Jun 16Wednesday, June 17, 2026 · all days
1.GrapheneOS has been ported to Android 17(discuss.grapheneos.org)
926 points by Cider9986 19 hours ago | 486 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Satisfied long-term GrapheneOS users praising the privacy and computer-like experience
  • ~Users report minor friction or missing features but remain happy overall
  • ~Concerns about compatibility with banking, payment, and region-specific apps
  • ~Frustration that GrapheneOS is limited to Pixel hardware, which isn't globally available
  • Newcomers asking practical questions about hardware choices and alternatives
2.Running local models is good now(vickiboykis.com)
1461 points by jfb 1 day ago | 559 comments | permalink
tl;dr: Local LLMs have crossed a usability threshold, with GPT-OSS and Google's Gemma models now enabling agentic coding workflows on a 64GB M2 Mac at roughly 75% the accuracy of frontier models. The author runs Gemma-4-12b-qat via LM Studio with the Pi agent harness in a Docker sandbox, successfully using it for refactoring, unit tests, and bootstrapping repos. Limitations remain—slow inference, small context windows, and occasional prompt template issues—but the tooling ecosystem (LM Studio, llama.cpp, HuggingFace) has matured enough to make local agentic development genuinely viable.
HN Discussion:
  • Local models still have significant usability issues with speed, memory, and quantization tradeoffs
  • Local models are still far behind frontier models for serious agentic coding work
  • Local models offer a genuinely preferable experience over frontier models for some users
  • Growing local model viability threatens the pricing power of cloud LLM providers
  • Cloud-based workflows remain preferable due to hardware constraints and convenience
3.Humiliating IIS servers for fun and jail time(mll.sh)
328 points by denysvitali 17 hours ago | 78 comments | permalink
tl;dr: Misconfigured IIS servers remain a goldmine for bug bounty hunters, with attack vectors including internal IP disclosure via HTTP/1.0 requests, virtual host brute-forcing past HTTPAPI 2.0 404s, and IIS tilde (8.3 shortname) enumeration that can be resolved using LLMs, GitHub dorks, or BigQuery's public GitHub dataset. High-value targets include web.config (containing machine keys for ViewState deserialization RCE), bin directory DLLs accessible via cookieless session path tricks like `/(S(X))/b/(S(X))in/`, and trace.axd/elmah.axd debug endpoints. Other techniques cover reverse proxy path confusion, NTFS alternate data stream auth bypasses, file upload extensions that render as HTML, and HPP for WAF evasion.
HN Discussion:
  • Praise for the article's tone, content quality, and design
  • Adding technical context about Windows 8.3 filename behavior and IIS defaults
  • Sharing complementary defensive tactic of using IIS landing pages as honeypot bait
  • Questioning the relevance of IIS given its perceived declining usage
  • ~Critiquing the webpage formatting/sidebar overlap issues
4.TIL: You can make HTTP requests without curl using Bash /dev/TCP(mareksuppa.com)
501 points by mrshu 23 hours ago | 215 comments | permalink
tl;dr: Bash can open raw TCP sockets via `/dev/tcp/host/port`, letting you make basic HTTP requests with just `exec 3<>/dev/tcp/...` and `printf` — useful when you're stuck in a stripped-down container with no curl or wget. Caveats: it's not a real HTTP client (no TLS, redirects, chunked encoding, etc.), requires `Connection: close` to avoid hangs, is bash-only (not POSIX), and depends on bash being compiled with `--enable-net-redirections`.
HN Discussion:
  • Nostalgic recollection of manually speaking protocols like HTTP/SMTP via telnet
  • Plan 9 offered this network-as-filesystem concept more elegantly and broadly
  • Confirms the technique works with practical demonstration against example.com
  • Article overstates the case; bash doesn't speak HTTP, and this shouldn't be used in production
  • Shares real-world use cases (container health checks, initramfs) validating the technique's usefulness
5.Calvin and Hobbes and the price of integrity(therepublicofletters.substack.com)
511 points by pseudolus 1 day ago | 219 comments | permalink
tl;dr: Bill Watterson spent six years fighting his syndicate to prevent Calvin and Hobbes from being licensed into merchandise, plush toys, or animated adaptations, viewing artistic integrity as inseparable from craft. He won that battle, secured two unprecedented sabbaticals, and forced newspapers to accept a non-standard Sunday strip format—but the cumulative creative toll led him to end the strip in 1995 after just ten years. Watterson has since maintained near-total silence about the work, giving only a handful of interviews and refusing to engage with fans about the comic.
HN Discussion:
  • Admires Watterson's integrity and dedication to artistic purity over commercial gain
  • Personal reflection on Calvin and Hobbes' profound influence on their life and identity
  • Disagrees with Watterson's stance, arguing refusal to merchandise causes the work to fade from cultural memory
  • ~Sympathizes with selling out, viewing commercialization as a reasonable choice most would make
  • Shares supplementary resources and historical context about Watterson and the strip
6.Has AI already killed self-help nonfiction books?(tim.blog)
377 points by imakwana 22 hours ago | 428 comments | permalink
tl;dr: Tim Ferriss shares BookScan data showing his catalog of bestselling self-help books (4-Hour Workweek, etc.) declined 5% in 2023, 13% in 2024, 46% in 2025, and is tracking -57% in 2026—closely correlating with the rise of ChatGPT and LLMs. He argues prescriptive nonfiction is the canary in the coal mine, since "how-to" content is essentially a lookup table that chatbots now deliver faster, cheaper, and personalized for free—threatening YouTube tutorials, podcasts, courses, and journalism next. His bet: information businesses collapse into chatbots, while transformation, storytelling, voice, and personality remain the only durable moats.
HN Discussion:
  • Self-help books are bloated filler that LLMs distill better, supporting the article's thesis
  • ~Self-help is dying due to its own grifty mafia-like nature, not primarily AI
  • Self-help is a poor canary for other genres since it lacks artistic value
  • The decline has multiple causes like economic factors, not just AI
  • ~Market fragmentation via YouTube/TikTok/podcasts is splintering the audience, with AI as accelerant
7.GPT‑NL: a sovereign language model for the Netherlands(tno.nl)
244 points by root-parent 22 hours ago | 277 comments | permalink
tl;dr: The Netherlands is building GPT-NL, a sovereign Dutch language model trained from scratch to avoid copyright, privacy, and data provenance issues inherited from existing models, with source code released as open source and weights under a controlled license. The project, backed by €13.5 million in public funding from the Ministry of Economic Affairs, includes a Content Board giving data providers a say and revenue share, emphasizing transparency, lawful data sourcing, and energy efficiency.
HN Discussion:
  • Building sovereign models from scratch wastes money; better to fine-tune existing open baselines
  • Europe needs its own sovereign, open-source model trained on local languages and renewable energy
  • Supporting national/European AI independence from US and China dominance is valuable
  • Countries should focus on controlling compute infrastructure rather than building their own models
  • Skepticism that €13.5M is sufficient to build a competitive model with fair compensation
8.Stop Using JWTs(gist.github.com)
457 points by dzonga 23 hours ago | 266 comments | permalink
tl;dr: JWTs are unsuitable for user sessions: the spec is designed for short-lived (~5 min) tokens, "stateless" auth isn't practically secure, and the specification itself has known flaws distrusted by security experts. Regular cookie-based sessions are simpler, more flexible, and purpose-built for keeping users logged in. If you genuinely need signed short-lived tokens, use PASETO instead—and never store auth credentials in localStorage.
HN Discussion:
  • ~JWTs are fine for service-to-service auth but shouldn't be used for browser sessions
  • JWTs can be made secure with short lifetimes, refresh tokens, and revocation nonces
  • The article is sensationalist and lacks evidence; no real-world attacks demonstrate JWTs are broken
  • Cookie-stored JWTs are essentially equivalent to session cookies, so the distinction is overblown
  • Agreement that JWTs are the wrong tool for web auth, suitable only for service-to-service
9.Stop Killing Games fails to secure EU law despite 1.3M signatures(dexerto.com)
321 points by slymax 14 hours ago | 243 comments | permalink
tl;dr: The European Commission declined to propose legislation requiring publishers to keep discontinued games playable, despite the Stop Killing Games initiative gathering 1.3M verified signatures. The Commission called such a mandate "not proportionate," citing IP, cost, and cybersecurity concerns, and will instead pursue a voluntary industry code of conduct starting in 2026. Campaigners say they'll now push to amend their proposals into the Digital Fairness Act via the European Parliament.
HN Discussion:
  • Headline mischaracterizes the outcome; ECI only requires response, and SKG anticipated this and shifted venue to Parliament
  • Voluntary code is disappointing; game ownership is becoming meaningless rental
  • ~SKG mishandled strategy by not engaging industry lobbyists and seeking compromise first
  • Industry lobbyists captured the EU process, contradicting copyright justifications given older games still work
  • Skepticism about regulation's effects, such as potentially reducing online games or revealing EU regulatory inconsistency
10.SpaceX to buy Cursor for $60B(reuters.com)
1090 points by itsmarcelg 1 day ago | 1592 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Cursor is inferior to Codex/Claude Code and users have moved on
  • The acquisition valuation is absurd and makes no sense
  • SpaceX pivoting to AI/IDE acquisition is bizarre and concerning
  • Concerns about IP theft via data access for training competitor models
  • Cursor still has unique strengths like autocomplete and plan mode
11.But yak shaving is fun (2019)(parksb.github.io)
287 points by parksb 1 day ago | 88 comments | permalink
tl;dr: The author recounts building their own static site generator from scratch instead of using Jekyll or Hugo, illustrating "yak shaving"—a term coined by MIT's Carlin Vieri after watching a Ren & Stimpy episode—where chained sub-tasks derail the original goal. While yak shaving usually wastes limited project budgets, it's intrinsically fun and occasionally pays off spectacularly, as when Donald Knuth created TeX, METAFONT, Computer Modern, and the literate programming paradigm just to typeset a book. The author argues it's especially valuable for learning, since the detours themselves teach you something even if you never finish.
HN Discussion:
  • ~Yak shaving is fun but can derail projects and waste limited budgets, especially for indie developers
  • Personal anecdotes of decades-long yak shaving journeys that never reached the original goal but were enjoyable
  • AI tools have reduced the costs of yak shaving, making custom tooling more valuable for learning and performance
  • Shaming yak shaving constrains engineer creativity and breadth in an over-complex framework landscape
  • Yak shaving often happens because existing alternatives are genuinely unsatisfactory, justifying the detour
12.Mechanical Watch (2022)(ciechanow.ski)
715 points by razin 1 day ago | 122 comments | permalink
tl;dr: An interactive, deeply illustrated walkthrough of how a mechanical watch movement works, breaking it down into seven core timekeeping components (mainspring, gear train, escapement, balance) plus complications like date display, keyless works, and automatic winding. Each part is explained with draggable 3D animations showing how energy flows from the wound mainspring through the escapement and balance wheel to drive the hands at precise rates, ending with a scale demonstration showing the entire mechanism fits within a credit card's footprint.
HN Discussion:
  • Article inspired hands-on real-world projects related to watch movements
  • Praise for the article's educational clarity and ability to explain complex topics simply
  • Admiration for the author's clean, vanilla code and universal browser compatibility
  • Personal reflections on fascination with mechanical timepieces and watch ownership
  • Recommendations of complementary resources like books and older demonstration videos
13.Apple's weird anti-nausea dots cured my car sickness(theverge.com)
834 points by neilfrndes 23 hours ago | 248 comments | permalink
tl;dr: Apple's Vehicle Motion Cues, an accessibility feature introduced in 2024 for iOS, iPadOS, and macOS, displays moving dots around the screen's periphery that shift in sync with a vehicle's acceleration, braking, and turning to combat motion sickness. The author reports it effectively eliminated nausea while reading and writing in a moving car. It can be toggled via Accessibility settings, and mapped to a Back Tap gesture on iPhone for quick activation.
HN Discussion:
  • Excited to try the feature after suffering from lifelong motion sickness
  • Feature works remarkably well to combat car sickness
  • Tried the feature but it failed to prevent nausea
  • Sharing alternative remedies like Android apps, special glasses, or mental tricks
  • ~Motion sickness is a serious overlooked problem needing better solutions than this implementation
14.Is Meta destroying its engineering organization?(newsletter.pragmaticengineer.com)
615 points by throwarayes 23 hours ago | 568 comments | permalink
tl;dr: Meta's engineering culture has rapidly deteriorated since April 2025 after leadership forcibly reassigned 30-50% of engineers on core teams (~6,500 people) to data labeling and RLHF work for AI training, mandated keystroke/mouse tracking, and began measuring AI token usage in performance reviews. The changes—driven by Mark Zuckerberg and Scale AI's Alexandr Wang—gutted infrastructure and security teams, leading to a major Instagram account takeover exploit, the CISO's resignation, and a mass exodus of tenured engineers. The author argues this reflects a broader "AI psychosis" among tech leaders who are sacrificing engineering quality and culture in the rush to chase AI.
HN Discussion:
  • ~Meta's homegrown engineering culture was already bad; acquired orgs masked the dysfunction
  • AI psychosis is spreading industry-wide beyond just Meta, validating the broader concern
  • Skeptical of the article's specific claims, like 30-50% reassignment to data labeling
  • Blame falls primarily on Alexandr Wang for poaching top engineers for labeling work
  • ~Zuckerberg's drastic moves are a rational gamble against social media's decline
15.A backdoor in a LinkedIn job offer(roman.pt)
1573 points by lwhsiao 1 day ago | 301 comments | permalink
tl;dr: A developer received a LinkedIn message from a fake crypto startup "recruiter" asking him to review a GitHub repo and check a "deprecated Node modules issue" — bait to trigger `npm install`, which auto-runs a `prepare` script that executes a backdoor disguised as a test file, fetching and running arbitrary code from a remote server. Both the recruiter's LinkedIn profile and the repo's commit author identity were stolen from real people. He flagged the threat using a read-only AI agent on a throwaway VPS, which spotted the payload in seconds.
HN Discussion:
  • Shares similar scam experiences confirming this attack pattern is widespread
  • Calls for organized societal/law enforcement response to cybercrime
  • Criticizes LinkedIn and GitHub (Microsoft) for inaction on reported threats
  • Notes the attack mimics legitimate interview tasks, making it dangerously deceptive
  • Suspects the writeup was largely AI-generated, casting doubt on authenticity
16.I admire Fabrice Bellard. He is almost certainly a better overall programmer(twitter.com)
905 points by apitman 1 day ago | 443 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Bellard's genius lies in choosing impactful problems, not just technical skill
  • ~Bellard's strength is translating complex specs into efficient C implementations
  • Bellard's legacy in FFmpeg is overstated; his original code was poor and long replaced
  • Admiration for Bellard's reclusive, focused work ethic and avoidance of limelight
  • The article's claim that FFmpeg is the Internet's invisible engine is hyperbole
17.Apple is about to make Hide My Email useless(arseniyshestakov.com)
504 points by SXX 21 hours ago | 315 comments | permalink
tl;dr: Apple is moving Sign in with Apple and Hide My Email aliases to a new `@private.icloud.com` subdomain, making it trivial for services to block all relay addresses without affecting regular iCloud mail users. This undermines the privacy value of Hide My Email, since previously the aliases were harder to distinguish from real iCloud accounts. Users wanting to preserve existing `@icloud.com` aliases should generate them before the change takes effect (rate-limited to ~30/hour).
HN Discussion:
  • Sites that block privacy emails aren't worth using anyway, undermining the article's concern
  • Hide My Email is becoming useless or impractical due to Apple's change
  • Use your own domain with catch-all forwarding or third-party aliases instead of Apple
  • ~Hide My Email already had fundamental design flaws beyond this change
  • ~Determined sites could already detect aliases via patterns, so impact is limited
18.Iroh 1.0(iroh.computer)
1365 points by chadfowler 2 days ago | 446 comments | permalink
tl;dr: Iroh 1.0 is the first stable release of a networking library that replaces IP-based addressing with public keys, enabling secure, direct device-to-device connections that persist across network changes and NAT boundaries. The release locks in wire protocol and API stability, adds official bindings for Python, Node.js, Swift, and Kotlin alongside Rust, and includes QUIC multipath, NAT traversal, WASM support, and pluggable transports like BLE and Tor. The project's public relays have handled 200M+ endpoints in the last 30 days, with ~95% of connection traffic typically flowing directly peer-to-peer.
HN Discussion:
  • Iroh is best understood as Tailscale at the application layer for app developers
  • Production users praise Iroh's reliability, team responsiveness, and developer experience
  • Decentralized peer-to-peer networking is the future and Iroh advances that vision
  • The announcement lacks clarity on fundamentals like what 'keys' mean and how relays work
  • Skepticism that Iroh solves a real problem given IP, DNS, IPv6, and QUIC already exist
19.The time the x86 emulator team found code so bad they fixed it during emulation(devblogs.microsoft.com)
492 points by paulmooreparks 1 day ago | 168 comments | permalink
tl;dr: A Windows x86-32 binary translation emulator team encountered a program that allocated 64KB on the stack and initialized it by unrolling the init loop into 65,536 individual byte-write instructions—256KB of code to zero 64KB of data. Rather than faithfully translating this monstrosity, the team added special-case detection to recognize the pattern and replace it with an equivalent tight loop during translation.
HN Discussion:
  • Sharing similar war stories of inefficient code discovered while working on tooling or emulation layers
  • Compatibility/translation layers fixing broken software is a recurring and valuable pattern
  • Questioning the article's framing—loop unrolling may actually be a legitimate optimization choice
  • ~Correcting or nuancing the article's claim about the 'standard way' of stack allocation
  • Broader lamentation about pervasive unoptimized code in modern software
20.Feds freaked over Fable 5 after 'fix this code', not jailbreak, say researchers(theregister.com)
583 points by _tk_ 1 day ago | 349 comments | permalink
tl;dr: The Trump administration's export controls on Anthropic's Fable 5 and Mythos 5 models were reportedly triggered by a research paper describing a "jailbreak" that amounted to prompting the AI with "fix this code" on vulnerable code samples, according to Luta Security's Katie Moussouris, who claims to be the only outside expert to read it. Moussouris and 100+ cybersecurity leaders signed an open letter urging reversal, arguing the models were simply performing standard defensive security work (find, fix, test) and that the ban hurts defenders while adversaries' models catch up anyway.
HN Discussion:
  • The 'fix this code' jailbreak is trivial yet fundamentally unfixable without crippling the model
  • The export controls are politically motivated retaliation rather than genuine security concerns
  • Anthropic's safety strategy is internally contradictory and untenable given how LLMs work
  • ~There's an inherent unresolvable tension between defensive and offensive security uses of LLMs
  • The article omits important context about manual steps involved beyond just 'fix this code'