SKIP LOCKED), time-series data (Clickhouse via TimescaleDB), vector DB for AI (pgvector), caching (Redis via UNLOGGED tables), graph data (LTREE), and even microservices (JSON query results). The argument: it's rock-solid, easy to run everywhere, and consolidating on Postgres dramatically simplifies your infrastructure—start there and only adopt specialized tools when you actually outgrow it.
HN Discussion:
  • Real-world examples like Revolut prove Postgres can handle these varied workloads at scale
  • Start with Postgres by default and only adopt specialized tools when you hit actual limits
  • Postgres is not a real replacement for specialized tools like Elastic or Kafka beyond basic use cases
  • The 'one tool for everything' framing is a hammer-nail fallacy; best-of-breed tools exist for good reasons
  • ~Agrees in principle but flags caveats around message queues, time-series, vectors, and ecosystem tooling support
14.Air Theremin – A browser theremin you play by waving at your webcam(theremin.bizibah.com)
297 points by gurov 2 days ago | 99 comments | permalink
tl;dr: Air Theremin is a browser-based instrument with three control modes: webcam hand-tracking (hand distance for volume, height for pitch, palms together to mute), phone gyroscope (tilt for volume/pitch), or mouse fallback. Phones auto-calibrate to your grip on start.
HN Discussion:
  • Appreciates the app's responsiveness and UX quality
  • ~Prefers physical theremins over virtual/browser-based ones
  • Raises privacy concerns about webcam access and potential data misuse
  • Notes similar projects exist, suggesting the idea is not novel
  • Suggests gesture-based controls have broader potential beyond this app
15.OpenLogi(openlogi.org)
1608 points by amatheus 3 days ago | 425 comments | permalink
tl;dr: OpenLogi is a local-first, Rust-based alternative to Logitech Options+ that talks directly to Logitech mice over HID++ to remap buttons, control DPI, and configure SmartShift, with no account or telemetry. Bindings live in a plain TOML config file, and it supports Bolt, Unifying, Lightspeed, Bluetooth, and wired connections across signed macOS, Linux, and Windows builds. Per-app profiles, Options+ import, and Flow-style cross-machine features are on the roadmap; users must quit Options+ (or Solaar on Linux) to avoid HID++ conflicts.
HN Discussion:
  • Logitech's official software is bad and needs open source alternatives
  • Pointing to existing alternatives like Solaar, BetterMouse, or BetterTouchTool
  • AI/LLM-generated website copy and vibe-coded projects undermine trust in the tool
  • AI is empowering developers to replace poor vendor software, sharing similar reverse engineering efforts
  • Hardware vendors should adopt better approaches like WebUSB or offline offerings
16.Remote workers report the highest well-being in study of 7,700 employees(colorado.edu)
623 points by downbad_ 2 days ago | 341 comments | permalink
tl;dr: A study of 7,704 healthcare employees found that fully remote workers reported the highest well-being, hybrid workers came second, and onsite workers scored lowest—with remote employees also using more positive language about teamwork and culture. Higher well-being correlated with lower turnover a year later, suggesting return-to-office mandates may hurt retention. Researchers attribute the gap to autonomy, flexibility, and the elimination of commute-related stressors.
HN Discussion:
  • ~Remote work well-being is bimodal; some thrive while others flame out from isolation
  • Study methodology is flawed—didn't control for occupation, pay, or job type
  • Eliminating commute time is a major life-changing benefit of remote work
  • Remote work enables better family life and personal freedom
  • Governments should promote remote work for broader societal benefits
17.Children's stunted lungs show recovery in ultra low emission zone(bbc.com)
409 points by dabinat 3 days ago | 403 comments | permalink
tl;dr: A five-year study of 3,400 children found that London kids' previously stunted lung growth accelerated to match those in less-polluted Luton after the 2019 introduction of the Ultra Low Emission Zone, with clinically impaired lung capacity dropping from 14% to 9%. Researchers linked the improvement to faster reductions in nitrogen dioxide exposure in London, though independent experts caution that Covid-era changes and behavioral shifts (like more walking/cycling) could also be factors. Published in Lancet Public Health, the study is being called the strongest evidence yet that clean air zones can reverse pollution damage in children.
HN Discussion:
  • Personal anecdotes confirming pollution reduction improves respiratory health
  • Broader examples from China and Japan supporting pollution reduction benefits
  • Skeptical of study methodology and confounding factors like COVID lockdowns
  • Findings are obvious and shouldn't require debate given historical precedent
  • Sharing the original study data and technical details neutrally
18.A decades-old bug in Knuth's long division (TAOCP Vol II, Algorithm 4.3.1D)(kolja.rs)
223 points by nk_kolja 8 days ago | 50 comments | permalink
tl;dr: While implementing Knuth's long division (Algorithm 4.3.1D from TAOCP Vol II), the author found a decades-old bug introduced in 1995 when Knuth switched from a saturated (x86-style) trial quotient to an unsaturated (ARM-style) one without updating the correctness bounds—step D3's two corrections can fail when q̂ = b+2, though this only manifests with odd bases (smallest case: base 3). Knuth acknowledged the bug, paid the customary hexadecimal dollar, and published the correction in the 2026 errata. The author also flagged a related outdated check (`qp == b` vs `>= b`) in LLVM's APInt.cpp that happens to work but isn't justified by the book's proofs.
HN Discussion:
  • Author of the bug find shares context and confirms Knuth's acknowledgment and errata publication
  • Awe and celebration of finding a decades-old bug in a Knuth algorithm
  • Personal anecdotes about implementing division algorithms without hitting this bug
  • Fellow Knuth-check recipients share their own experiences of finding errors in TAOCP
  • Technical clarification question about whether the bug affects MIX/MMIX implementations too
19.Moderna reports first positive Phase 3 for mRNA neoantigen therapy in melanoma(twitter.com)
589 points by heydenberk 2 days ago | 286 comments | permalink
tl;dr: Moderna and Merck's Phase 3 trial of intismeran autogene, an individualized mRNA neoantigen therapy, combined with Keytruda met its primary endpoint of recurrence-free survival in patients with resected stage IIB-IV melanoma, as well as a key secondary endpoint of distant metastasis-free survival. This is the first positive Phase 3 result for both an individualized neoantigen therapy and an mRNA-based cancer therapy.
HN Discussion:
  • Excitement about mRNA therapeutic vaccines finally delivering on long-anticipated potential
  • Personal cancer experiences underscore the significance and urgency of this breakthrough
  • Skepticism due to lack of actual Phase 3 data being presented in the announcement
  • Curiosity about whether this targeted approach can generalize to other cancer types
  • Market reaction and financial validation supports the significance of the result
20.Supersonic Trebuchet [video](youtube.com)
253 points by CharlesW 7 days ago | 112 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Praise for Tom Stanton's engineering skills and iterative prototyping approach
  • ~Technical critique questioning efficiency of energy transfer and friction losses
  • ~Suggestions for improvement via modeling to boost efficiency and optimize spool design
  • Related links and references to similar projects like SpinLaunch and walking trebuchets
  • Lighthearted interest in building one's own trebuchet