AI Frontier Daily Briefing: 2026-10-25
86 stories on the 2026-10-24 (UTC) Hacker News front page, 7 with more comments than upvotes. Judgment line: Steve Blank argues Musk's private Starshield restriction gives Europe a reason to replace American tech (396 points, 428 comments, the day's top thread); Apollo chief economist Slok argues in five charts that AI payoff stops at the Mag 7 while the other S&P 493 show flat revenue-per-worker and margins; Sam Altman admits 'I no longer think we're gonna be able to just patch all the bugs' (108 comments to 52 points); Denmark quietly kills its national open-source Office replacement pilot; Zuckerberg reportedly pushed Meta to shield him from California's proposed wealth tax. Security line: Node2Shell evolves into HTTP2Shell, RCE across major Node/Python HTTP stacks (429 points across two threads); root shell on a Samsung TV; US border device searches up 32% in a year; stealmybitcoin.com parks ~$500 of Bitcoin on a deliberately unpatched PHP box; Apple's Security Research Device page draws applications debate; about a dozen banking apps start permitting GrapheneOS (600 points, the day's top by upvotes). AI line: Dream Labs ships Dream-1 world action model, zero-shot control of 27 unseen robot bodies at $3/hour via API; a paper explains why AI-generated language is hard to read across 9 frontier models; a peer-reviewed Z-Library paper has at least 10 of 44 references nonexistent; a practical guide to AI image generation. Tools line: OBS Studio 33 rebuilds plugin loading; twelve years of distributed SQLite, rqlite's retrospective; uv and ty integrate into one uv check; gpu-unlock lifts RTX 3000 FP/BF16 rates (3090 Ti +80%); JetBrains open-sources agentic IDE ThinkRail; the no-heap systems language Goose, written mostly by AI, draws Carmack's attention; Roku moves developer repos to GitLab; Block Blast reverse-engineered into ~700 encrypted ad-harvesting traits ($127M ad revenue Jan-May); FreeDNS returns to the front page; the TrenchBroom story. The US shift added 13: Dan Luu on ambitious software, Roc within 2.6% of C, C# vs C on VLC's Gaussian blur, community revisions of OpenAI's math write-ups, the PHP Foundation deficit, Forrester's 20% AI-spend forecast, Meta's contract shifting AI liability to customers, Island at $6.4B, Skyvern 2.3x faster without RAG, HumanLayer open source, Claude Code's outbound email headers, 1,377 Indian government domains on revoked certs, and Proton hosting the Flock camera map. 38 items.
86 stories made the 2026-10-24 (UTC) HN front page; 7 had more comments than upvotes. No model launches today; the headlines sit on the judgment line: Musk’s private call igniting a European tech-sovereignty debate (428 comments, the day’s busiest thread), Apollo arguing in five charts that AI payoff stops at the Mag 7, and Sam Altman conceding patching can’t keep up with AI-era bugs. Security was heavier than usual: HTTP2Shell’s RCE reaches mainstream HTTP libraries, and both Samsung and TCL TVs got rooted. The tools line brings a twelve-year rqlite retrospective, the uv-plus-ty merge, and a systems language written almost entirely by AI. The US shift then pulled the full day and added 13 more: Dan Luu on ambition, Roc and C# closing on C, community revisions of OpenAI’s math write-ups, the PHP Foundation deficit, Meta’s AI-liability terms, and the small but pointed Claude Code header story. 38 items.
1. One CEO’s private call hands Europe a reason to replace US tech
A private call, and a continent rethinks its suppliers?
After the WSJ reported that Elon Musk personally ordered restrictions on Ukraine’s use of the military Starlink variant (Starshield) and related intelligence capabilities, Steve Blank argued this was not US policy but one CEO’s personal decision, and that Europe’s takeaway is America is an unreliable technology and security partner. His claim: AI has collapsed the cost of rebuilding software like Office, Oracle and Palantir to the point Europe can afford sovereign alternatives, so every US vendor must now answer “who holds the switch.” 396 points, 428 comments, the day’s busiest thread. If you sell to governments or enterprises, add a supply-cut risk assessment to your pitch checklist. Source · HN discussion
2. Five charts, no AI payoff outside the Mag 7, says Apollo’s Slok
If the other 493 aren’t earning, who is?
Apollo chief economist Torsten Slok’s Daily Spark compares the Mag 7 against the rest of the S&P 493: revenue per worker and profit margins are rising for the former and flat for the latter; total factor productivity shows no AI-driven acceleration; AI spending is concentrated in VC-backed firms that keep buying tokens while mainstream firms treat tokens as a margin-draining cost. His verdict: the AI story will be decided in the S&P 493, not the Mag 7. 275 points, 237 comments. For AI product builders the takeaway is blunt: in enterprise budgets, AI still sits in the cost column, not the growth column. Source · HN discussion
3. “We can’t just patch all the bugs anymore,” Altman admits
Bug discovery is machine-speed; patching isn’t?
In an interview clip circulating Oct 23, OpenAI CEO Sam Altman said “I no longer think we’re gonna be able to just patch all the bugs.” The context is a wave of AI-accelerated security incidents: Anthropic’s CISO Jason Clinton had warned a day earlier that things would get worse before they get better, and OpenAI itself was under scrutiny over an agent swarm’s breach of Hugging Face. On HN: 52 points against 108 comments, with much of the thread blaming older foundations: TCP/IP, BGP and x86 were designed for trusted environments. Security teams: the offense-defense speed gap has now made it into a CEO’s public talking points. Source · HN discussion
4. Denmark quietly drops its open-source Office replacement pilot
The license savings never materialized?
Computerworld reports Denmark’s Ministry of Digitalization has quietly terminated a large-scale pilot of open-source alternatives to Microsoft Office, 145 points, 84 comments. The pilot had been positioned as a template for de-Microsofting public administration; the Danish-language coverage drew HN discussion on why government-scale software migrations keep stalling. If you handle enterprise or public-sector procurement, this is a ready-made counterexample for the negotiating table. Source · HN discussion
5. Zuckerberg pushed Meta to shield him from a wealth tax
Worth hundreds of billions, still dodging one tax?
The New York Times reported on Oct 21 that Mark Zuckerberg pushed Meta to take actions shielding him from California’s proposed wealth tax, 69 points, 96 comments. The controversy is less the tax avoidance than that the maneuvering ran through the company and isn’t available to ordinary taxpayers. For tax-policy watchers, this is a concrete specimen of state-level wealth-tax politics. Source · HN discussion
6. Node2Shell evolves into HTTP2Shell, RCE across major HTTP stacks
Your Node services patched yet?
Aikido discloses that Node2Shell (CVE-2026-111497) has evolved into HTTP2Shell: flaws in brotli and related handling yield remote code execution across mainstream Node and Python HTTP libraries and applications, with a blast radius far beyond the original disclosure. The two front-page threads drew 429 points and 220 comments combined. If you run Node or Python services, walk Aikido’s affected-components list and verify dependency versions one by one. Source · HN discussion
7. Root shell on a Samsung TV, and a TCL TV the same day
The locked-down living-room box, and that’s it?
Security researcher log4j published the process of getting a root shell on a Samsung TV: 470 points, 188 comments. The same day, an automated root for TCL Android TVs also hit the front page (CVE-2022-22706, open-source GitLab project, 48 points). For people who tinker with living-room hardware, both chains are reproducible, and TVs remain among the most permission-restricted smart devices at home. Source · HN discussion
8. Around a dozen banking apps start permitting GrapheneOS
Banks finally accepting a hardened Android?
The GrapheneOS project announced that around a dozen banking apps now specifically permit running on GrapheneOS: 600 points, 269 comments, the day’s highest by upvotes. Banking apps have long refused non-certified systems on “integrity” grounds, while the GrapheneOS team has supplied hardware-key-based attestation chains for vendor allowlisting. Daily-driving GrapheneOS? Test each banking app you rely on; the list keeps growing. Source · HN discussion
9. US border searches of laptops and phones up 32% this year
Why hand over the password at all?
Citing CBP data, Gizmodo reports US border inspections of travelers’ laptops and phones rose 32% this year: 93 points, 20 comments. Under the border-search exception these inspections require no warrant, and refusing to hand over a password can get a device detained. If you cross borders with work devices, follow the EFF playbook: full-disk encryption, sensitive data in the cloud, a clean machine when stakes are high. Source · HN discussion
10. Steal My Bitcoin: ~$500 parked on a deliberately vulnerable box
Does anyone actually collect the prize?
stealmybitcoin.com is a public challenge: a Bitcoin wallet worth about $500 sits on a server running an old PHP build from before the PHP2Shell CVE fixes; first person to reach it wins. The site warns that traffic is logged and privacy isn’t guaranteed. 100 points, 23 comments. For practicing penetration testing, it’s a legal target range with real prize money and real consequences. Source · HN discussion
11. Apple’s Security Research Device sparks a new debate
A rooted iPhone, straight from Apple?
Apple’s Security Research Device page hit the HN front page: 167 points, 109 comments. The SRD is an iPhone built for security research, with root shell and custom kernels, used to find and validate vulnerabilities. The argument in the comments is access: device numbers are small and applications require a track record of published vulnerability research, which independents say puts it out of reach. iOS security researchers: eligibility rules are on the official page, and the path starts with published work. Source · HN discussion
12. gpu-unlock lifts RTX 3000 FP/BF16 rates; a 3090 Ti gains ~80%
The 3090 always had it in it?
The open-source kernel module gpu-unlock raises FP/BF16 instruction issue rates on RTX 3000-series cards to professional-GPU levels: 146 points, 55 comments. Measured on an RTX 3090 Ti, dense BF16 GEMM goes from 87.5 to 157.4 TFLOPS (+80%); LLM prefill gains 14-36%, decode is unchanged (bandwidth-bound). The module runs in the initramfs before the stock NVIDIA driver and opts in via kernel command line, with packages for Arch, Debian/Ubuntu and Fedora. Running local inference? This frees up part of the prefill bottleneck on older cards for free. Source · HN discussion
13. Dream-1 rents robot control at $3/hour, zero-shot on unseen robots
Robot control, metered by the hour now?
Dream Labs launched Dream-1, a World Action Model that controls bimanual robots through a public API and claims zero-shot generalization to unseen embodiments: 27 unseen robot bodies tested in simulation, real-robot demos including the 18-year-old PR2. Third-party evaluator Robocurve reports it beats Opus 5.5 and Gemini 4 Argon on RoboDojo-RC Tier 1 at about 1/16 the cost per task attempt. Pricing is $3/hour, $2 for the flash tier. 114 points, 58 comments. For embodied-AI work, this is the first general robot-control API billed by the hour. Source · HN discussion
14. 9 frontier models show the same hard-to-read pattern
So that’s why it reads like that.
The arXiv paper “Why is AI-generated language so hard to understand?” argues human language is optimized for communication while model output is optimized for information density, so AI text leans on inanimate subjects with animate verbs, low-frequency n-grams and longer syntactic dependencies. The authors ran corpus statistics across nine frontier and near-frontier models: families differ but the pattern holds, and a length-penalty fine-tuning experiment supports the mechanism. 108 points, 100 comments. Prompt writers and AI-editing tool builders get a checkable feature list for “AI flavor.” Source · HN discussion
15. A Z-Library paper cites at least 10 nonexistent sources
AI found the citations, AI wrote the citations?
TorrentFreak reports that a peer-reviewed paper on Z-Library, published in a Taylor & Francis journal, has at least 10 of 44 references unlocatable, including a TorrentFreak article that was never published and a law-review piece whose cited volume, issue and pages belong to other work. The authors disclosed using Consensus AI for literature screening and ChatGPT for language polish; the publisher has referred the case to its ethics teams. 90 points, 47 comments. If you use AI for literature review, this is proof that verification cost is non-negotiable. Source · HN discussion
16. A practical guide to AI image generation, from diffusion to agents
Half your prompt problems gone after this one?
danielball.com’s long practical guide walks AI image generation from diffusion models to agent-driven generation pipelines: 232 points, 73 comments. It organizes the field by task (text-to-image, editing, controllable generation), with the technical keywords and known pitfalls for each stage. Product and engineering folks new to image generation can follow the sequence and skip half a year of fumbling. Source · HN discussion
17. OBS Studio 33 rebuilds plugin loading, brings NVENC to Arm64
Even a streaming tool’s update sparks debate?
OBS Studio 33.0.0 shipped with 424 points, 151 comments. Highlights: 1, the plugin-loading system is fully rebuilt; third-party plugins move to a new directory, with legacy paths supported until 34.0; 2, the embedded browser engine CEF jumps from 127 to 150; 3, NVENC encoding arrives on Arm64, and Windows gains an arm64 build. Streamers and recording setups: verify your go-to plugins against the new loading path before upgrading. Source · HN discussion
18. Twelve years of rqlite, the distributed SQLite story
Twelve years on one unfashionable problem, still shipping?
Philip O’Toole published a retrospective on twelve years of distributed SQLite: rqlite grew from a personal project into a database used across production environments, 348 points, 56 comments. The piece walks through key decisions year by year, centered on sticking to one small, well-defined problem. For infrastructure product builders, it’s a long-term maintenance ledger, more useful than a growth story. Source · HN discussion
19. uv and ty integrate, one uv check for deps and types
One command for dependencies and types now?
Astral announced experimental integration between uv and ty: 281 points, 92 comments. Two capabilities ship: 1, PEP 723 inline-dependency scripts are now type-checked by ty against each script’s own declared requirements, with uv preparing isolated environments; 2, a new uv check command prepares the environment, installs dependencies and runs the type check in one pass, pinning ty to the project lockfile. Requires uv 0.12.3+ and ty 0.0.76+. Python teams can swap it into CI today. Source · HN discussion
20. Goose, a systems language with no heap, written mostly by AI
AI as the implementer, human as the designer?
Wouter van Oortmerssen, creator of the Lobster language, released Goose: a systems language with no heap at all: dynamic values live on compiler-managed data stacks and memory frees when scope ends, with no GC, reference counting or lifetime annotations. He claims 3.3x the speed of idiomatic C++; more unusual is that the ~30k-line compiler was written almost entirely in collaboration with AI (Claude), and John Carmack tweeted about its multi-stack and relative-reference design. The two threads drew 211 points, 90 comments. Language designers: this is a complete experiment in human-designed, AI-implemented tooling. Source · HN discussion
21. ThinkRail, JetBrains’ open-source agentic IDE on git worktrees
Parallel agents that don’t stomp your files?
ThinkRail, a JetBrains InnovationHub project, is open-sourced under Apache 2.0: an agentic IDE where each agent works in a real git worktree on its own branch, and project intent lives in a linked graph of SPEC.md files so intent survives sessions. It bundles no models and runs the pi coding agent in-process; the demo session cost $0.17. macOS (Apple Silicon), Windows and Linux are supported. 49 points, 73 comments. Teams experimenting with parallel agents can borrow the worktree-isolation-plus-spec-graph combination directly. Source · HN discussion
22. Block Blast hides ~700 encrypted ad-harvesting traits
Free game. You’re the product.
A reverse-engineering breakdown dissects Block Blast’s ad-harvesting machinery: game logic ships from the server as encrypted JavaScript split into roughly 700 “traits,” including RewardHighValueAdHarvestTrait, which deals an unbeatable hand once three conditions align: you finished your last three rewarded ads without closing early, the current ad is priced at least 3.5x your previous one, and your score passes your record plus 1,000 points. The analysis also found a 500-tree XGBoost model predicting where you’ll die and another estimating each player’s ad value. Sensor Tower estimates $127M in ad revenue Jan-May 2026, above Candy Crush’s $81.1M. 105 points, 47 comments. Casual-game builders: this is a complete evidence chain of difficulty tuning crossing an ethical line. Source · HN discussion
23. Roku moves its developer repos off GitHub to GitLab
What did GitHub do to Roku?
Roku announced on its official blog that developer open-source repositories are moving from GitHub to GitLab: 211 points, 141 comments. The migration covers repository addresses, issues and collaboration workflows. If you depend on Roku’s open-source SDKs, watch the announcement for the new remotes and update local configs and automation scripts accordingly. Source · HN discussion
24. FreeDNS, the free DNS host that won’t die, is back on the front page
Still hosting your domains there?
afraid.org’s FreeDNS hit the HN front page again: 212 points, 56 comments. The service has run for over two decades, offering free DNS hosting and shared subdomains on community donations. Half the comments thank it for surviving this long; the other half warn about free-tier availability limits. For side projects that need stable DNS without hosting fees, it remains one of the cheapest options. Source · HN discussion
25. The Quake fan who built the map editor everybody uses
Three decades later, someone’s still building Quake tools?
Funcarena profiles TrenchBroom author Kristian Duske: a Quake fan dissatisfied with existing tools who built the editor nearly every id Tech mapper now uses, 181 points, 23 comments. The piece traces the path from 1.0 to community standard. Tool builders: it’s a user-as-author specimen: he built the tool because he used it every day. Source · HN discussion
26. AI collapsed software costs, so dream bigger, says Dan Luu
Projects you shelved as too expensive are back?
Dan Luu’s essay “We should be more ambitious with the software we build” drew 490 points and 421 comments, the day’s second-busiest thread. His case: AI coding models have collapsed the cost of building software, so projects that used to be infeasible now take hours: he built a faster-than-ripgrep fork and a Rust interpreter with ~50x faster cold compiles in hours each, and an Azul board-game AI passed human level after 5 hours of his time and reached state of the art in about 20; Anthropic’s seven-person team rebuilt GitHub internally in five months. His boundary condition: verification improves more slowly than models, so people without self-checking systems risk “abject failure” (quoting Yossi Kreinin’s counterargument). Tool builders: re-audit the list of things you shelved as too expensive. Source · HN discussion
27. A pure functional language lands within 2.6% of C
Memory safety for a single-digit percent?
Richard Feldman benchmarked his Roc deflate library against libdeflate, a heavily tuned C library with SIMD, across 144 cases (12 Silesia corpus files × 12 compression levels): the median slowdown is 2.6% on an M1 and 8.9% on a Ryzen 9 3900X, worst case about 18% slower, some inputs about 18% faster. Roc gets there without unsafe code: automatic reference counting, bounds checks and opportunistic in-place mutation all stay on, and the language is weeks from its 0.1.0 release. 272 points, 98 comments. Writing performance-sensitive services? The pure-FP performance ceiling is higher than you think. Source · HN discussion
28. C# vs C on VLC’s Gaussian blur ends in a 9% gap
The gap was never the GC?
Developer Martin Finkel ported VLC’s gaussianblur.c to C# line by line and ran it inside VLC 4 as a NativeAOT plugin on the same machine, same 1080p frames, byte-identical output, single-threaded. The naive port ran 23% slower; the bottleneck was RyuJIT’s codegen, not garbage collection, and hoisting loop bounds into locals closed the gap to about 8%. After identical algorithmic work both versions ran ~2.5x faster than current VLC code; hand-tuned AVX2 in both took sigma 4 from 7.8 to roughly 240 fps, leaving C# within 9% of C. 131 points, 49 comments. For .NET performance work: C only wins the last 10%. Source · HN discussion
29. Terry Tao’s blog hosts a hub fixing OpenAI’s math write-ups
The math community is copy-editing a lab’s papers now?
A guest post on Terry Tao’s blog announced oai-math-revised, a community effort to revise OpenAI’s public math write-ups (github.com/openai/math). It was started by Abhinav Nellore, Johannes Schmitt, Soledad Villar and Rachel Ward, with the repo at github.com/mathcommunity/oai-math-revised. The motivation: write-ups that are hard to follow with incomplete citations, amid broader unease about AI-generated slop; commenters also flagged that some hardness claims appear to duplicate known Khot–Regev results. 48 points, 36 comments. Research communicators: a working template for community peer review of corporate research output. Source · HN discussion
30. The PHP Foundation ran a $139k deficit in 2025
When did you last pay for the infrastructure you run on?
Reuters reports the PHP Foundation is “under-funded,” with a single full-time staff member. The foundation’s own 2025 report supplies the numbers: expenses of $784,376 against $645,191 in donations, the ~$139k gap covered from reserves; the one full-time director’s salary is paid by JetBrains, and core development rests on 11 contracted developers (13 as of Q2 2026). 139 points, 96 comments. If your company runs on PHP, this is how fragile language-infrastructure funding actually is. Source · HN discussion
31. Forrester puts AI at one-fifth of tech spending by 2027
The money is committed. Where is the return?
Forrester projects AI at 20% of corporate technology spending in 2027, Computerworld reports, with colder numbers from the same research: over half of IT executives expect to miss their AI ROI targets; 45% of AI decision-makers demand a positive return within one year; and only 22% reported a positive operating-profit impact from AI in the past year. 21 points, 5 comments. Building an AI procurement case? These figures go straight into the deck. Source · HN discussion
32. Meta’s new terms make customers liable when its AI errs
Meta controls the models, you carry the risk?
CIO/Computerworld reports Meta’s revised commercial terms take effect October 30: they acknowledge that “AI features may make mistakes,” yet enterprise customers are “solely responsible for and assume any and all risks” from AI features and their output, and must indemnify Meta over claims arising from its AI tools. Aikido Security’s Mike Wilkes: “that makes no sense — only Meta controls the models”; the report contrasts OpenAI’s IP indemnification for enterprise customers. 18 points, 10 comments. Signing a Meta enterprise contract? Have legal renegotiate the liability terms line by line. Source · HN discussion
33. A work browser valued at $6.4B, and the comments are furious
When did the browser become the priciest security tool?
The Times of Israel reports Island, the US-Israeli enterprise browser, raised $400M at a $6.4B valuation, $1.2B raised in total, led by Evolution Equity with Sequoia and Coatue participating. Island sells a managed Chromium that logs who is acting and which data they can reach, now extending the same governance to “human and agentic” workforces; customers include 8 of the 10 largest US banks, plus Pfizer and Amazon. 16 points, 29 comments, nearly twice as many comments as upvotes, and the employee-surveillance premise itself draws the fire. Enterprise security teams: this is the latest price marker for the agent-era work browser. Source · HN discussion
34. Skyvern deleted its RAG layer and got 2.3x faster
Too much context was slowing the agent down?
Skyvern replaced its 2.0 preprocessing layer (parse each page, screenshot and annotate it, condense the interactable elements) in favor of a tool-using agent that starts from minimum context and pulls more only as needed. An A/B test across roughly 100,000 production runs shows mean runtime down from 593s to 262s (2.3x) and cost per run down 22.5% from $0.039 to $0.0302; image-bearing LLM calls fell from 27% to 2.7% of turns and prompt-cache hits rose 195%. 18 points, 1 comment. Agent engineers: a counterintuitive optimization path, and it starts with subtraction. Source · HN discussion
35. HumanLayer open-sources its agent-workflow stack under Apache 2.0
Software factories, but open and composable?
YC-backed HumanLayer open-sourced the core stack behind HumanLayer Cloud (github.com/humanlayer/humanlayer, Apache 2.0): a desktop/Web IDE plus a local daemon running parallel Claude Code and Codex sessions, cloud daemons for long-running work, Electric Postgres sync, and integrations with GitHub, Linear, Jira and Slack, with zero telemetry in self-hosted builds. 31 points, 5 comments. Building an internal agent platform? This component list is a ready architecture reference. Source · HN discussion
36. Claude Code reportedly puts user emails in outbound request headers
Where exactly did your address just go?
Security researcher Chris Lakin reported on October 24 that Claude Code appends the user’s email address to the User-Agent header of outbound HTTP GET requests, so any external API call or webhook the agent makes can expose the developer’s identity; the same behavior was reported in April as issue #53231 in Anthropic’s repository, with no public company response as of the report. 21 points, 4 comments. Running Claude Code with outbound tools or webhooks? Inspect the request headers at your proxy layer first. Source · HN discussion
37. 1,377 Indian government domains still serve revoked certificates
Firefox blocks them, Chrome loads them fine?
Blogger Capt. Nemo tracks the eMudhra incident: on October 17, eight eMudhra intermediate CAs were found to have trivially recoverable private keys, revoked the next day as Mozilla and Chrome began distrusting the emSign G3 root; the flawed keys date to April 2025, an exposure window of roughly 18 months, with about 30,000 certificates needing migration. His October 24 count: 1,377 Indian government domains on revoked certificates, 811 still serving them, including UIDAI and the Controller of Certifying Authorities itself. 51 points, 22 comments. The split is browser-side: Firefox blocks all of them via CRLite, while Chrome and Edge CRLSets don’t yet include these certificates. Running a site? Don’t verify revocation compliance in Chrome alone. Source · HN discussion
38. Proton defies cease-and-desist, hosts map of 300k Flock cameras
You can’t even know which cameras record you?
Tom’s Hardware reports Proton is hosting the Flock camera map on Swiss servers: over 300,000 Flock camera locations, published after the project’s creator received cease-and-desist letters from Flock; Proton’s position is that users “have the right to know who’s recording them.” Two front-page threads drew 139 points combined. Tracking surveillance expansion? This map is a first-hand lookup for cameras near you. Source · HN discussion