Today’s thread is trust in technical systems: Midjourney’s new full-body ultrasound scanner leads with big claims and bigger questions about proof, safety, and the path from spa scans to medicine, while security and product reliability show up in the report on 10,000 GitHub repositories spreading Trojan malware and in complaints that Microsoft’s new Outlook turns instant tasks into multi-second waits. AI news follows that same line, with Noam Shazeer joining OpenAI, arguments for local Qwen as a distinct coding tool, and DeepSeek adding vision. Elsewhere: Switzerland reopens the nuclear debate, AMD quietly drops a Ryzen feature, and Ubiquiti ships a ZFS NAS.
Midjourney announced a new medical project: a full-body ultrasound tomography scanner aimed first at spa-style body scans, with a long-term pitch of frequent cheap imaging and eventual medical use. The reaction was split between curiosity about the hardware and deep skepticism that the images, safety claims, regulatory path, and screening vision are anywhere near proven.
Switzerland’s parliament voted to remove its long-standing ban on building new nuclear plants, but any real change still faces a public referendum and years of financing and siting fights. The comments split between seeing this as overdue energy realism for a winter-short power system and seeing it as symbolic politics for a technology that is now too slow and expensive to matter.
A security researcher documented roughly 10,000 GitHub repositories that impersonate legitimate projects and point users to Trojan-filled ZIP files, then tied the behavior to a broader campaign of search manipulation and fake freshness. The useful signal is that this is not a one-off repo scam. It is a scalable distribution pattern that keeps resurfacing across GitHub, search engines, and adjacent software directories.
A widely upvoted post benchmarks Microsoft’s new Outlook against classic Outlook on Windows and finds basic actions taking seconds instead of feeling instant. The comments mostly agree the app is badly engineered, but the strongest signal is that people blame product and platform decisions around web wrappers, enterprise lock-in, and security tooling more than the browser tech itself.
Noam Shazeer, a key Google researcher behind the transformer architecture and a recent Gemini co-lead, posted that he is joining OpenAI. The reaction treated it as a real talent loss for Google and a signal about where top AI researchers think they can move fastest.
A Tom’s Hardware report says newer AMD AGESA firmware disables transparent DRAM encryption on some consumer Ryzen systems, leaving the feature only on Ryzen Pro parts even though many boards had exposed it before. Commenters mostly saw this less as a must-have consumer security tool and more as an ugly case of post-sale feature removal with no explanation from AMD.
A Realtor.com piece argues the old rule of spending no more than 30% of income on rent no longer matches reality because many workers still feel squeezed even when they stay under that line. Most of the conversation rejected the framing and said the core problem is not budgeting advice or wages alone, but housing costs that have outrun incomes because supply is constrained and housing has become a protected asset.
A blog post argued that running Qwen locally should not be judged as a cheaper, weaker version of Anthropic Opus, but as a different kind of coding tool with better privacy, control, and some niche strengths. The comments mostly agreed on the framing, then got into what actually makes local models useful in practice: unstable behavior, harness design, serving choices like llama.cpp vs vLLM, and when local beats cloud despite being clearly worse at long, hard tasks.
Ubiquiti introduced a $3,999 16-bay rackmount NAS that uses OpenZFS, dual 25 gigabit networking, ECC memory, and NVMe cache, aimed at small and mid-sized business storage rather than home users. The comments mostly treated it as a credible new option for teams already in the UniFi ecosystem, while arguing over whether Ubiquiti’s software quality and product follow-through are strong enough to trust with critical data.
DeepSeek’s chat app appears to have added real image understanding, not just OCR, letting users upload pictures and ask what is in them. The comments focused less on the launch itself than on price, missing API support, and how vision is becoming table stakes for coding agents and mobile voice workflows.
A Works in Progress article argues Madrid expanded its metro unusually cheaply by keeping technical expertise inside government, standardizing designs, and pushing projects through with political focus. The comments mostly agree that state capacity matters, but they also point out omitted caveats like easy geology, specific failed lines, and the fact that Madrid may be an exception rather than a general Spanish model.
A King’s College London writeup says hospitals and universities are finding new uses for existing drugs after patents expire, often at a fraction of the original price, because there is little commercial incentive for pharma to run those studies. The comments mostly agreed on the broken incentives, but added that off-label use is already common and that safety, reimbursement, and regulation still decide whether cheap repurposing actually reaches patients.
A blog post argues that Windows 2000’s interface got core desktop UI basics right: obvious buttons, consistent controls, clear labels, and fast response on weak hardware. The comments mostly agree, then push the point further by blaming later web, mobile, and cross-platform design trends for making software less discoverable and less reversible.
A popular Emacs user wrote up the Emacs 31 features they are already using daily, including built-in Tree-sitter grammar installation, editable cross-reference results, better window layout controls, and terminal improvements. The comments turned into a broader check-in on whether Emacs is still gaining ground as a serious modern editor, especially now that AI tools make its configuration and extension model much easier to use.
A blog post about adding a JavaScript fallback to a WebAssembly proof-of-work system turned into a war story about reproducible builds, after the author found Clang was generating different binaries depending on memory layout. Readers mostly treated it as a compiler determinism bug plus a broader reminder that low-level toolchains leak hidden environment state in surprising ways.
A MAME developer wrote up how he used an AI coding agent to revive dynamic recompilation for PowerPC emulation and get Apple’s old Graphing Calculator running again. The post landed as a concrete example of AI speeding up gnarly legacy systems work, with readers split between excitement about the result and unease about what it says about programming as a craft.
Valve released SteamOS 3.8 as a stable update for Steam Deck and newer AMD handhelds, and the release notes also add early support for third-party devices. Readers cared less about the version bump itself than what it signals: SteamOS is edging beyond Valve hardware, but it is still rough as a general-purpose desktop OS.
A Norwegian privacy advocate documented how electronics retailer Elkjøp required customers to accept marketing in order to stay in its loyalty club, then learned five years later that Norway’s data regulator fined the company €1.8 million. Readers focused less on the individual win than on what it says about GDPR enforcement: real penalties exist, but they arrive slowly and only if someone bothers to push the complaint through.
A Mindgard blog post showed ChatGPT’s image tool producing violent and sexual images from a fake “restore this attached photo” prompt, sometimes even when no image was attached. Commenters mostly agreed the writeup was sensationalized, but many still treated the behavior as a real safety and product design failure rather than a fake issue.
Australia’s communications regulator will require SMS and MMS sender IDs to be registered starting July 1, with unregistered branded senders labeled “Unverified” instead of appearing as trusted business names. The post drew strong support from people drowning in scam texts, but the useful detail was that this mainly targets phishing and impersonation, not the full spam problem.
IEEE Spectrum profiled Modos, a two-person startup building a 13.3-inch color e-paper monitor that aims to make e-ink responsive enough for real computer use, with touch and much faster refresh than typical readers. The interest was less about the article itself than about whether this points to a real new category of low-power, sunlight-readable displays or just an expensive niche gadget.
A blog post argues that W Social, a new Europe-branded social network now attracting public institutions and EU politicians, is a poor fit for “digital sovereignty” because it is closed-source, for-profit, and built around invasive identity checks. Comments mostly agreed that Europe already has more open options like Mastodon and Eurosky, and saw W Social’s early political adoption as a marketing and access story, not a technical one.
A blog post argues that RTK, a tool that rewrites command output to save LLM tokens, is being sold on eye-catching “token savings” while saying too little about accuracy, real bill impact, and maintenance risk. Commenters mostly agreed that token reduction needs better measurement, but several users said the concept is still useful when applied narrowly or benchmarked inside a specific workflow.
Cornell’s CS 6120 self-guided compilers course puts a graduate-level optimizing-compiler class online, with lectures, papers, and projects. Readers mostly treated it as a strong free resource, then dug into what “advanced” really means and whether its dynamic compilation material reflects what actually works in production.
A blog post walks through Git’s lesser-known ways to ignore files beyond a repo’s `.gitignore`, including per-user global ignores, per-repo local excludes, and `git check-ignore` for debugging. The comments largely agree those features are useful, but split hard on when ignore rules belong in the repo versus in each developer’s own config.
A January report says Meta struck a deal tied to eight planned TerraPower Natrium nuclear reactors, a sodium-cooled design that would power AI data centers if it ever ships. The comments mostly treated it as financing theater around an unproven reactor and a very long construction timeline, not near-term power capacity.
Quanta covers a new math result on riffle shuffling: if real people do imperfect, uneven cuts before the usual random interleaving, about 14 shuffles are enough to randomize a 52-card deck. The comments mostly corrected the article’s framing, especially what “perfect shuffle” means and what the proof actually changes.
An Ask HN post asked whether anyone is actually using Google’s A2A agent-to-agent protocol. The answers were mostly “not really” outside enterprise pilots, with practitioners saying MCP is winning real adoption today and A2A still feels overbuilt or aimed at a narrower future use case.
TesterArmy launched on Hacker News as an AI-driven service that tests web and mobile apps by navigating them like a human tester instead of relying on brittle scripted end-to-end tests. The comments focused on whether this is actually better than letting coding models generate Playwright or Cypress tests in-house, and on whether the economics and reliability hold up at scale.