Today’s thread is digital trust and control: the Xbox outage turned physical game discs into another reminder that software ownership is contingent, while an active npm compromise and a critique of misleading web security design showed how brittle everyday developer and user trust has become. AI trust runs through the rest, from skepticism toward AI-generated blog art to Apple and OpenAI escalating their dispute over alleged confidential data and trade secrets. Elsewhere, Waymo’s Dallas expansion marked robotaxis becoming ordinary infrastructure, the Iran war exposed strain on U.S. precision-missile stockpiles, and local AI got a boost from DeepSeek on one MI300X and sparse Qwen on consumer devices.
-
Xbox goes down. You can't play games you own on disc
-
birchtree.me
- 617 comments
A blog post used the latest Xbox outage to argue that physical game discs no longer guarantee access, because many Xbox games still need online license checks or downloads before they will run. Commenters largely agreed and broadened it into a bigger point about software ownership, DRM, and why console gaming is starting to feel worse than PC, emulation, or piracy.
-
-
U.S. used 'virtually all' of its long-range precision missiles during Iran war
-
cnbc.com
- 602 comments
A Reuters report said the U.S. has burned through nearly all of its Army stock of ATACMS and newer PrSM precision missiles during the Iran war, raising readiness questions. Commenters mostly said the headline overstates the specific stockpile issue, but agreed the bigger signal is ugly: the U.S. is spending scarce stand-off munitions in a war it cannot cheaply or politically escalate.
-
-
Waymo in Dallas
-
waymo.com
- 295 comments
Waymo has opened its driverless taxi service to anyone in its Dallas operating zone, adding another large, car-centric US city to its rollout. The comments were mostly less about Dallas itself and more about a shift in attitude: for many riders and pedestrians, robotaxis have moved from novelty to clearly better-than-human urban driving, while cost, liability, surveillance, and transit tradeoffs still look unresolved.
-
-
Keyv and friends compromised in active Shai-Hulud supply chain attack
-
aikido.dev
- 119 comments
Aikido reported an active npm supply-chain attack that compromised Keyv-related packages and then spread worm-style into more maintainers’ packages via stolen credentials. The comments mostly treated it as another proof that default JavaScript workflows are too trusting, with the most useful advice centering on slowing dependency adoption and separating code execution from publish credentials.
-
-
AI-Generated Images Discourage Me from Reading Your Blog
-
nelson.cloud
- 432 comments
A short blog post argued that obvious AI-generated images make the author less likely to read a blog, because they signal low effort and raise suspicion that the text is AI-generated too. The comments turned that into a wider argument about trust, taste, and whether AI visuals are useful illustration or just a spam marker.
-
-
Apple says more ex-employees may have taken confidential data to OpenAI
-
techcrunch.com
- 249 comments
A new Apple court filing says more former employees who joined OpenAI may have kept Apple devices, accessed internal systems after leaving, and taken confidential hardware documents. The comments mostly treated this less as normal employee mobility and more as a trade secret case, while also noting Apple’s own offboarding and device-management practices look surprisingly weak.
-
-
Apple is getting this wrong
-
openai.com
- 281 comments
OpenAI published a public blog post attacking Apple’s trade-secrets lawsuit over former Apple employees now at OpenAI, using emails and message transcripts to argue Apple mishandled access and overreached. Most readers zeroed in less on the evidence than on the fact that OpenAI chose to litigate on its corporate blog in a tone many found reckless and juvenile.
-
-
DeepSeek V4 Flash on a Single AMD MI300X
-
github.com
- 87 comments
A GitHub repo shows how to run DeepSeek V4 Flash, a large open-weights language model, on a single AMD MI300X GPU without quantizing the model weights. The appeal is not lower token cost than DeepSeek’s own API, but getting full-quality local inference with decent speed, strong privacy, and hardware that a small team can realistically rent.
-
-
Show HN: Run an 80B Qwen in 4.3 GB of RAM on a Mac, and a 35B on an iPhone
-
- 134 comments
A new open source project called Swiftlet shows how to stream sparse large language model weights from SSD so an 80B-parameter Qwen model can run on a Mac using about 4.3 GB of RAM, and a 35B model can run on an iPhone. The pitch is not speed so much as pushing very large local models onto consumer devices, which kicked off debate about whether this is a real path to on-device AI or just a slow demo with harsh bandwidth limits.
-
-
Web Security is Too Hard
-
textslashplain.com
- 105 comments
A blog post walked through why Cloudflare’s new wallet site looked exactly like a phishing scam even though it was real, using a separate `cloudflare.pay` domain, weak trust signals, and confusing login flow to show how easy it is to train users into bad security habits. The comments largely agreed this is a company design and org problem, not a purely technical one, and used it as a broader indictment of how marketing and speed routinely override web trust.
-
-
Online ad giant Adform was hacked, proving once again why ad blockers are needed
-
this.weekinsecurity.com
- 78 comments
A security incident at ad platform Adform let attackers use ad-delivered JavaScript to tamper with users’ clipboards and swap cryptocurrency wallet addresses. The comments largely treated it as another reminder that third-party ad tech is a live security boundary, not just an annoyance, and argued ad blockers are now basic protection.
-
-
Harness engineering for self-improvement
-
lilianweng.github.io
- 68 comments
A new post argues that a coding agent’s "harness" is now a bigger lever than model weights for improving real-world performance, and sketches how agents can iteratively improve their own prompts, tools, and workflows. The comments mostly treated this as practical field notes, with a lot of concrete advice on evals, context bloat, and keeping the optimization target small and measurable.
-
-
Oxide Computer raises $445M (SEC Form D)
-
sec.gov
- 69 comments
An SEC filing shows server-and-software startup Oxide has raised another $445 million, on top of rounds it announced in 2023, 2025, and earlier this year. Comments focused less on the filing itself than on what this says about demand for an integrated on-prem cloud alternative, and whether Oxide is building a durable hardware business or an acquisition target.
-
-
FFmpeg 9.0
-
- 94 comments
FFmpeg 9.0 is out with a grab bag of new media plumbing, including animated WebP support, more Vulkan and hardware-accelerated paths, ONNX Runtime integration, and niche format work like Playdate video. The reaction was less about any one headline feature and more about how central FFmpeg has become, plus a side debate over security burden and light AI use in the release process.
-
-
All of Winona Police Department's Flock cameras cut down and stolen
-
valleynewslive.com
- 206 comments
A local news report says every Flock license plate reader used by Winona, Minnesota police was cut off its poles and stolen in a coordinated hit. The comments mostly turned the incident into a broader indictment of Flock as a nationwide surveillance system with weak privacy guarantees, loose data sharing, and a long abuse history.
-
-
Amazonian civilization had estimated 3M people in 3% of forest area
-
science.org
- 196 comments
A Science report says lidar scans are uncovering hundreds of large earthen structures in the southwestern Amazon, supporting estimates that millions once lived there and managed parts of the rainforest far more intensively than the old “untouched wilderness” story suggests. The comments mostly ran with that broader reframing, while pushing back on easy romanticism and on the paper’s bolder population extrapolations.
-
-
Twenty Years of Pandoc
-
pandoc.org
- 49 comments
A 20th-anniversary post about Pandoc, the long-running open source document converter, sparked a wave of praise from users who rely on it for everything from academic writing to contract redlines and static site generation. The useful signal was not just nostalgia: people surfaced why Pandoc’s design has held up, where it fits better than LLM-style tooling, and what its Haskell roots have meant for contributors.
-
-
That time when I failed the Microsoft interview
-
ochagavia.nl
- 296 comments
A blog post recounts a 2015 Microsoft internship interview where the author told the interviewer he already knew a classic marble riddle, asked for another question, and was rejected. The comments turned it into a broader indictment of puzzle-heavy tech interviews, with some useful nuance about what these questions are really testing and why the process often feels arbitrary.
-
-
"Clean" Code, Horrible Performance (2023)
-
computerenhance.com
- 116 comments
A 2023 essay argues that several “Clean Code” habits, especially object-heavy indirection and tiny decomposed methods, can impose real runtime costs compared with simpler procedural designs. The comments mostly agreed that the benchmark proves a tradeoff, but pushed the discussion toward when extensibility and maintainability are worth paying for and when they are just dogma.
-
-
Why Large Language Models Fail at Tabular Prediction
-
arxiv.org
- 32 comments
A paper tests frontier large language models on plain tabular prediction tasks and finds they lose badly to old-school methods like boosted trees, even after checking common suspects like CSV formatting. The comments mostly treat that result as unsurprising in practice, but useful as a clean demonstration that language models break down on structured numeric data for reasons deeper than prompt formatting.
-
-
Show HN: Fine-tune an 8B model on a 4 GB laptop GPU
-
- 23 comments
Soup is an open source CLI for fine-tuning small language models with LoRA on very low-end GPUs by streaming most of the frozen base model from system RAM instead of keeping it all in VRAM. The key signal in the comments is that the author backed the headline with detailed memory and correctness numbers, while also clarifying the limits: this does not eliminate VRAM requirements, and it is better suited to adapting behavior than teaching new facts.
-
-
Thanks FedEx, This Is Why We Keep Getting Phished (2024)
-
troyhunt.com
- 54 comments
A blog post shows FedEx Australia sending customers to a third-party payment page on an unrelated domain that looks exactly like phishing. The comments land on a blunt point: companies keep teaching users the wrong instincts by sending real messages that are indistinguishable from scams.
-
-
You don't need React: creating a minimal UI library in Vanilla JavaScript
-
pedroth.github.io
- 105 comments
A blog post shows how to build a tiny React-like UI layer in plain JavaScript and argues many sites do not need a full frontend framework. The comments mostly agreed React is overused for simple sites, but they also pointed out that the demo sidesteps the hard parts frameworks actually solve, especially state, DOM updates, and preserving input behavior.
-
-
Show HN: Simple algorithm and color space to generate diverse skin tones
-
toneyalexander.github.io
- 87 comments
A developer posted an interactive color space and generation algorithm meant to make it easier to pick realistic, varied human skin tones for games, digital art, and character creators. Readers liked the usability and math, but the useful signal was where the model breaks down: lighting, rendering, and real-world reference standards still matter as much as the picker itself.
-
-
Mistral's Shieldstral: 3B open-weights model for multimodal moderation
-
mistral.ai
- 69 comments
Mistral released Shieldstral, a 3B open-weights model for text-and-image content moderation aimed at cheap, customizable policy enforcement. Comments mostly treated it as a practical B2B tool for triage and compliance, while arguing over how flexible, explainable, and politically loaded AI moderation can really be.
-
-
There Will Come Soft Rains (1950) [pdf]
-
users.wpi.edu
- 376 comments
A Hacker News post resurfaced Ray Bradbury’s 1950 short story “There Will Come Soft Rains” because its in-story date is August 4, 2026. Readers mostly treated it as an eerie check-in on two things that now feel newly close again: smart-home automation and the return of nuclear-war anxiety.
-
-
Dates That Don't Exist (2015)
-
blog.yossarian.net
- 76 comments
A 2015 blog post walks through the Gregorian calendar’s skipped days in 1582 and how different programming languages handle dates that never existed. The comments mostly argue the post overstates a universal rule, since most date libraries intentionally use a simplified "proleptic Gregorian" calendar and real historical cutovers varied by country, church, and use case.
-
-
Nobel Disease
-
en.wikipedia.org
- 102 comments
A Wikipedia article on “Nobel disease” sparked a debate over whether some Nobel winners really become cranks after fame, or whether the article just documents a broader habit of treating prestige as expertise outside its lane. Most of the useful discussion landed on selection bias, celebrity amplification, and the sloppy mixing of outright racism with merely eccentric beliefs.
-
-
In Memory of My Wife, Elise Cawley, with Thanks for 36 Wonderful Years
-
writings.stephenwolfram.com
- 44 comments
Stephen Wolfram published a long memorial essay about his wife Elise Cawley, who died suddenly in 2026, recounting their 36-year relationship, family life, and her work. The comments were mostly raw condolences, with a few useful side threads on Wolfram’s obsessive life-logging and on why the post’s unusual length felt earned rather than self-indulgent.
-