HN digest

Jensen Huang resignsapnews.com · 3984 points · 3385 comments · 4dbusinessaiJensen Huang has resigned as NVIDIA's president, CEO, and board member effective immediately, and CFO Colette Kress is interim CEO while the board searches for a permanent successor.

Key points

  • Kress will keep her CFO role while serving as interim CEO, and the board will consider both internal and external candidates for the permanent job.
  • Harvey C. Jones becomes the company's first-ever board chair; he previously was CEO and Executive Chairman of Synopsys.
  • Kress joined NVIDIA in September 2013, after nearly 25 years in finance roles at Cisco and Microsoft.
  • NVIDIA says it will still report Q3 fiscal 2027 results on November 17, as previously scheduled, followed by an earnings call.

What HN thinks

  • Commenters widely debate whether the affair or the reported export-control circumvention drove the decision, with many arguing the China sales and reported spy link are the more serious issue.
  • Several commenters suspect political motives, pointing to the post-midterm shift and the administration's earlier knowledge, though this remains speculation without evidence.
  • Concern centers on NVIDIA's lack of a succession plan and whether the stock and AI sentiment can hold up without Huang, with some noting Colette Kress is a capable interim choice.
Service degradation on Microsoft 365status.cloud.microsoft · 1125 points · 947 comments · 4dsystemstoolsOutlook went down for thousands of users, according to Downdetector reports, and Microsoft's status page was also reported unavailable during the outage.

Key points

  • The article was not retrievable, so the scope and duration of the outage are unconfirmed beyond the headline's report of thousands of affected users.

What HN thinks

  • Commenters largely treated the Microsoft status page updates as a sign of trouble: servers failing health checks, restarts making things worse, and a rollback of an earlier update.
  • Several commenters noted that Downdetector spikes during big cloud outages are unreliable because users misattribute failures, such as Outlook problems showing up as Gmail or AWS outages.
  • One commenter with on-premise Exchange experience said classic Outlook clients still depend on Microsoft's cloud for some connections, so self-hosted setups were affected too.

The article couldn't be read, so this summary is based on the discussion.

Anthropic IPOipo.anthropic.com · 980 points · 886 comments · 5daifinanceAnthropic is offering Class A common stock to the public on Nasdaq under ticker ANT, with pricing expected November 18 and trading starting November 19.

Key points

  • Roadshow begins November 10 (EST), and the final price depends on regulatory clearance, investor demand, and market conditions.
  • Q3 2026 revenue was $17.0B, over 12x year over year, with about 11,900 customers paying at least $100K trailing twelve months.
  • Q3 2026 net loss was $23.3B, driven largely by a $21.9B non-cash charge from revaluing convertible notes, derivatives, and SAFEs.
  • Adjusted operating margin was 15% while GAAP operating margin was negative 3%, a gap the prospectus reconciles.

What HN thinks

  • Commenters are skeptical of the valuation, with several pointing to the roughly $23 billion quarterly net loss and the large non-cash adjustments in the adjusted margin figures.
  • A recurring thread argues that model switching is cheap and open or Chinese models are close enough for most work, so Anthropic's pricing power and moat are questioned.
  • One commenter notes the filing's inference margin claim is gross margin, while operating and net margins remain negative, and another argues subscriptions are still subsidized.
Rust 1.100blog.rust-lang.org · 865 points · 311 comments · 4dprogrammingtoolsRust 1.100.0 stabilizes the never type, the core Allocator API, a DropGuard type, and a dedicated function for creating zero-sized types, plus Cargo cooldown and unused-dependency warnings.

Key points

  • Writing ! in arbitrary positions such as Result<T, !> is now allowed, and Infallible becomes an alias for it.
  • Editions 2015, 2018 and 2021 change never-type fallback from () to !, a breaking change that can alter some code.
  • Cargo's min-publish-age setting, such as global-min-publish-age = "3 days", skips recently published versions during resolution.
  • cargo install ignores the minimum publish age for both the crate and its dependencies, so pin versions when needed.

What HN thinks

  • Commenters largely welcomed the cargo change making the lockfile in the .crate tarball the default for cargo install, calling the old behavior a long-standing gotcha.
  • Dependency cooldowns drew debate: several commenters argued they guard against supply chain attacks, while others noted they can delay urgent security patches, and the consensus was to use cooldowns alongside immediate patching of advisories.
  • Compile times remained a recurring complaint, with one commenter noting that faster compilation and smaller build directories can conflict, and maintainers listed planned reductions in debug info and build-dir size.
Show HN: My $300 embroidery machine needed a subscription, so I built my own appgithub.com · 712 points · 231 comments · 4dtoolsopen-sourceZigZag is a free, Apache-2.0 embroidery editor for Mac and Windows, written in Rust, that turns SVG and traced logos into stitches and sends designs to a Brother Skitch over Bluetooth.

Key points

  • Exports PES, DST, EXP, JEF and VP3 files, and can send designs directly to a Brother Skitch (PP1) over Bluetooth.
  • Includes an MCP server so AI agents such as Claude Code can edit the open design, with each agent call as one undo step.
  • The Skitch Bluetooth protocol was reverse-engineered, and the author notes the app has not yet been run against a real Skitch.
  • Offers satin, tatami, running, contour, appliqué and other stitch types, plus Ink/Stitch and Hatch feature comparisons in the README.

What HN thinks

  • Commenters split on whether AI-built tools like this are a win for ownership and against subscriptions, with some arguing the real value is in support rather than the software itself.
  • The strongest criticism is that the project builds directly on Respira, an existing open source project, and the README and post did not credit it clearly enough.
  • Several commenters pointed to InkStitch as the established free alternative, noting it works but is awkward because it lives as a plugin inside Inkscape.
Socratic Tutorgithub.com · 641 points · 146 comments · 7daitoolsJane Street published a SKILL.md file that tells an AI agent to run a structured Socratic tutoring session, placing the learner on a 0-5 skill map before teaching.

Key points

  • The skill triggers only on explicit requests like 'teach me' or 'quiz me', and it asks for confirmation when intent is unclear.
  • Each session starts with a 3-5 question diagnostic, then the agent asks one question at a time and withholds answers.
  • Levels 0-5 are earned only after demonstrated performance on at least two checks, one of them unprompted recall.
  • Users can save session state as a compact block and resume later, with commands like 'show progress' and 'just tell me'.

What HN thinks

  • Several commenters report the Socratic tutor skill works well for dense material, such as a philosophy text or learning Lean, Logo, or Python, and it pushes learners to think rather than receive answers.
  • Commenters note much of the effect can be reproduced with a single prompt line, such as asking narrower questions when the user is wrong, so the skill format itself adds little.
  • Skeptics point out that these tutors struggle to build progressive curricula from scratch, and context degradation over long sessions can make later questions less useful, while the idea itself is long established and many similar repos already exist.
Gemini 4 Hydrogenblog.google · 546 points · 340 comments · 4daitoolsGoogle released Gemini 4 Hydrogen, a faster, cheaper model for everyday coding and agentic work, priced at $1.50 per million input tokens and $7.50 per million output tokens.

Key points

  • Scores 51 on the Artificial Analysis Intelligence Index at default high thinking, ahead of Gemini 4 Argon's 50, and 54 at a new xhigh setting.
  • Claims 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash, and 2.5x faster task completion at default settings.
  • Orchestrator in Google Antigravity switches between Carbon and Hydrogen in one session, reportedly using 24% less quota and finishing 19% faster on Terminal-Bench 4.0.
  • Pricing is flat across the full 1 million-token context window, with up to 1 million output tokens supported.

What HN thinks

  • Commenters broadly agree the element-based naming scheme is confusing, since Hydrogen, Argon and Carbon do not map intuitively onto a size or capability ordering like Haiku-Sonnet-Opus or Luna-Terra-Sol-Astra.
  • Several commenters note the pricing is puzzling: Hydrogen costs the same as the non-discounted 3.8 Flash and roughly twice the price of Argon's introductory rate, raising worries about whether Google is abandoning the cheap tier.
  • One commenter reports that Hydrogen's cost per task was much lower than 3.8 Flash on Artificial Analysis, attributing the gap to a much higher cache hit rate, while another found it 37% pricier per task but 40-60% faster.
Gemini 4 Hydrogen and Orchestratorantigravity.google · 489 points · 160 comments · 4daitoolsGoogle Antigravity adds Gemini 4 Hydrogen and an Orchestrator model option that switches dynamically between Hydrogen and Gemini 4 Carbon within a single task.

Key points

  • Orchestrator lets Carbon plan and read-only explore, then hands edits and tests to Hydrogen, returning to Carbon when Hydrogen hits friction.
  • On Terminal-Bench 4.1, Orchestrator reaches near-Carbon task success while using on average 24% less quota and finishing 19% faster.
  • Hydrogen scores 51 on the Artificial Analysis Intelligence Index, above Argon's 50, and 57% on Terminal-Bench 4.1 versus 18.2% for 3.8 Flash.
  • Hydrogen offers four thinking levels: low, medium, high (default), and a new xhigh tier.

What HN thinks

  • Commenters largely focused on the confusing model naming scheme (Hydrogen, Carbon, Argon) and how it does not map to size or capability.
  • Several commenters said the Antigravity harness, with its heavy file-gathering before tasks, is slower and more costly than the underlying models, so harness quality matters more than model choice.
  • Commenters compared the Orchestrator approach to Anthropic's advisor mode and third-party multi-model pipelines, with one reporting roughly 60% savings from pairing a large model with a cheap subagent model, though others questioned caching and cost tradeoffs.
I'm making LLMs work on small and old devicesgithub.com · 460 points · 173 comments · 4daitoolssmolmaxxing is an MIT-licensed llama.cpp patch and kit that runs LLMs from 260K to 16B parameters on a Xiaomi TV Box S and a 2013 iPad Air with under 1 GB usable RAM.

Key points

  • On the TV box, Qwen3.5-2B runs at 7.7 tok/s in chat, and a 16B mixture-of-experts model runs at 5.1 tok/s by paging expert weights from flash.
  • The 1-bit Bonsai-8B goes from 1.2 to 5.2 tok/s on the TV by using a headless mode that force-stops Android launchers and Google services to free RAM.
  • On the 2013 iPad Air, Bonsai-1.7B reaches 24-26 tok/s, and the 8B and 16B MoE models load in 556-625 MB via an expert pager.

What HN thinks

  • Commenters broadly see the project as a fun hardware-limits challenge rather than a practical tool today, with several noting that 0.5B to 1B models remain too weak for serious agentic or tool-calling work.
  • Some argue that Qwen models are scaled down poorly for tiny sizes and that purpose-built small models such as LFM, Gemma E2B, or Granite do better, and the author says LFM became the main model after Qwen degraded.
  • A recurring thread questions how much of the work was hand-directed versus generated by an AI agent, with one commenter noting that a coding agent can now port inference engines to new hardware, shifting the difficulty to sourcing the device.
Former Nvidia CEO Huang had a relationship with a second employee, sources sayreuters.com · 436 points · 465 comments · 4dbusinesscultureReuters reports, citing unnamed sources, that former Nvidia CEO Jensen Huang had a relationship with a second employee.

What HN thinks

  • Commenters widely treat the affairs themselves as less important than the alleged link to a Chinese spy and the export-control and Singapore shell-company questions raised by the reporting.
  • Several commenters argue that the timing of the disclosures, close to the midterms, Trump's China trip, and Nvidia's earnings, looks coordinated, though others say it reflects one investigation prompting more leaks.
  • A recurring point is that affairs with subordinates violate company rules, and some note that open-secret tolerance among executives may explain why the relationships were known inside the company.

The article couldn't be read, so this summary is based on the discussion.

Ask HN: Microsoft/Azure Outage?388 points · 195 comments · 4dcultureExchange Online mailbox access is failing for many Microsoft 365 users in the Americas, with Microsoft's status page itself slow or unreachable during the outage.

Key points

  • Microsoft's status page lists a service degradation affecting Outlook on the web, MAPI, Exchange ActiveSync, and the REST API.
  • Microsoft says connectivity errors are on a portion of Exchange Online infrastructure that receives and processes user access requests.
  • Microsoft's next update was promised by 8:00 PM UTC, and the incident is tracked as EX1550543 in the admin center.

What HN thinks

  • Commenters report the outage hits Exchange Online and Outlook broadly across organizations, with some saying Teams, OneDrive and other Microsoft 365 services are also affected while Azure itself appears mostly fine.
  • Several commenters speculate about causes such as a self-inflicted change, AI-related traffic load, or a DDoS, but the thread offers no confirmed cause beyond Microsoft's own statements about elevated load and unavailable servers.
  • One commenter notes that Microsoft's status page is hosted on infrastructure that is also failing, so it became slow or unreachable, and that Downdetector spikes can be misattributed to unrelated services like Gmail or AWS.
Chinese frontier AI models carry the fingerprint of export-controlled GPUsunemploymentlabs.com · 384 points · 165 comments · 4daihardwareA forensic analysis of published model weights reports that six recent Chinese models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on NVIDIA Blackwell-layout hardware, which export rules restrict.

Key points

  • The method reads the seed and thread-count layout of the random initialization preserved in weights, matching PyTorch's cuRAND behavior on NVIDIA GPUs.
  • Blackwell-class readings use 148 SMs (B200/B300) or 152 SMs (GB200/GB300), while earlier models from the same labs read as 132-SM Hopper.
  • On a physical MI300X, DeepSeek-V4-Flash weights matched NVIDIA rather than AMD in 363 of 377 differing values, excluding the AMD alternative for that model.
  • Estimated training runs needed roughly 500 to 3,000 Blackwell GPUs for two to three months, with the largest clusters costing $150 million to $200 million.

What HN thinks

  • Commenters broadly note that renting Nvidia GPUs in Singapore or Malaysia is legal under current US rules, so the findings do not by themselves show any violation.
  • A recurring caveat is that the method shows only that a Blackwell GPU initialized the weights, not which hardware ran the bulk of training, so the claim is weaker than the headline suggests.
  • One commenter notes DeepSeek already published Blackwell-only inference kernels, so the Blackwell usage was not really a hidden fact.
Dwarf planet candidate appears which takes roughly 280 years to orbit our sunsmithsonianmag.com · 383 points · 172 comments · 4dspacescienceAstronomers using Vera C. Rubin Observatory data have found 2026 ED138, a dwarf planet candidate about 46 times Earth's distance from the sun with an orbit of roughly 281 years.

Key points

  • Estimated diameter ranges from about 370 to 930 miles, but brightness alone cannot settle whether it is large enough to be round.
  • Confirming size would require infrared thermal measurements or timing how long it blocks light from a background star.
  • The object was found in earlier sky images across the past decade and in Harvard photographic plates from 1937.
  • It has been crossing a region in front of the dense Milky Way core that is hard to survey, and its orbit is tilted about 25 degrees.

What HN thinks

  • Commenters noted the object was found partly because it sits in the Milky Way's dense star field, which usually hides moving objects, and that the Vera Rubin Observatory's sensitivity made the detection possible.
  • Several commenters defended the 2006 reclassification of Pluto, arguing Pluto is an outlier among planetary-size measures, while a minority argued Ceres could have been promoted instead.
  • One commenter who previously computed asteroid orbits shared an interactive orbit viewer, and others pointed out that a Planet X would need to be far larger than this dwarf planet candidate.
Writing correct software is hard and math alone won’t help us (2016)pron.github.io · 359 points · 169 comments · 5dprogrammingscienceA long essay argues from computability and complexity theory that verifying general programs is fundamentally infeasible, so software correctness always carries an unavoidable cost.

Key points

  • Bounded halting is shown to be Omega(|S|): no method can decide whether a program halts within n steps faster than effectively running it.
  • Verification of finite-state machines is PSPACE-hard, and a 2005 result suggests modular decomposition does not make correctness proofs tractable.
  • A 10-line Java example that crashes iff Goldbach's conjecture is false illustrates how simple composed code can resist proof.
  • A 1919-state Turing machine is cited whose termination is independent of ZFC, so no proof of its behavior exists even in principle.

What HN thinks

  • Several commenters argue the article's theory is largely beside the point for practical work, since real programs have bounded inputs and most failures come from shifting or disputed definitions of correct rather than undecidability.
  • One commenter notes that Rice's theorem only rules out a general decision procedure for all programs, not proofs about specific simple programs, and that the article seems to conflate the two.
  • A recurring caveat is that sensor noise, adversarial inputs, and business politics limit what any formal proof can guarantee, so safety cases and lightweight methods are often the practical route.
A live tree of systemd jobs on the boot consolecodeberg.org · 355 points · 140 comments · 7dsystemstoolsbootvine is a Rust program that replaces the scrolling boot log with a live tree of pending systemd units on a dedicated virtual terminal, plus a timeline of the slowest units.

Key points

  • It also follows shutdowns and answers LUKS passphrase prompts in place, so the tree stays usable through unlock.
  • When a unit fails, it shows that unit's last journal lines directly on the boot screen.
  • It requires systemd 246 or newer as PID 1, libdbus and libsystemd, and a systemd initrd to cover early boot.
  • A Nix command boots a sample NixOS VM with slow and failing units, so it can be tried without changing the host's boot.

What HN thinks

  • Several commenters note that on fast-booting machines the 12-second demo is unrepresentative, and much of the delay is firmware or memory training that the tool cannot shorten.
  • Commenters point out that systemd-analyze critical-chain and systemd-analyze plot already show blocking units, and that watch with systemctl status gives a rough live view.
  • Some worry the graphical boot view hides the plain text log that is needed when a machine fails, such as over a serial console, and suggest keeping a fallback.
Critical RCE Vulnerability in Apache HTTP Serverwiz.io · 341 points · 183 comments · 5dsecuritywebApache HTTP Server 2.4.55 through 2.4.70 has a race condition in mod_http2's assign_headers() that can corrupt heap memory and allow remote code execution; fixed in 2.4.71.

Key points

  • The bug is a synchronization flaw in assign_headers() that a remote, unauthenticated attacker could trigger to corrupt heap memory.
  • Affected Apache HTTP Server versions are 2.4.55 through 2.4.70, and the patched version is 2.4.71.
  • Temporary mitigation is to set Protocols to http/1.1, removing h2 and h2c, which disables HTTP/2 at the cost of that feature.
  • Wiz reports 64% of cloud environments have vulnerable resources and 51% have public instances, but this ignores configuration.

What HN thinks

  • Commenters note the vulnerability is in mod_http2, so disabling HTTP/2 (a2dismod http2) or hard-coding HTTP/1.1 is a practical mitigation that some say has few downsides.
  • Several commenters point out that prefork-based installs, common in cPanel and WordPress images, may be less exposed than the default event MPM, though this was disputed.
  • Debate continues over whether memory corruption alone counts as RCE; one commenter argues the 9.8 CVSS rating and Red Hat's claim of a reliable exploit justify treating it as RCE, while others want a public PoC first.
Jensen Huang resigns as Nvidia CEO immediatelyreuters.com · 334 points · 174 comments · 4dbusinessaiThe headline reports that Jensen Huang resigned as Nvidia CEO with immediate effect, but the article could not be fetched, so this is unverified. The Reuters URL indicates a live board meeting over his fate after affair reports.

Key points

  • The Reuters live-blog URL indicates the board was meeting about Huang's position after reports of an affair.
  • The article itself was unavailable (HTTP 401), so no details on the board's decision or any stated reasons can be confirmed.

What HN thinks

  • Commenters widely speculate the board acted over the sanctions and spy allegations rather than the affairs alone, though the reasons remain unconfirmed.
  • Several commenters doubt the resignation will hurt Nvidia's core business, arguing its chip demand and roadmap persist, while others warn that losing Huang's vision and sanctions fines could matter.
  • Some commenters argue the timing near the earnings report and an IPO looks suspicious, though these claims are speculative and unsupported.

The article couldn't be read, so this summary is based on the discussion.

Abandoned Quake 4 expansion finished by modders and now playablemoddb.com · 298 points · 118 comments · 4dgamingopen-sourceopenQ4 0.14.1 is a free open-source replacement engine and game code for Quake 4 that plays the recovered, previously unreleased Awakening expansion, which requires your own Quake 4 copy.

Key points

  • The Awakening's 13 maps run on openQ4's own code; the alpha's game DLL is never loaded, and the expansion's content must be supplied by the user.
  • Where the alpha was broken, openQ4 fixes it, including the Valkaryne finale, the beach death-ray turret, and Pain Lord weak spot.
  • Vulkan on Windows is now the default for new installs, tested across all stock single-player maps and 20 multiplayer maps, with OpenGL fallback.
  • Multiplayer requires matching 0.14.1 clients and servers; retail Quake 4 and mods built for its original game DLLs are incompatible.

What HN thinks

  • Several commenters pushed back on crediting AI for the work, arguing the human modders and the tool's users did the finishing, while others noted AI output can look polished even when it is wrong.
  • A commenter noted the release is a fan-driven effort, so the value comes from community labor rather than the company, and some questioned whether modders would accept paid AI-based modding models.
  • Commenters recalled Quake 4 as a divisive game, with one noting its German release was heavily censored and another recalling its early DVD release as notable at the time.
Rust 1.100.0github.com · 291 points · 62 comments · 4dprogrammingtoolsRust 1.100.0 is released, stabilizing the never type, custom allocators, and extern "custom", and adds Cargo dependency cooldowns and a built-in unused-dependency lint.

Key points

  • Cargo can now ignore crate versions published more recently than a configured minimum age, unless they are already pinned in Cargo.lock (RFC 3923).
  • cargo install now uses the lockfile bundled in the .crate tarball by default, making --locked a no-op.
  • Infallible is now a type alias to !, and the never type fallback changed to ! on all editions.
  • Several ARMv7-A hardfloat targets that wrongly required VFPv3-D32 can be fixed by passing -Ctarget-feature=+d32.

What HN thinks

  • Commenters broadly welcomed the release as large, with the never type and custom allocators singled out as long-awaited features.
  • Several commenters noted the blog post omits stabilized APIs such as bool::toggle, Box::map, and str::trim_prefix/trim_suffix.
  • Debate over a stable Rust ABI continued: some want prebuilt library distribution, while others argue a stable native ABI would be a forward-compatibility burden, and one commenter pointed to the crABI RFC effort.
Nintendo Takes Down Emulators and Firmware in Massive GitHub DMCA Sweepign.com · 260 points · 198 comments · 4dgamingpolicyNintendo sent four DMCA takedown notices to GitHub targeting the Yuzu and Citron Neo Switch emulators and mod chip firmware, with 144 firmware repositories removed.

Key points

  • A November 6 notice says Switch TPMs restrict the console to legitimate Nintendo game files, and the firmware is used to circumvent that protection on mod chips like Picofly.
  • TorrentFreak reported the firmware was still hosted elsewhere on GitHub after the removals.
  • In 2024 Nintendo took down about 8,500 Yuzu copies by targeting 12 mirrors of the emulator.
  • The takedown followed emulators running the remake of The Legend of Zelda: Ocarina of Time on PC a few days earlier.

What HN thinks

  • Several commenters say they will stop buying Nintendo products over its takedowns, citing the Bowser/Team Xecuter case and repeated DMCA actions.
  • Commenters note that emulators themselves are legal in many places, and that the legal risk lies mainly in DMCA anti-circumvention trafficking rules and in distributing copyrighted games or firmware.
  • Some commenters point out that emulators for older consoles remain widely available and that Nintendo has so far targeted the newest Switch emulators, and that projects have been removed largely because they were hosted on GitHub under identifiable accounts, while self-hosted forges and Radicle are suggested as alternatives.
Humble Bundle just killed most of their security through vibecodingforums.spacebattles.com · 257 points · 129 comments · 5dsecurityaiA forum thread alleges Humble Bundle seriously weakened its security through AI-generated ('vibecoded') code, but the claim itself is unverified here.

What HN thinks

  • Many commenters doubt the vibecoding attribution itself, noting the article offers no direct evidence of which tools or teams built the new site, and that similar sloppiness predates AI.
  • Commenters broadly agree the reported flaws, such as no 2FA on login, no confirmation for password changes, and one-click payments, would be caught by basic testing or security review regardless of how the code was written.
  • One commenter with firsthand experience said the cancellation flow used a dark pattern and that recurring payments were reinstated after they paused, though another said the pause was clearly described as one month.
Foundation Models frameworkdeveloper.apple.com · 256 points · 35 comments · 4daitoolsApple's Foundation Models framework is a native Swift API that lets apps call Apple's on-device model, Private Cloud Compute, or third-party models like Claude and Gemini through one interface.

Key points

  • Server models like Claude and Gemini are reachable through the same Swift API, with tool calling and guided generation, once a provider ships a Swift package conforming to the Language Model protocol.
  • Developers in the App Store Small Business Program with fewer than two million first-time downloads can use Apple Foundation Models on Private Cloud Compute at no cloud API cost.
  • Dynamic Profiles let a single session switch model, temperature, reasoning level, and toolset while sharing one transcript across switches.
  • A new Evaluations framework tests prompts and features across changing conditions, and an fm command line tool ships in macOS 27.

What HN thinks

  • Commenters point out that the framework was already discussed when announced last year, and one notes Apple may be preparing new local models, which could explain the renewed attention.
  • One commenter reports running a local Apple model on an M1 with 8GB of RAM for terminal error correction and command suggestions, finding it surprisingly capable for that use.
  • A caveat raised is that on-device model safety restrictions can cause refusals; one commenter had to stop sending full firewall events to a local model because it refused some requests.
Talking Zen 6 with Mike Clark at AMD In-Depthchipsandcheese.com · 229 points · 7 comments · 4dhardwareaiIn an interview, AMD's Mike Clark explains Zen 6 changes: the dense Zen 6c core gets full 4MB-per-core L3 and more IPC, and Zen 6 adds more VNNI units and wider vector loads.

Key points

  • Zen 6c's dense design now gets the full 4MB of L3 per core, so a 32-core CCD carries 128MB of L3.
  • The classic core becomes a high-frequency design, while the dense core becomes the top-performance option.
  • Integer multipliers drop from three to two because distributed schedulers replace the more unified Zen 5 scheduler.
  • Two extra VNNI units were squeezed in late in the design, despite the floor plan already being set.
Turning a 22-year-old Postgres type-confusion primitive into RCEmehmetince.net · 224 points · 41 comments · 4dsecurityprogrammingA researcher shows a type-confusion flaw in PostgreSQL core, dating to 2004, where altering a composite type's attribute after a prepared statement is built lets a crafted UUID be read as a bytea and leads to remote code execution.

Key points

  • Changing a composite attribute from uuid to bytea with ALTER TYPE after PREPARE, without rewriting stored bytes, makes byteaout() treat controlled UUID bytes as an expanded-object pointer and segfault.
  • The author chains this with an unnest() confusion on a bigint[] to leak a heap pointer, defeating ASLR, then forges a fake expanded-object method table pointing at system().
  • The mutable-row-type primitive dates to a 2004 commit, while the expanded-object technique used dates to 2015.
  • The fix, CVE-2026-6460, is an umbrella patch that adds composite attribute versioning, and the report covers 21 variant cases.

What HN thinks

  • Commenters disagree on severity: some argue it is really privilege escalation or authenticated arbitrary code execution rather than remote code execution, since an attacker needs a database login first.
  • Several note that the exposure is mainly a lateral-movement and multi-tenant risk inside organizations, with one pointing out that Postgres instances are sometimes exposed directly, as with a public OpenStreetMap query service.
  • One commenter argues the design of Postgres storing type identity separately from datums is fragile and hard to make safe, while another replies that compact representations are necessary for performance and that safety can still be preserved.
Our software engineering interviews are changing a littleblog.janestreet.com · 223 points · 212 comments · 4dcareeraiJane Street is trialing new software engineering interview rounds for campus onsites that emphasize software design, code verification, and using LLMs, while keeping most coding rounds LLM-free.

Key points

  • The new rounds began as a trial this summer in onsites for campus candidates; not every onsite includes them yet.
  • Most rounds remain traditional live coding interviews conducted without LLM assistance.
  • Candidates are not required to have a particular LLM workflow; the company says it wants clear thinking about what tools are good at and where they fail.
  • The company deliberately withheld details, saying formats are still evolving and may look different in six months.

What HN thinks

  • Commenters largely agree that git and GitHub PR mechanics are a weak thing to test in interviews, with some arguing LLM use for such mechanics is acceptable if the resulting commits and PRs meet team conventions.
  • Several commenters question whether interview methods are validated at all, noting that little evidence is shown that these process changes predict hiring quality, though others reply that long-term outcome data for new LLM-era formats cannot exist yet.
  • One commenter who runs interviews describes a live coding process where candidates need not finish, but must explain everything written, with LLM use judged by a higher bar for senior candidates.
Relax: JIT compiler for Rust to reduce build timesmakepad.github.io · 171 points · 95 comments · 5dtoolsprogrammingRelax is an experimental live compiler and JIT for Rust that runs a Cargo package from source, compiling functions on first call and patching edits into the running process.

Key points

  • A large 3D game went from clean checkout to first frame in 11 s on an M4 Mac, versus about 3 minutes for cargo and rustc.
  • Relax's optimised code runs slower than rustc --release: geomean 1.12× time on x64 and 1.24× on arm64 across 21 workloads.
  • Function body edits patch live with state kept, but changes to struct layouts, imports, traits, impl headers or macros need a restart.
  • The code cache in target/makepad-run takes 120 to 190 MB for a hello-world package, and JIT code is append-only in memory.

What HN thinks

  • Commenters see a trend of several AI-assisted Rust iteration tools appearing within a month, and some argue the rapid build of such ambitious projects is what AI agents now make possible.
  • Several commenters tie the idea to older live-programming traditions such as Lisp, REPLs and Smalltalk-style environments, and argue that Rust's slow builds are a matter of language design and priorities rather than the borrow checker.
  • The main criticisms are that the project is an AI-generated re-implementation with questionable maintainability, that moving errors from compile time to runtime is a real trade-off, and that the website looks like low-effort AI output.
AI is turning Apple's app review process into a nightmare for synth developerssynthanatomy.com · 162 points · 172 comments · 4daitoolsSynth developers report iOS App Store submissions stalling for weeks, with some apps rejected repeatedly by what appears to be automated review, including AI-hallucinated rejection reasons.

Key points

  • Apple's own guidelines say at least 50% of submissions are reviewed in under 24 hours and 90% in under 48 hours, but developers report weeks-long waits.
  • Unfiltered Audio's Sinusoidal Run Rhythm was rejected four times, including a reviewer allegedly hallucinating that a term needed a documented license.
  • Marek Dendes's CRUX Synth was rejected four times within seconds of entering review, and his appeals to the App Review Board got no reply after about five days.
  • Developers say the iOS music app market now sees an estimated 10 to 20 new apps per day, driven by AI-assisted vibe coding.

What HN thinks

  • Commenters broadly blame Apple's review capacity rather than AI alone, pointing to the roughly 500 reviewers cited against over 100,000 apps weekly and noting Apple's large App Store revenue.
  • Several developers report that wait times and rejections have grown, with one saying a significant update took about 80 hours versus under 12 in the past, and others say reviews now complete within a day.
  • Commenters attribute the slowdown mainly to a flood of vibe-coded and template-based apps, and one noted that Steam review times have also grown from days to weeks for similar reasons.
Data centers need power – lower income, rural residents pay the pricenews.cornell.edu · 158 points · 161 comments · 4dpolicyaiA Cornell study finds data centers paid only 38% of the infrastructure costs tied to their growth across four major U.S. grids in 2025, shifting the remaining 62% largely onto residents.

Key points

  • Of an $8.25 billion infrastructure bill in 2025, data centers paid $3.12 billion and households carried $5.13 billion.
  • Lowest-income households saw extra energy costs of up to 28% of their income.
  • Wholesale prices rose nearly 10% or more in the most affected hubs by 2025, reaching 15% in Virginia and 20% in southern Bay Area.
  • Cost allocation often follows past consumption, so new data centers that have not yet drawn power escape the upfront costs.

What HN thinks

  • Commenters broadly agree that the permitting and regulatory regime, not data centers alone, constrains supply, and that a bipartisan permitting bill (BAAJA) is being debated, with critics warning it could weaken protections for historic sites and tribal lands.
  • One commenter argues the study's method is misleading because it allocates transmission upgrade costs by peak demand share rather than by the actual cost of upgrades, and notes large customers are often billed for their own upgrades.
  • Several commenters say local benefits such as jobs and tax revenue are limited, with construction-phase employment fading and tax breaks reducing the payoff, while others note Australia charges connecting loads for upstream upgrades.
Nvidia CEO Jensen Huang resigns after reports of second affair deepened scandalbbc.co.uk · 158 points · 84 comments · 4dbusinessaiA BBC live page reports that Nvidia CEO Jensen Huang has resigned amid reports of a second affair, but the article itself could not be retrieved.

What HN thinks

  • Commenters largely tie the resignation to the reported China chip sales and a suspected spy connection, not the affairs alone, with one saying the affair reports made the earlier chip-sales concerns public.
  • Several commenters doubt the framing, noting that 'stepping down' is often a forced exit and that the affairs with a direct report may have mattered more than the scandal itself.
  • Some dispute the cultural explanation, arguing American reaction was driven by the foreign spy link rather than general prudishness about affairs.

The article couldn't be read, so this summary is based on the discussion.

Singapore Says U.S. Not Helpful on Investigation into Suspected Spy's Companynytimes.com · 139 points · 50 comments · 4dpolicyhardwareSingapore says U.S. authorities have not been helpful in its investigation into a suspected spy and the company linked to that suspect, reportedly MegaSpeed, which is tied to Nvidia.

What HN thinks

  • Commenters point out that Singapore's main complaint is that US authorities did not compel Supermicro to cooperate fully, and that DOJ already has its own charges and subpoenas, so legal limits or focus on its own case may explain the gap.
  • Several commenters suspect political favoritism, citing Nvidia's donations to Trump and the Wyden and Neal investigation into an alleged scheme taking a 25% tariff cut on Nvidia chip sales to China.
  • One commenter with technical knowledge notes that Huawei Ascend chips are already used for inference by some Chinese model providers, though training still relies on Nvidia hardware and CUDA.
Libreagi – Save OSS Modelslibreagi.com · 136 points · 25 comments · 4daiopen-sourceLibreAGI is an invite-only BitTorrent tracker where members seed copies of open-weight AI models, with seeding credit weighted toward models that have few copies.

Key points

  • Credit grows with file size and seeding time, and a model with only one copy earns four times the base rate.
  • Only safetensors and GGUF files under Apache-2.0, MIT or the Qwen Community License are accepted.
  • Each file is checked against its source commit, and staff review every model before it is listed.
  • A signed public catalog lists every model, hash and source so swarms can be rebuilt if the site disappears.

What HN thinks

  • Commenters questioned the invite-only gate and login requirement, asking what it protects against since anyone could otherwise seed open-weight files.
  • One commenter asked why only safetensors and GGUF are accepted rather than ONNX, which is commonly used for interchange.
  • A reply suggested the restriction serves accountability, but it was unclear to others what threat, such as swarm poisoning, the gating would actually address.
Making a €40 USB-C hub work as a Nintendo Switch 2 dockblog.nozomi.space · 127 points · 44 comments · 7dhardwaretoolsA €40 Cable Matters USB-C hub was reflashed so a Nintendo Switch 2 treats it as the official dock, enabling TV mode and HDR over HDMI while charging.

Key points

  • The Switch's docking handshake is Nintendo vendor-defined USB-PD messages (SVID 057E), and the hub's PD controller had to answer them with the official dock's identity and version records.
  • Returning only Nintendo's vendor ID was not enough; the Switch also checks the full identity, and the first working version needed the dock's complete identity data.
  • HDR looked washed out because the HDMI AVI InfoFrame was labeled BT.709; a 108-byte patch in the RTD2142 marks it BT.2020 whenever HDR metadata is present.
  • VRR was not made to work: enabling it blanked the screen once frame timing varied, and the author disabled VRR in Switch settings.

What HN thinks

  • Commenters debate why Nintendo deviates from standard USB-C, with some arguing it is a licensing or accessory-control choice and others saying the console could simply refuse non-matching chargers with a message.
  • One commenter notes that the Switch 2 charges from any standard USB-C PD charger, and another says the handshake does not block standard PD chargers, so the proprietary behavior is limited to the dock path.
  • A commenter reports that the Switch 1 had a similar quirk where most USB-C hubs failed, with only one obscure hub working, and the author says the Switch 1 uses a different but related handshake.
Show HN: AgentRun – open-source, Self-improving Managed Agentsgithub.com · 125 points · 29 comments · 4daitoolsAgentRun is an Apache-2.0 platform that wraps existing agent harnesses like Claude Code, Codex, OpenCode and pi with self-improving learning, crash-durable runs, approvals and spend caps.

Key points

  • After each run, the agent proposes notes, skills or workflows, which code checks and one grader reviews before committing them to a per-agent git folder.
  • Learned skills that serve 3 runs can be converted into a workflow draft, which is dry-run without a model or tools before being kept.
  • Runs survive server restarts from one store, and a dead runner's runs are picked up by another after a 30-second lease expires.
  • Interrupted writes are marked interrupted rather than retried, so they can be checked manually.

What HN thinks

  • One commenter says that for teams on subscriptions, a local orchestrator handling PR reviews is far cheaper and higher quality, and points to their own open-source tool for this.
  • Several commenters say PR review is now a major bottleneck and cost driver, and one shares a code-review agent template the author reports cut review costs by about 80 percent.
  • One commenter asks about adding a proprietary corporate registry data source with about 200 government integrations and suggests MCP-based plugging, which the author says is planned.
ExcalidrawZ: Local Native Excalidraw on macOS/iOSgithub.com · 117 points · 26 comments · 4dtoolsdesignExcalidrawZ is an MIT-licensed native SwiftUI client for Excalidraw on macOS, iPadOS, and iOS, adding file management, iCloud sync, and an AI drawing assistant.

Key points

  • Drawings are stored as app documents with iCloud sync, per-file history checkpoints, and support for .excalidraw, .excalidraw.png, and .excalidraw.svg files.
  • Locked files and their checkpoints are encrypted, with a Recovery Key for fallback access and content hidden from AI.
  • The app can run as an MCP server over a local HTTP endpoint, with a basic and an optimized mode.
  • Math features include LaTeX formulas and function graphs, and the Excalidraw core it uses is open source.

What HN thinks

  • Commenters pointed out that an Emacs package, excali-mode, already offers Excalidraw editing, so this is not the only alternative.
  • Several commenters questioned the name, noting it closely borrows Excalidraw's branding, and one said other projects do the same.
  • One commenter raised concern about trusting the app with API keys, since the AI assistant requires them, and noted the app is about four years old.
I gave an agent a lakehouse and control of my gardenjatorre.github.io · 114 points · 55 comments · 8daitoolsA hobbyist ran a two-square-metre vegetable bed through a LakehouseBox data lake, with a hosted agent that reads eight soil probes and a forecast each morning and adjusts sprinkler minutes via an old phone.

Key points

  • Eleven Ecowitt sensors and a camera post straight into Iceberg tables, with no middleware, accumulating about four million readings since 18 September in 3.22 MB.
  • Cameras that only speak FTP were handled by adding an FTPS drop box that indexes each JPEG into a queryable images table.
  • The camera missed caterpillar damage on leaves; close-up phone photos revealed the frass, showing full-frame views are too coarse for small pests.
  • The watering actuator is an old Android phone tapping the official Rain Bird app, capped at 20% change per run and 1 to 15 minutes.

What HN thinks

  • Several commenters found the fully automated setup less appealing than hands-on gardening, arguing that automating watering removes the intuition and satisfaction people get from tending plants themselves.
  • One commenter with experience noted that the author's post seems heavily AI-written and asked for disclosure, and others said the writing was grating to read.
  • A practical lesson from one commenter: give physical-control agents a small set of named actions with hard limits, a log of past actions, and daily caps on each action, so repeated or misread decisions become harmless no-ops.
I'm Working on a FOSS Linux Pinbal Simulation Enginegetlazarus.org · 104 points · 35 comments · 5dtoolsopen-sourceA developer is building a FOSS Linux pinball simulation engine with table logic in TypeScript, tables designed in OnShape CAD, and an in-game editor, aiming for an initial release with three demo tables.

Key points

  • The engine imports tables as 3D meshes from parametric CAD, simplifying many physics bodies into primitives with configurable mesh conversion options.
  • A new sound system attaches multiple sounds with individual volumes to parts, and ambient sounds are attenuated by applied force.
  • Recent work added multiple camera modes including a portrait mode, script-controlled cameras, and an Asteroids video mini game on the demo table.

What HN thinks

  • Commenters raised compatibility questions, asking whether the engine can load Visual Pinball tables, since most existing pinball tables are made for that platform.
  • One commenter asked how physics is handled, noting that the author's Tiny.Sim library was described as 2D, and suggested pinball is a good demo for physics engines.
  • Several commenters expressed interest in Linux-native pinball games, and one asked for force feedback support via DualShock controllers and another for dual-screen machine support.
Researchers discover rare star system that emits radio signals every 8 minutescurtin.edu.au · 102 points · 53 comments · 5dspacescienceAstronomers link a long-period radio transient, GPM J071507−114153, to a white dwarf orbiting a Sun-like G-type star, the first such pulsing pair detected.

Key points

  • The white dwarf spins every eight minutes and its orbital motion is visible in the radio pulses, suggesting the emission comes from near the white dwarf itself.
  • Previous long-period transients with white dwarfs were mostly paired with cooler red dwarfs, making this a new configuration.
  • The system was found in an MWA survey of the Milky Way visible from Australia, then localized with MeerKAT and confirmed optically by Keck Observatory.
  • The team processed more than 100 hours of radio data at Pawsey Supercomputing Research Centre to validate the white dwarf motion.

What HN thinks

  • Commenters note the headline is misleading: the source emits continuously, and the 8-minute period comes from a beam sweeping past Earth as the object rotates.
  • Several commenters point out that 8 minutes is an unusually long period, far slower than typical pulsars, which often spin hundreds of times per second.
  • One commenter says the title's wording confuses emission with detection, a distinction they see as central to understanding the object.
Adapting CPython for Yktratt.net · 74 points · 5 comments · 4dprogrammingsystemsLaurence Tratt adapted CPython 3.15.1 to the yk meta-tracing JIT framework in 3-4 days, adding about 450 lines, and reports it beating stock CPython on most of 13 benchmarks.

Key points

  • ykcpython was built on CPython's tier 2 interpreter after the tier 1 version suffered trace invalidation from quickening transitions.
  • Making opcode loads idempotent cut the Mandelbrot benchmark from 13.7s to 5.0s, reducing HIR output from 20,772 to 7,648 lines.
  • On DeltaBlue, adding a YkLocation at each executor's first opcode raised traced execution from about half to most of runtime, a 4.23x speedup.
  • Results are wall-clock on 30 process executions; ykcpython is 2.5% slower than CPython on DeltaBlue and the author says it is not ready for real use.
PostgreSQL 18.7, 17.12, 16.16, 15.20, 14.25, and 19 RC 2 Releasedpostgresql.org · 73 points · 8 comments · 4dprogrammingsecurityPostgreSQL releases update versions 18.7, 17.12, 16.16, 15.20 and 14.25 fixing 52 security vulnerabilities, and ships the second release candidate of PostgreSQL 19.

Key points

  • The release fixes 52 CVEs, many rated CVSS 8.8, mostly memory-safety bugs such as use-after-free, type confusion and buffer overflows enabling arbitrary code execution.
  • PostgreSQL 14 reaches end-of-life with this release and will receive no further security or bug fixes.
  • Several fixes may require reindexing, including GIN indexes, pg_trgm GiST indexes, and B-tree indexes on bpchar columns without a length.
  • Incremental backups may be corrupt from earlier versions, so the project recommends taking a new full backup after updating.
Did Life Begin in an RNA World?quantamagazine.org · 73 points · 45 comments · 7dsciencehistoryQuanta explains the RNA world hypothesis for life's origin and reviews evidence for and against it, including a 45-nucleotide self-copying RNA called QT45.

Key points

  • Gerald Joyce's 2019 lab evolution of ribozymes found most mutations destroyed catalytic activity, so self-sustaining RNA evolution appeared hard.
  • Spontaneously assembled RNAs found so far are about 10 to 15 nucleotides long, while Joyce's ribozymes were several hundred.
  • Holliger's team found a 45-nucleotide RNA, QT45, that can copy another RNA strand, and two other small active RNAs in a pool of about 10 trillion molecules.
  • Holliger's team showed RNA replication driven by cycles of acidity change and freezing and thawing, a scenario suited to icy poles.

What HN thinks

  • A recurring objection is that prebiotic chemistry yields a messy soup of junk rather than the pure activated monomers that RNA experiments rely on, and commenters say this gap is not addressed by lab results.
  • Commenters note that the hard step is going from occasional replication to a self-sustaining system, and that RNA's fragility and need for protocell compatibility make this harder still.
  • Counterpoint from commenters: combinatorial-explosion arguments may be too reductive, since cycles of cooperating catalysts across varied environments could build an evolutionary ratchet before any single perfect replicator forms.
Bad ideas that aren't bad anymoredevonestes.com · 70 points · 53 comments · 5daicareerA engineering leader describes four previously 'bad' practices his team now uses successfully because AI lowered their costs: skipping human PR review, splitting the team into separate streams, de-emphasizing deep expertise, and rewriting or building in-house.

Key points

  • Engineers open PRs with Claude and the human owning the ticket reviews them, merging once CI is green and Greptile reports clean, a practice running about one month.
  • Getting up to speed on a colleague's stream now takes a few minutes of prompting Claude, which the author says cut the cost of context switching.
  • The author says rewrites that take days or weeks instead of years avoid the main risk of rewrites, since the old system no longer changes underneath them.
  • He now recommends engineers prioritize fundamentals like abstraction design and domain modeling, plus product and user research, over deep framework expertise.

What HN thinks

  • Several commenters argue that code review still matters, citing edge cases, architectural drift, and shared understanding that AI-generated code does not guarantee.
  • Skeptics note that a single month of data cannot show long-term maintenance costs, and that gains may come from eating technical debt.
  • One commenter with experience suggests the real value of review has shifted toward architecture and design decisions, and that reviewing agents' conversations or decision logs could replace some lost learning.
Running web sites makes me not want tolowkey.party · 68 points · 41 comments · 4dculturecareerA personal blog post, unavailable to fetch, that says running websites makes its author not want to keep doing it.

What HN thinks

  • Commenters broadly agree that bot and scraper traffic is the main burden of running a site, with one reporting traffic shifting from cloud IPs to residential proxies after blocks, so blocking becomes endless whack-a-mole.
  • Several practical tips came up: serve mostly static pages behind a CDN, return fake 200 or 404 responses instead of explicit 403 blocks so bots don't retry elsewhere, and tarpit or rate-limit suspected bots.
  • One commenter argued that proof-of-work challenges punish legitimate customers, especially outliers hit by repeated challenges, and that operators rarely track that collateral impact.

The article couldn't be read, so this summary is based on the discussion.

Nvidia CEO Jensen Huang resigns after reports of extramarital relationshipsnbcnews.com · 68 points · 16 comments · 4daibusinessJensen Huang resigned as Nvidia CEO after reports of affairs, including one with a businesswoman alleged to be a Chinese spy. CFO Colette Kress is interim CEO.

Key points

  • Nvidia named its first chairman, Harvey Jones, a former Synopsys CEO, a governance change for a company that had long relied solely on Huang.
  • Shares fell more than 11% after the board's announcement, following an 11% drop earlier that same week.
  • The board hired a former acting U.S. attorney and his law firm to investigate whether Huang was targeted by a foreign intelligence plot.
  • Alice Huang held a senior role at Megaspeed, a Singapore firm under U.S. criminal investigation over alleged chip smuggling to China; the two Huangs are not related.

What HN thinks

  • Commenters largely focused on whether the affair angle matters, with one arguing the press emphasizes it for clicks while the real concern is export controls and conflicts of interest.
  • One commenter contrasted the response with Elon Musk's controversies, arguing that Huang was treated more harshly for personal conduct.
  • A shareholder-minded commenter argued Huang should be judged mainly on value creation, citing his long-term investments in Tegra and CUDA despite earlier write-offs.
Fitting Four 32 GB Resizable Bars Behind a PCIe Switch on Linuxdoug.sh · 65 points · 23 comments · 4dhardwaresystemsA Linux kernel module sets 32 GB Resizable BARs on four Radeon AI PRO R9700s behind a PCIe switch, working around a kernel window-sizing bug that left most cards without them.

Key points

  • Linux 7.0 sized the switch's memory window as the plain sum of child windows, leaving a gap of almost 32 GB per card and only two of four R9700s got their BARs.
  • The module raises each smaller 64-bit prefetchable BAR's alignment to the largest on its bus segment, making each R9700 window exactly 64 GB.
  • Resizing and rescanning must happen from the initramfs before GPU drivers load, because amdgpu cannot grow a window other cards already occupy.
  • With P2P enabled, every R9700 pair reached 14.0 GB/s and a four-card ring moved 54.7 GB/s.

What HN thinks

  • One commenter asked why 48-bit physical addressing (hundreds of terabytes) still causes a problem, suggesting the limit may come from other hardware or firmware rather than CPU address width.
  • Another commenter argued the underlying task is trivial, just writing values to registers, and that the difficulty comes from hidden firmware and BIOS layers standing in the way.
  • The author noted that P2P between the cards roughly halved prefill speed when disabled, which is the concrete payoff for doing the workaround.
AWS Lambda Web Functions keep billing memory ~7 minutes after the last requesteasel.sh · 61 points · 28 comments · 4dculturesystemsEasel.sh dropped AWS Lambda Web Functions for its own servers after finding Web Functions kept memory billed 5–7 minutes past the last request, with no configurable idle timeout.

Key points

  • Six measured environments stayed alive 313 to 411 seconds after their last request, and the processes remained active while idle.
  • Memory accounted for about 88% of the calculated compute and request costs in Easel's experiment, despite a 160–210 MB application footprint.
  • Easel reports 1.5–3× faster median warm response times and about 3.5× faster cold starts after 12 minutes idle, from a single Ashburn client.
  • Cold-start results rest on only two paired trials, and the platforms ran on different hardware, so the comparison is not normalized.

What HN thinks

  • Commenters largely agree that Lambda environments stay alive for several minutes after the last request, and one noted this is undocumented as a specific duration, though AWS documents that memory is billed while the environment is active.
  • The main caveat is that the cost only bites for low-traffic or bursty workloads, since sustained load reuses environments and amortizes the idle time, as the author also conceded.
  • Some commenters argue that the idle billing is a poor fit for rarely invoked functions and that standard Lambda event functions bill only for request duration, making them cheaper in that case.
Words of warning about radiocarbon dating medieval human bonesnature.com · 60 points · 12 comments · 5dsciencehistoryResearchers report that freshwater fish diets make radiocarbon dates of medieval Danish monastic bones appear centuries too old, with six of 27 samples predating their monasteries.

Key points

  • Bone collagen with δ15N above about 8‰ signals substantial freshwater protein intake and a possible reservoir effect that can skew radiocarbon ages.
  • Freshwater reservoir offsets vary from zero to over 1000 years, unlike the roughly uniform 400-year correction used for marine diets.
  • Six of 27 Danish monastic burials (about 22%) yielded dates predating their institutions, including one about 200 years before Viborg Maria's founding.
  • The authors say measuring δ15N first is the only viable way to assess whether a medieval skeleton's radiocarbon date needs a reservoir correction.

What HN thinks

  • Commenters note the freshwater reservoir effect on radiocarbon dating is already well known, and one says it has appeared in popular archaeology books for years.
  • One commenter who works with radiocarbon labs says the effect is routinely studied and accounted for, and suggests non-radiocarbon methods such as electron spin resonance dating for human remains.
  • One commenter had expected the effect mainly from marine sources and was surprised it matters meaningfully in freshwater systems.
Why You Should Expect an AI Crasheconomist.com · 56 points · 30 comments · 4daibusinessAn Economist leader argues that an AI market crash should be expected, but the full text was unavailable, so its specific reasoning cannot be verified here.

What HN thinks

  • Commenters mostly could not assess the Economist's argument because the full text was unavailable, so the thread is largely about access rather than substance.
  • One commenter jokingly observes that wide talk of an AI crash suggests it is already priced in, while others point to excessive optimism as the reason a crash is likely; the thread offers no evidence either way.
Electric ship batteries promise compact, cooler powerspectrum.ieee.org · 53 points · 60 comments · 4dhardwareenergyXing Mobility unveiled Immersio CTP, a marine battery pack whose cells are fully submerged in a flammable coolant, claiming about twice the energy of conventional marine packs.

Key points

  • Lab tests show about 200 Wh/kg at pack level, versus 168 Wh/kg listed for Corvus Energy's Dolphin NxtGen, a 19 percent edge measured only in the lab.
  • The doubling claim is measured against LFP packs, which store less energy per kilogram than the nickel-based cells the CTP uses.
  • Xing claims 330 to 350 Wh/L, against 212.5 Wh/L for Corvus, aided by omitting module boxes, cold plates and cooling pipes.
  • Xing says immersion can cool a punctured cell from about 1,000 °C to below 100 °C within 10 seconds, but has not released the test results behind that figure.

What HN thinks

  • Commenters broadly agree batteries cannot practically power long-haul ocean container ships given energy density, weight, and cost, with ammonia or other fuels seen as likelier decarbonization routes.
  • Several note batteries suit short-route ships such as ferries, feeder ships, and river transport, citing Washington State Ferries' partial battery hybrid refit as an example.
  • One commenter raises a safety concern that fire at sea is extremely dangerous, since abandoning ship in open water can be deadly, which undercuts the idea that abundant water makes large battery packs safer.
Show HN: Fable52.games – Play and inspect Fable 5.2 single-file HTML gamesfable52.games · 51 points · 45 comments · 4daigamingFable52.games is a 32-game arcade of browser games, mostly single HTML files, that Anthropic's Claude Fable 5.2 built autonomously from one prompt each.

Key points

  • Featured Saltline was built in about 8 hours and is a 1.4 MB single file with zero external assets.
  • Games are self-contained, with Three.js and WebGL for 3D, Web Audio for sound, and levels and audio generated in code at runtime.
  • Each game lists its original prompt beside it, and source can be inspected or downloaded as a .html file.
  • Some games are hosted on creators' own sites in sandboxed iframes, while others are served from the arcade itself.

What HN thinks

  • Commenters largely enjoyed the output quality, with the BLAMELESS game singled out as fun and humorous given the tiny prompt, though some called it derivative.
  • Several commenters questioned the value of AI-generated low-effort games, and one former browser game developer said the capability highlights what skilled, hand-built game development has lost.
  • Practical problems raised include the site going down under traffic, a mobile joystick bug in the Saltline game, and many prompts being missing, which one commenter felt undermined the site's purpose.
Lean Grimoireterrytao.wordpress.com · 51 points · 8 comments · 5dtoolsmathTerence Tao rebuilt his 2018 browser textbook QED as "Lean Grimoire", a proof-building logic game with an option to show an in-progress Lean proof.

Key points

  • The game teaches propositional and first-order logic by applying rules to reach a conclusion from hypotheses, in the style of a crafting game.
  • A new fantasy mode renames game elements, so mathematical statements appear as potions and expressions as scrolls.
  • Tao built the rebuild with a modern coding agent and is asking for playtesting to improve its use for teaching proofs and Lean.

What HN thinks

  • One commenter found the post hard to parse, calling the introduction LLM-sounding and the controls like an unfinished personal note.
Show HN: Kapso – Iterative learning from enterprise data49 points · 11 comments · 4daiopen-sourceKapso is an open-source engine that runs experiments on a task, verifies results with a separate verifier, and stores learned lessons in a git-based bank that can be shared and A/B tested.

Key points

  • Lessons carry a credence score, conditions on where they apply, and supporting evidence, and ideation re-ranks proposed ideas by expected return against budget.
  • A verifier module that experiments cannot access checks results, and experiments can propose changes to the success metrics for approval.
  • Updated lesson banks can be A/B tested against previous banks using otherwise identical builds.
  • The author claims first place on RelBench predictive modelling and first on OpenAI's MLE-bench at submission time.

What HN thinks

  • A commenter reports using Kapso for hyperparameter optimization on RelArena across 21 structured tasks, saying it beat TabPFN on most larger tasks with about 1% of its parameters and no in-context learning.
  • Another commenter argues RelArena is the more credible benchmark because it reruns submitted implementations under shared protocols, whereas the author's RelBench leaderboard results rely on self-reported predictions.
  • Several commenters note the name clashes with an existing popular WhatsApp integration product called Kapso, and the author says a rename is being considered.
Legocam: My Lego webcam from 2000, working on a modern Macblog.benjamin.ie · 40 points · 29 comments · 7dhardwareopen-sourceA user-space macOS driver now lets a 2000-era LEGO Studios webcam, a Logitech QuickCam Web inside, work in Zoom, FaceTime and browsers without a kernel extension.

Key points

  • The camera's STV0610 bridge and VV6410 sensor were supported only by the Linux gspca_stv06xx driver, whose register protocol the author copied.
  • A libusb command reads isochronous frames and writes raw 356x292 Bayer frames to stdout, with no kernel extension required.
  • A second program pushes frames into the signed OBS Virtual Camera via CoreMediaIO, working around the need for a paid Apple developer signature.
  • Limits: about 5 fps due to USB 1.1, no auto exposure, microphone and snapshot button not working, and apps show it as OBS Virtual Camera.

What HN thinks

  • Commenters pointed out that OBS's signed virtual camera can push frames from arbitrary apps, so Apple's signed camera extension requirement offers less protection than intended.
  • One commenter read the requirement as a money-making measure rather than a good-faith security step, while another criticized Apple's fee for signing one's own software.
  • A commenter recalled using the webcam as a spy car on a LEGO RC car, a reminder of its original MovieMaker stop-motion purpose.
Show HN: Build a PC in 3D in the Browserbuildcores.com · 39 points · 13 comments · 8dhardwaretoolsBuildCores is a browser-based 3D PC builder with over 5,000 models, live price comparison across retailers, and compatibility checks, built by a former Google Earth engineer.

Key points

  • Compatibility checks flag issues like CPU socket, RAM type, GPU length, and cooler height as each part is added.
  • Game FPS estimates are labeled as illustrative rather than live benchmark results.
  • A Pro tier removes ads while building and unlocks AR, airflow, and estimated FPS; a desk setup configurator is slated for October 2026.

What HN thinks

  • Commenters asked for airflow modeling, and one raised the desire for a parts database that ships with RGB LEDs off by default, citing the annoyance of default lighting.
  • One commenter with direct experience described RGB RAM whose control software was buggy and failed to turn the LEDs off, so they removed the LED power traces.
  • Several commenters praised the visual approach as useful for people building their first PC, though one called the interface cluttered.
Slavdom Unearthed: Marian Wawrzeniecki's Neopagan Paintings (1891–1923)publicdomainreview.org · 37 points · 9 comments · 8dhistorycultureMarian Wawrzeniecki was a Polish archaeologist who also produced a large body of neopagan paintings and illustrations of Slavic ritual, mostly from 1891 to 1923.

Key points

  • He discovered the first obsidian artifacts on Polish lands and early earthen shelters, ceramics, and steles, and his archaeology shaped his painting.
  • Painted in Kraków's academic history-painting style, his early oils later gave way to bright, textbook-like illustrations in his final decades.
  • Recurring motifs include faceless female nudes tied to trees or pillars and a basilisk, which critics at the time compared to pulp erotica.
  • Maria Janion's 2006 book Niesamowita Słowiańszczyzna drew on his speculative theories, and Stanisław Szukalski is named as an artistic descendant.

What HN thinks

  • One commenter responded to the piece mainly with enthusiasm for the kind of obscure material the site surfaces, with no substantive debate about the artist.
Distributed Ray Tracerray.hatchet.run · 36 points · 12 comments · 4dsystemstoolsA web page presented as a distributed ray tracer, though its contents could not be retrieved, so its implementation details remain unverified.

What HN thinks

  • Commenters point to signed distance field resources (Inigo Quilez's articles and a Zed editor write-up) as background for the rendering approach.
  • The main concern raised is cost: one commenter worried that a public web endpoint could run up large Cloudflare bills, and the author said infrastructure credits cover it.
  • One commenter questioned whether a GPU would be simpler, and the author replied that GPUs are costly and the project is mainly an exploration, with GPU rendering being a relatively easy swap in the Hatchet setup.
The Learjet Repo Man (2009)salon.com · 36 points · 12 comments · 8dbusinesscultureSalon's 2009 profile covers Popovich, a former Braniff pilot who makes a living repossessing aircraft, flying seized planes out of the hands of owners who have defaulted.

Key points

  • The article is unavailable, so details here come only from the comment thread, which links related profiles of repo men who recover seized ships.
  • Commenters note the story was picked up again later, with a San Diego Union-Tribune piece from September 2009 following the Salon interview.

What HN thinks

  • Commenters pointed to related repo stories, including a profile of Max Hardberger who recovers stolen cargo ships, and a memoir he wrote about the work.
  • A commenter with a family history in Australian sheriff's work described a similar lien-enforcement job, holding a supertanker in port for days until fees were paid.
  • One commenter questioned whether a repo pilot could fly an unfamiliar airliner type on short notice, noting that type ratings and preparation normally take days.

The article couldn't be read, so this summary is based on the discussion.

How to make a CNC plasma cutter with high-frequency startincoherency.co.uk · 35 points · 7 comments · 8dhardwaretoolsA hobbyist converted a £80 Lidl-sold Parkside plasma cutter with high-frequency start into a CNC machine by running FluidNC on an ESP32 over WiFi to avoid USB disconnects from arc interference.

Key points

  • Igniting the HF pilot arc disconnected the USB serial link to the Grbl controller, but Grbl kept executing buffered gcode without interruption.
  • The ESP32 outputs were level-shifted to 5V with a 74HCT245 on a homemade stripboard breakout, which the author admits he is unsure is legitimate.
  • Leadshine easy servos ran on a 12V supply despite documentation calling for 18V, and their alarm output falsely triggered at every arc start.
  • The machine's frame is box section steel with 608 bearings running directly on the steel and printed GT2 pulleys, with noticeable play and no adjustment.

What HN thinks

  • One commenter suggested that testing USB cables with a multimeter for good end-to-end shield continuity might solve the disconnect problem without switching to WiFi.
  • A commenter asked whether kerf width or taper had been tuned, noting that plasma's typical weakness is precision compared with its speed.
Pulling the Lever in a Subway Problemjefftk.com · 32 points · 35 comments · 5daicultureA blogger who pulled a subway emergency brake when a passenger's arm seemed caught in a door now thinks it was probably only a bag, and uses the episode to test philosophical dilemmas against reality.

Key points

  • Witnesses inside the car disagreed about whether a hand was in the door, since the passenger was holding a bag and the view was unclear.
  • The author estimates a one-hour delay for about 25,000 riders equals roughly 4.3 years of waking life in total.
  • Pulling the brake may not have avoided delay, because an injury would likely have prompted someone else to pull it anyway.
  • Pointing at one specific bystander and asking for help did not work; the author then ran down the train to pull the brake personally.

What HN thinks

  • Several commenters argued the arm-in-door scenario is not a deep dilemma, since someone shouting that a limb is trapped would normally prompt an immediate brake pull without moral calculation.
  • One commenter criticized the blogger's life-years arithmetic, arguing that minutes of waiting on a train are not equivalent to losing that time from one's life.
  • A commenter pointed to reexaminations of the Kitty Genovese case, suggesting the bystander-effect framing popularized by that story is less accurate than commonly believed.
Safety in Algae Home Experimentschillphysicsenjoyer.substack.com · 31 points · 3 comments · 6dsciencehealthA home hobbyist argues that three algae strains (Chlorella vulgaris, Scenedesmus quadricauda, Dunaliella salina) are low risk to culture at home, relying on supplier hazard ratings and sparse case reports.

Key points

  • Chlorella vulgaris is sold to high school suppliers and is absent from the UK HSE pathogen list, which the author treats as proxy evidence of safety.
  • The author's tally of chlorella illness reports is low and mostly tied to specialised circumstances, such as a 1994 Singapore factory asthma case.
  • CCAP and ATCC both self-rate Scenedesmus quadricauda and Dunaliella salina as their lowest hazard class, with the author checking the LLM-scraped ratings manually.
  • The author identifies contamination with other organisms as the biggest remaining risk, and says supplier trust is the only check before microscope inspection.
Darling Development Update – Q4 2026blog.darlinghq.org · 28 points · 6 comments · 5daiopen-sourceDarling, the project running macOS apps on Linux, adopts an AI-assisted contribution policy and reports initial ARM64 support, with forks merged back and three architectures planned.

Key points

  • The policy makes contributors fully responsible for every line of AI-assisted code and requires them to explain it in their own words.
  • Clean-room rules bar leaked material, Apple restricted source like corecrypto, and disassembly-based matching code.
  • The post cites macOS 27 with 464 public and 2,754 private frameworks, versus 273 and 878 in macOS 10.14, as a scale problem.
  • Agent-driven merge logic has reportedly fixed many bugs in forked code and pre-existing Darling bugs.

What HN thinks

  • One commenter argues AI coding works best where a detailed executable oracle exists, and macOS app behavior and game porting fit that well.
  • Another commenter questions how the project can review PRs at its ambitious pace within two months.
Stingray Fish Laserstingray.no · 28 points · 26 comments · 7dhardwarescienceStingray describes a submersible unit that uses stereo cameras, software, and lasers to shoot sea lice on farmed fish while the fish stay in the pen.

Key points

  • Lasers fire up to five pulses per second, around the clock, aimed at individual lice on fish skin.
  • Blue 450 nm LED lighting is used to boost contrast between fish and sea lice, working from surface to 25 meters.
  • Each unit has a 360-degree thruster for remote heading control, and fish farmers see live camera streams via the Stingray Online portal.
  • The company says the units are designed and manufactured in Norway.

What HN thinks

  • Commenters questioned whether lasers can work well underwater, since water attenuates light, though the thread offers no first-hand answer on this.
  • One commenter raised the prospect of parasites evolving resistance to the laser approach, and another joked about switching to blue lasers.
  • Much of the thread focused on the website's poor scrolling and layout rather than the technology itself.
Show HN: AI Red Lines – write down what would make you act on AIyourairedlines.com · 26 points · 15 comments · 4daipolicyA web tool lets people privately write down AI-related red lines, share them to a public wall, and set up a witness, reminder, and news check for crossed lines.

Key points

  • Red lines stay in the browser until shared, and sharing puts them on a public wall with a random ID.
  • The site cites a 2016 meta-analysis of 138 trials, where tracking a goal helped more when others could see it (0.55 vs 0.19 private).
  • Users can name a witness who gets an email with every finished line, including private ones.
  • Nearly all the underlying trials concerned health goals, and the site itself admits no one has tested red lines for AI.

What HN thinks

  • Commenters questioned how the term "safety" is defined, with one saying the site's concrete examples are clearer than its vaguer wording and asking for a clear definition.
  • Some commenters dismissed the project as lobbying or regulatory advocacy, and others worried it could be parodied, while one noted that red lines are easy to state in the abstract but hard to judge until actually confronted.
  • One commenter suggested a stronger example than the one used on the site, pointing to Stanislav Petrov as someone who acted effectively on a personal line.
Before Pixels: Speed-Minderunsung.aresluna.org · 22 points · 1 comment · 5dhistorytoolsIllinois State Police reportedly fielded a 1950s or early 1960s "Speed-Minder" police car that displayed its own speed on a non-digital dial, with no radar involved.

Key points

  • The car did not use radar; it simply reported its own measured speed to the driver and onlookers.
  • The author could not find evidence of how effective the device was at slowing drivers.
  • It may have been intended as a calibration aid for checking a car's own speedometer rather than as enforcement.
Microsoft Status: "We are re-routing the protocols"admin.cloud.microsoft · 21 points · 13 comments · 4dculturesystemsMicrosoft 365 incident MO1550821 is degrading access to Exchange Online, Teams, SharePoint, Copilot, Defender XDR and Purview, with users unable to connect to multiple services.

Key points

  • Microsoft's status update says it validated an alternative health check method on a subset of affected infrastructure, which let those servers return to service.
  • The fix is being applied more broadly, and an automated process to detect and restart unresponsive servers is being prepared.

What HN thinks

  • Commenters questioned why the public status page requires login, with one pointing out that a static status page should not be affected by load on core services, while another noted a public version exists at status.cloud.microsoft.
  • One commenter speculated the cause was increased load that might look like a DDoS, with the login redirect filtering traffic at the network edge.
  • Users reported being unable to sign in at all, including emailed verification codes not arriving and 2FA failing, and one saw an error page in the wrong language.

The article couldn't be read, so this summary is based on the discussion.

Vibes Don't Maintainroll.bybacon.com · 21 points · 7 comments · 4daiprogrammingA restaurant owner's Claude Code web app worked on the happy path but lacked responsive layout, login, translations, and error handling, so the author argues its maintenance depends on recorded intent.

Key points

  • The app had no responsive layout, login, settings, translations, error states, or handling for empty, duplicate, or concurrent edits.
  • The author lists failure cases like file corruption, dependency breakage, public exposure, and daylight saving time shifting every schedule by an hour.
  • The original reasoning lived in a chat session that was lost, leaving no spec, tests, or commit messages explaining intent.
  • The author's team keeps markdown specs, stories, and decisions in git alongside the code so humans or agents can read them later.

What HN thinks

  • Commenters recommended having an LLM generate a markdown context file or writing a detailed markdown request before prompting AI.
  • One commenter who built a diagnostic app with AI reported that asking for changes often broke unrelated functionality.
  • Another commenter argued that AI will write tests and maintain code, and that human architects who understand how parts fit together are the remaining bottleneck.
Paths in Piecesstoppels.ch · 20 points · 0 comments · 4dprogrammingsystemsGCC splits short string constants copied into buffers into overlapping 16-to-64-byte vector pieces, so grep and strings can miss embedded absolute paths that a binary still uses at runtime.

Key points

  • A 53-byte literal path is stored as two overlapping 32-byte pieces, so the full path never appears contiguously in the binary.
  • Conda-style _placehold padding gets shrunk to 176 bytes because GCC exploits the cyclic repetition of the padding.
  • Clang keeps the literal intact in .rodata.str1.1 and generates nearly identical copy instructions, so the splitting gains little.
  • Disabling the splitting is hard: -Os works, and -mstore-max=128 only restores older defaults for strings up to 256 bytes.
Creating Worlds: Runes and Spells Prototypefractalkitty.com · 20 points · 4 comments · 7dmathtoolsA mathematics educator's post walks readers through a playful magic system built from grid-drawn runes and spells that transform them, with a prototype tool for experimenting.

Key points

  • Runes can sit on a 4x4 square grid, a hexagonal grid, or a polar grid, and spells such as reflections, flattening, and complements transform them.
  • A spell book is defined as closed when its spells can only generate runes already inside it, so no surprise runes appear.
  • Example spell book structures include an involution pair toggling two runes, an orbit, a sink where every rune collapses to one, and a cycle-and-tail pattern.
  • The author notes the system maps onto sets, groups, and category theory, and musical notes can serve as runes.

What HN thinks

  • One commenter points to a Minecraft mod that implements a similar glyph-drawing system as stack-based programming.
  • Another commenter wonders whether combining primitives like this could work for procedural game design, but worries the results might feel samey.
Grote Reber and the 'Square Kilometre Array' Near Bothwell, Tasmania (2017)researchgate.net · 20 points · 2 comments · 8dspacehistoryA 2017 paper by historian Ron Ekers and colleagues on Grote Reber's proposed low-frequency radio array near Bothwell, Tasmania, in the 1960s and 1970s, which was never built as envisioned.
Jensen Huang resigning effective immediatelystocktitan.net · 14 points · 12 comments · 4dbusinessaiNvidia says CEO Jensen Huang resigned effective immediately, with CFO Colette Kress named interim CEO while keeping her finance role and a permanent search underway.

Key points

  • Harvey C. Jones was named the board's first-ever chair, and the board says it will consider internal and external candidates for CEO.
  • Nvidia still plans to report Q3 fiscal 2027 results on November 17 at 1:20 p.m. PT, followed by an earnings call at 2 p.m. PT.
  • The filing gives no reason for the resignation, and the press release says the board expects operations and the product roadmap to continue without interruption.

What HN thinks

  • Commenters linked the resignation to a relationship scandal, but the official announcement itself gives no reason and no allegation is confirmed in the provided material.
  • Several commenters contrasted the response with Elon Musk's, and one argued the reaction reflects whether supporters prioritize ethics or leadership results, which turned into a heated and partly racially charged exchange.
The Lost World of Aquaterratheatlantic.com · 14 points · 2 comments · 5dhistoryscienceThe Atlantic profiles Aquaterra, a submerged or lost landmass offered as a site of ancient human migration, but the article could not be fetched and no discussion exists yet.
I caught Microsoft Defender uploading my unreleased software without askinggarrettstimpson.ca · 13 points · 4 comments · 5dsecuritytoolsA developer says Microsoft Defender uploaded 217 of his unreleased programs to Microsoft over three months, mostly test builds, without a consent prompt he could find.

Key points

  • Defender's Operational log Event 2050 recorded 217 sample submissions covering 216 different SHA-256 hashes, 201 of them builds of one Go test program.
  • The MpCmdRun.exe process sent 224 MB to Azure blob storage in one day, in bursts of about 28 MB per minute.
  • A Event 5007 log shows sample submission switched from never (2) back to safe samples automatically (1) about two hours after he disabled it.
  • Defender's support log held 85 'command line reported as lowfi' entries, 78 of which contained his user and project folder names, though it's unclear whether they were hashed before sending.

What HN thinks

  • Commenters were skeptical, with several saying the post reads like an advertisement for the author's Hodometer tool.
  • One commenter pointed to other HN threads about Defender sample uploads as possible context for why the story took several submissions to gain traction.
Making the Iron Giant Liveanimationobsessive.substack.com · 13 points · 1 comment · 8dgamingdesignThe Iron Giant's original 3D models of the title character surfaced on the Internet Archive this summer, and fans have started animating them again.

Key points

  • Bird wanted the Giant to be a 3D character living seamlessly in a 2D world, with CGI as emotive as the hand-drawn animation.
  • Crewmember Brian Gardner's Gartoons system added wobbly, variable-thickness lines and frame-rate modulation so the Giant could move on twos like the 2D art.
  • Gardner also made 3D frames printable onto paper with peg holes so 2D animators could guide their drawings to interact with the Giant.
  • Animation ran on Silicon Graphics machines using a beta of Maya, which was not yet on retail sale until 1998.
Decoding REXPaint Files in Odintylerelton.com · 12 points · 1 comment · 7dprogramminggamingA step-by-step Odin tutorial that decodes REXPaint .xp ASCII art files by gunzipping them and parsing their little-endian layers and 10-byte cells, with bounds checks against corrupt input.

Key points

  • Files are gzipped; after decompression the header is a negative version number, a layer count, then per-layer width and height followed by cells.
  • Each cell is 10 bytes: an i32 CP437 character index, then foreground and background RGB triples.
  • Cells are stored column-major, so the cell at (x, y) sits at index x * height + y.
  • The parser checks counts against remaining bytes before allocating, guarding against truncated or corrupted files claiming billions of layers.

What HN thinks

  • Only one comment was posted, a brief thanks, so the thread adds little beyond the article.
AI Models Know What Women, Peace and Security Is. They Just Don't Bring It Upblog.cip.org · 11 points · 4 comments · 4daipolicyA benchmark of 30 scenarios finds frontier LLMs almost never raise gender considerations in sparse military-style prompts, with the frontier average falling from 0.82 to 0.16 when cues are removed.

Key points

  • On a terse intelligence summary with no gender cues, Gemini and GPT models produced staff-officer analysis that never mentioned women, girls, or sexual violence.
  • In a budget-cut scenario, Gemini 3.1 Pro recommended eliminating the gender advisor post and drafted a justification, while Claude Opus 5 pushed back citing UNSCR 1325 and funding clawback risks.
  • A multi-agent workflow with a mandatory verification pass scored 0.84 on GPT-5.6 Sol and kept gender sensitivity flat between cued and sparse prompts, while the bare model dropped from 0.86 to 0.21.
  • The adversarial tier has only three scenarios, so the authors treat resistance-to-pressure results as indicative rather than definitive.

What HN thinks

  • A commenter who worked on the benchmark says the same omission pattern would likely hold for other policy-required considerations such as civilian harm and legal review.
  • One commenter questions why WPS should apply, arguing it should only be enacted when women are drafted and die in war, as men are.
Europe has been living beyond its means. Now the bills are dueeconomist.com · 10 points · 2 comments · 5dpolicybusinessAn Economist essay argues Europe has spent beyond its means and that the resulting fiscal and economic costs are now coming due, though the article itself could not be retrieved.

What HN thinks

  • Commenters could not read the piece because it is paywalled, so their reactions rest on the headline alone.
  • One commenter attributes the problem to wars and energy shocks, including Russia's war in Ukraine, and to low taxes paid by US megacorps.
  • Another commenter reads the headline as an argument against spending on wars of any kind, weapons or economic.

The article couldn't be read, so this summary is based on the discussion.

Apify Security Incidentstatus.apify.com · 8 points · 0 comments · 4dNot summarized yet.
Gemini 4 Hydrogendeepmind.google · 8 points · 1 comment · 4dNot summarized yet.
Show HN: Apron Chat Protocolapron.chat · 8 points · 0 comments · 4dNot summarized yet.