HN Brief: 2026-05-16
Today’s front page was dominated by a wave of AI skepticism, from a viral thread arguing that companies are shipping buggy code under “AI psychosis” to the fallout of AI-generated slop reports killing a bug bounty program. A second throughline was the creeping expansion of surveillance—the DOJ demanding app-store data on 100,000 car-tinkerers, and London police deploying facial recognition at a protest for the first time. And amid the gloom, a few bright spots: Project Gutenberg getting a long-overdue redesign, and a decade-long effort to rescue 13,000 technical manuals finally reaching the finish line.
Threads most worth clicking into: “I believe there are entire companies right now under AI psychosis” — the thread that split HN on whether AI agents are genuinely productive or just creating illusions of progress; “U.S. DOJ demands Apple and Google unmask over 100k users of car-tinkering app” — a breathtakingly broad subpoena that swept up law-abiding tinkerers and sparked a privacy vs. enforcement debate; “Waymo updates 3,800 robotaxis after they 'drive into standing water'” — a stark reminder that mapping-dependent autonomy has a fundamental vulnerability no amount of pre-driving can fix; “We are retiring our bug bounty program” — a textbook demonstration of how AI slop submissions break the incentive structure of open-source maintenance; and “ABC News has taken all FiveThirtyEight articles offline” — a genuinely useful public resource nuked out of corporate spite, with no good reason.
I believe there are entire companies right now under AI psychosis [article]
1255 points · 598 comments · twitter.com · 11h ago
The linked article wasn't available to this summarizer; from the discussion, Mitchell Hashimoto’s tweet argues that entire companies have fallen into “AI psychosis,” shipping buggy code with the rationalization that AI agents will patch things faster than humans ever could. The thread split hard: one camp insisted that agents are already better than the average developer and that resisting this is like carriage operators fighting automobiles, while the other side pushed back with concrete examples of agents happily declaring work done while ignoring tests entirely, producing nothing but the illusion of productivity. A strong rebuttal pointed out that MTTR-optimized YOLO deployment is fine for recoverable errors but catastrophic when a bug silently corrupts data for months, and another comment nailed the core metric: speed means nothing if the defect-introduction rate exceeds the fix rate. Tangents spun off into VC pressure forcing every startup to claim AI, the inevitability of a crash like 2000/2008, and a few people arguing that formal verification and exquisite architectures are the only sane path forward—though most agreed that humanity has never done that for human output either.
Project Gutenberg – keeps getting better [article]
898 points · 189 comments · www.gutenberg.org · 15h ago
Project Gutenberg’s site has been getting a serious redesign after years of looking like a relic, and the thread is full of people noticing and thanking the volunteer developers for the work. A lot of the discussion immediately pivots to two pain points: the site is getting hammered by bot traffic (AI scrapers hitting it in DDoS-like patterns from thousands of IPs, which the team hasn’t fully solved), and the lack of a clean, officially supported machine-readable index for bulk downloading—though the volunteers point to the offline catalogs and OPDS feeds and say a new OPDS 2.0 is coming. Several commenters pushed for features like sorting books by original publication date (not just Gutenberg release date), git-backed version history for proofreading fixes, and better mobile menus (which the devs acknowledged and said they’re iterating on). There’s also a brief history refresher: the project started in 1971 when Michael Hart got a massive account on a mainframe that was one of the 15 nodes on the ARPANET.
Explore Wikipedia Like a Windows XP Desktop [article]
506 points · 113 comments · explorer.samismith.com · 23h ago
The site turns Wikipedia into a navigable file explorer styled after Windows XP, with articles as documents, media as viewable files, and categories as folders you click through. The thread quickly split between people charmed by the nostalgia and those grumbling about inaccuracies in the XP design—the title bars and orange buttons get the era right, but the flat desktop and odd icons lean closer to Temu’s version, likely to dodge copyright. Others dove into the underlying mechanics, pointing out that the folder hierarchy is lifted straight from Wikipedia’s category system, which is really a tagging system with messy overlaps rather than a clean tree, sparking a broader debate about whether nested folders are a good mental model for knowledge or a relic that search rightly killed. A few people went down rabbit holes chasing weird Wikipedia redirects (like “More milk” leading to Michael Jackson’s death, because propofol is nicknamed “milk of amnesia”), and the lack of working search drew mixed reactions—some called it useless, others called it authentic to the original Windows Explorer experience.
California bill would require patches or refunds when online games shut down [article]
461 points · 289 comments · arstechnica.com · 12h ago
A California bill that would force game publishers to offer refunds or offline patches when they kill an online game has cleared a key committee vote, with the Stop Killing Games group celebrating a win after the shutdown of Ubisoft’s The Crew. The HN thread quickly split: some argued the bill is a clumsy bandaid and that legalizing reverse engineering would let communities keep games alive themselves, though others pointed out that DMCA restrictions and the opacity of server-side logic make that impractical for modern online titles. A large chunk of the discussion worried about enforcement—publishers could spin up shell LLCs per title or simply go bankrupt, and several people suggested making platform holders like Steam jointly liable or requiring source code escrow upfront. Another recurring concern was that small studios would be crushed by the cost of disentangling third-party cloud services and proprietary middleware to release a standalone server, handing the market to big players, while a few commenters broadened the complaint to all "lifetime license" software, not just games.
U.S. DOJ demands Apple and Google unmask over 100k users of car-tinkering app [article]
416 points · 282 comments · macdailynews.com · 14h ago
The DOJ is demanding Apple, Google, Amazon, and Walmart hand over personal data on anyone who downloaded EZ Lynk’s Auto Agent app or bought its hardware, aiming to find witnesses in a years-long emissions defeat-device case. The HN thread immediately zeroes in on the privacy overreach—commenters argue the subpoena is absurdly broad, sweeping up users who just wanted to read their own trouble codes, and point out the government could just target obvious violators like the “rolling coal” crowd instead of dragging 100,000 innocent people into it. That rolling coal tangent takes off, with people sharing horror stories of being blinded by intentionally dumped exhaust and noting that the real problem is enforcement against the worst offenders, not mass surveillance. A separate meta-debate erupts in the comments over an AI-written article linked as context, with several people pushing back hard that AI slop undermines trust and shouldn’t be posted on HN, while the poster defends it as a curated compilation—illustrating how even a privacy-related thread can’t escape the site’s ongoing fight about AI-generated content. The core split is clear: the government says it needs the data to enforce emissions laws, but most of the thread sees this as a fishing expedition that will sweep up law-abiding tinkerers and set a dangerous precedent for app-store surveillance.
Bun Rust rewrite: "codebase fails basic miri checks, allows for UB in safe rust" [article]
385 points · 271 comments · github.com · 15h ago
The submission is a GitHub issue flagging that Bun’s AI-assisted Rust rewrite introduces undefined behavior in safe Rust code, failing the Miri checker. The HN thread quickly turned into a meta‑debate about the rewrite’s announcement and reception: many argued the Bun team never claimed the port was memory‑safe yet, and that the big flashy headlines were coming from outsiders, not the developers themselves. Commenters split between those calling the UB a predictable, fixable side effect of a mechanical 1:1 translation and those insisting that merging code with known UB into `main` undermines Rust’s core safety guarantees. Others dove into the asymmetry of attention—pointing out that a correction (this issue) gets far less reach than the original “Bun rewritten in Rust” buzz, a pattern they likened to marketing tactics and media sensationalism rather than a genuine failure of the project’s incremental approach.
A 0-click exploit chain for the Pixel 10 [article]
366 points · 181 comments · projectzero.google · 18h ago
Google Project Zero published a full zero-click exploit chain for the Pixel 10, building on their Pixel 9 work by porting the Dolby audio bug and discovering a new, trivially exploitable kernel vulnerability in the VPU driver that lets userspace map arbitrary physical memory—including the kernel image—with just five lines of code. The HN crowd latched onto how shallow the bug was: a missing bounds check in an mmap handler that was instantly noticeable during a cursory audit, which sparked a long thread about whether AI code reviewers could have caught it (several people pasted the vulnerable function into various LLMs and reported the models correctly identified the issue without web access, though others pointed out the base-rate fallacy problem). That led into a heated split over liability—some argued the developer and Google should face severe consequences like prison or massive fines, while others countered that such rules would kill software development entirely or just drive it underground, with comparisons to other engineering fields where liability exists but is managed through insurance and professional standards. A separate thread noted the article’s admission that this was the first Android driver bug from this researcher patched within 90 days, which prompted comparisons to Apple’s slower response times and a few people sharing their own experiences with Apple’s bug bounty program.
We are retiring our bug bounty program [article]
349 points · 278 comments · turso.tech · 18h ago
Turso shut down its $1,000 bug bounty program because AI-generated slop submissions overwhelmed their maintainers — people were submitting obviously fake bugs like manually corrupting the database header and claiming it proved a vulnerability. The HN thread largely agreed this is a textbook example of the harms AI skeptics have been warning about: low-effort, low-cost output that costs humans hours to review, and the incentives are completely misaligned since generating a fake report takes a minute but evaluating it takes an expert. A big split emerged over whether the solution is to “just shut it down” (move to invitation-only contributions) or accept that AI is here to stay and figure out some other mechanism like charging a fee per submission. Several commenters pointed out that banning users doesn’t work because creating a new identity is free, and the whole dynamic is a Sybil attack problem that no individual project can solve on its own. A smaller but sharp tangent compared this to the “tactical tornado” engineer who writes huge, sloppy PRs that slow the team down — except now that engineer has twenty AI agents running in parallel.
Amazon workers under pressure to up their AI usage are making up tasks [article]
345 points · 388 comments · www.fastcompany.com · 18h ago
The linked article wasn't available to this summarizer; from the discussion, it's about Amazon managers pressuring workers to increase their AI usage, leading employees to invent pointless tasks just to burn tokens. The HN thread immediately invoked Goodhart's Law—when a metric becomes a target, it ceases to be a good measure—and split hard over whether this is pure waste or an intentional, if blunt, strategy to force experimentation and discovery of real AI use cases. A vocal camp argued that burning millions on busywork is insane and that the real payoff is inflating adoption metrics to impress investors, while others countered that the 20% of employees who figure out genuinely productive uses justify the 80% waste. The conversation ricocheted into side debates about LLM-generated documentation being worse than none, Cursor gaming its "AI-edited" line count, and the profound tension between using AI to replace knowing a single command and the long-term risk of losing operational understanding entirely.
'No way to prevent this,' says only package manager where this regularly happens [article]
335 points · 156 comments · kevinpatel.xyz · 7h ago
The article is a satirical Onion-style takedown of the npm ecosystem’s hand-wringing over yet another supply chain attack, mocking the refrain that “there’s no way to prevent this” while pointing out that Go, Rust, and other ecosystems with leaner standard libraries and built-in verification don’t have this problem. HN immediately pushed back on the premise, with several people pointing out that PyPI/pip is actually *more* dangerous than npm because it lacks lockfiles, though adoption of uv is slowly fixing that. Others argued that Rust’s `build.rs` and proc macros are essentially the same attack surface as npm’s postinstall scripts, and that the real difference is just that JavaScript’s culture of tiny, deeply-nested dependencies makes the problem more visible. A long thread dug into the root cause: underfunded, volunteer-run registries can’t afford even basic security measures like namespacing or publisher verification, while well-funded ecosystems like Maven Central (backed by Sonatype) and NuGet (backed by Microsoft) have had guardrails in place for years. The consensus was that removing install scripts alone wouldn't fix anything—the malware would just shift to build-time execution—and that the real solution is either sandboxed build environments or a fundamental shift toward fewer, larger, vetted dependencies.
ABC News has taken all FiveThirtyEight articles offline [article]
329 points · 147 comments · twitter.com · 12h ago
ABC News has fully nuked the FiveThirtyEight archives, pulling down every old article from a site that was already put out to pasture over a year ago. Most of the thread is just frustration at a big media company buying something and then deliberately erasing it out of spite — especially after Nate Silver said ABC refused to sell the IP back to him because he’d criticized their management. A good chunk of the discussion is people re-litigating 538’s prediction accuracy, with the usual split: defenders hammering that a 30% chance happens plenty (and that the 2016 call was actually pretty good), while critics argue Silver drifted into punditry and got too punchy for his own good. Some side tangles about whether this is a breach of fiduciary duty (mostly no, because corporate spite isn’t illegal) and a joke about Big Burrito suppressing the burrito bracket. The overall mood: this was a genuinely useful resource that’s now gone for no good reason.
Show HN: Find the best local LLM for your hardware, ranked by benchmarks [article]
279 points · 63 comments · github.com · 22h ago
The tool detects your hardware and suggests which local LLM to run based on merged benchmarks, but the thread quickly turned into a referendum on AI-generated slop projects hitting the front page. Several people noticed the README used 48 em-dashes, the author’s replies appeared copy-pasted from Claude, and a now-deleted `marketing.md` file literally instructed the LLM to post on HN and Reddit. On the technical side, the memory estimation can’t handle sliding-window attention (overestimates for Mistral/Gemma), and users found it recommending outdated Qwen2.5 models while missing newer Qwen3.x and GPT-OSS variants. The main split: a few thought the ranking logic was useful for deciding which fitting model to run, but the dominant takeaway was that this is a vibecoded project you shouldn’t install, especially since existing web tools like llmfit or canirun.ai do the same without requiring a local Python script of unknown provenance.
The Zulip Foundation [article]
262 points · 67 comments · blog.zulip.com · 13h ago
The Zulip founder announced he's stepping back to join Anthropic, donating the company to a newly formed nonprofit foundation modeled after Mozilla and Signal. The HN crowd split sharply: plenty of people praised the governance setup and long-term commitment to open-source values, but a vocal chunk called out the move to Anthropic as either a cash grab disguised as altruism or a naive belief that working on AI is "for humanity's benefit." Defenders pushed back hard, pointing to the founder's decade-long track record of doing right by the project and the fact that he specifically structured the transition so Zulip could survive without him. Some side discussion erupted over whether Zulip's threading model is actually too complex for small teams compared to Discord, and a few people flagged the Friday-afternoon timing as a classic "bury bad news" move—though board members said that was just when the paperwork landed.
Radicle: Sovereign {code forge} built on Git [article]
236 points · 81 comments · radicle.dev · 19h ago
The submission is Radicle, a peer-to-peer code collaboration stack built on Git that aims to break away from centralized hosts like GitHub by letting users run their own nodes and keep full control of their data. The thread largely cheered the spirit of the project but landed hard on its rough edges—multiple people described the UX as deeply unintuitive, with one detailed complaint that creating issues or submitting patches requires bouncing between CLI and web UI with no clear onboarding path, making it a non-starter for low-information contributors. A few commenters pushed back on the documentation itself, arguing the FAQ doesn’t even explain what a “forge” is or how Radicle differs from plain Git, while others pointed out that its CI story is still rudimentary (just webhooks to external systems), and a potential killer feature like VCS-agnostic support is planned but not there yet. There was also a tangential dive into using Radicle with Jujutsu for stacked PRs, and a mention that the project recently moved from .xyz to .dev because ISPs often block .xyz domains.
O(x)Caml in Space [article]
233 points · 56 comments · gazagnaire.org · 21h ago
A team got a pure-OCaml CCSDS protocol stack with post-quantum key rotation actually booted on a satellite in low Earth orbit, running as a guest on a hosted payload module. The thread immediately produced someone claiming to have put OCaml in space first back in 2016 on GHGSat-D, and a current GHGSat engineer confirmed their constellation of 16 satellites still mostly runs OCaml, though newer components are in Rust. The performance angle got real attention: OxCaml’s stack-allocation annotations dropped p99.9 dispatch latency from 29 ns to 9 ns and eliminated GC pressure entirely over 25 million packets, with a commenter noting they’d replicated the same pattern in their own HTTP stack. A split emerged over whether reinventing CCSDS security made sense versus using TLS, but the author pushed back that CCSDS is what the industry has and there’s no good open-source implementation of the full stack, while another space engineer argued SDLS is itself battle-tested and handles high-latency links where TLS breaks. The hiring complaint came up—several people said they’d kill to work on OCaml embedded systems, but the reply was that the pool shrinks further when you filter for domain expertise plus language skill.
Bitwarden scrubs 'Always free' and 'Inclusion' values from its site [article]
225 points · 126 comments · www.fastcompany.com · 19h ago
The linked article wasn't available to this summarizer; from the discussion, Bitwarden has removed “Always free” and “Inclusion” from its website, and its longtime CEO has stepped down to an advisory role while a new CEO with heavy M&A and private equity experience takes over. The thread largely treats this as the company being prepped for a sale, with multiple people pointing to the incoming CEO's LinkedIn history and drawing direct parallels to the Lastpass enshittification arc. A big split emerges around self-hosting: half the room says just switch to Vaultwarden (a compatible open‑source server) and keep using the Bitwarden clients; the other half warns that once Bitwarden's new owners decide to block third‑party servers in the client code, Vaultwarden becomes a fragile fork project that loses Bitwarden's security reputation. Meanwhile, KeePassXC with Syncthing gets pitched as the actually‑free alternative, but gets pushback over sync conflicts and the difficulty of sharing passwords with non‑technical family members.
Steve Jobs in Exile – New book on his years at NeXT Computer [article]
203 points · 174 comments · spectrum.ieee.org · 21h ago
The IEEE Spectrum piece promotes a forthcoming book arguing that Steve Jobs’s decade at NeXT wasn’t just a wilderness exile but the period where he learned discipline and where the software foundation for Apple’s modern OS was built. The thread largely agreed that NeXT’s influence is hardly forgotten — every macOS developer knows the “NS” prefix — and that the real debate was over which biography captures those years honestly, with several people pointing to *Becoming Steve Jobs* as far superior to the Isaacson hagiography. A few pushed back hard on romanticizing Jobs, noting that NeXT was a money-losing mess saved only by patient VCs and that his abrasive style wrecked careers, while others argued that constraint and failure were exactly what matured him into an effective leader. People also dug into the technical what-ifs, like whether BeOS could have won the battle to replace classic Mac OS, and the broader lesson that founders chasing perfection often miss that their real gold mine is the software platform — a point the article itself makes about AI today.
Palantir has hired more than 30 senior UK Government officials [article]
196 points · 62 comments · www.thenational.scot · 11h ago
The article reports that Palantir has hired 32 senior UK government and public sector officials since 2012, including figures from the NHS, Ministry of Defence, Home Office, and Downing Street, with transparency experts warning of an "acute risk" of corruption from the revolving door. HN mostly shrugged at the number, with one top comment linking to a story about Boeing hiring 840 former government officials to say "only 30?" — the real reaction was that this is business as usual for the UK civil service, not a scandal. The thread immediately turned into a heated moral debate about working for defense contractors: several people argued that cutting off friends who take such jobs is the right thing to do, while others pushed back hard, calling that approach toxic and noting that taxes fund the same military systems. Palantir’s own defense—that six of the hires are veterans and it's "inappropriate" to include them in a revolving-door narrative—got torn apart, with multiple ex-service members saying that veteran status shouldn't shield criticism of the military-industrial complex pipeline. A significant split emerged between commenters who see defense work as inherently corrupt and those who argue that without it, Ukraine or the Baltics would be overrun, leaving the conversation unresolved and ideological.
The sigmoids won't save you [article]
192 points · 187 comments · www.astralcodexten.com · 21h ago
The piece from Astral Codex Ten pushes back on the "all exponentials eventually become sigmoids" talking point used to dismiss AI progress, arguing it's technically true but useless without knowing when the curve will actually flatten. Most of the thread zeroed in on the specific graphs he used, especially the METR benchmark showing AI capabilities measured in hours of human work per task—several people noted that metric is suspect because it measures task completion time, not intelligence, and that recent improvements might just reflect better tooling or overfitting to the benchmark rather than real capability jumps. A significant split emerged between those who see this as a clever application of Lindy's law ("the trend will probably continue as long as it already has") and those who found the whole thing muddled and self-contradictory, with one person arguing the article fights its own assumptions and doesn't know what conclusion it wants. The deeper pushback came from skeptics who say the entire framing is wrong: model reasoning might be improving on an s-curve, but that's a different axis from human-like intelligence, which requires recursive self-reflection that LLMs simply don't have.
Waymo updates 3,800 robotaxis after they 'drive into standing water' [article]
186 points · 179 comments · www.cnbc.com · 14h ago
Waymo recalled 3,800 robotaxis after ones in Austin and San Antonio drove into flooded streets and stalled, prompting a NHTSA probe. The thread dove into the hard problem of telling wet pavement from standing water; a few commenters pointed out that Waymo’s dense HD maps could theoretically detect depth mismatch if water sits higher than the known road surface, but others pushed back hard on map freshness, noting how quickly roads change (brand new highways, sinkholes, earthquake damage) and that a database of every inch is impractical and stale. The conversation also spun off into humans’ own failures at judging flood depth—people trashing engines on fords in England—and a side debate on whether EVs handle deep water better or just float away. The consensus was that this isn’t a trivial software patch and that Waymo’s mapping-dependent approach has a fundamental vulnerability that no amount of pre-driving can fully eliminate.
ASCII by Jason Scott [article]
185 points · 22 comments · ascii.textfiles.com · 18h ago
Jason Scott posted a wrap-up on his decade-long project to rescue a warehouse full of technical manuals from the trash and get 13,000 of them scanned and uploaded to the Internet Archive. The thread is almost entirely people calling him a legend and sharing personal stories of chipping in donations or getting random manuals in the mail as perks—one person said they’d wondered for years what happened to that project and just found out it’s done. A few people noted his site was getting hugged to death (Scott acknowledged a bad vhost setting and said he was working on it). Somebody made a sarcastic joke about giving free training data to robot overlords, and someone else retorted that at least the LLMs might write better manuals than the originals.
Erlang/OTP 29.0 [article]
182 points · 29 comments · www.erlang.org · 8h ago
Erlang/OTP 29 shipped with a bunch of long-awaited improvements: native records (EEP-79, still experimental), SSH daemon now hardened by disabling shell and exec services by default, and a new `io_ansi` module for terminal styling. The thread immediately split on the SSH change—several people pointed out the daemon was never auto-started anyway, so it’s really about the defaults when you do start it, not a net new lockdown. The native records got the most heat, with Elixir folks wondering whether structs or maps could eventually compile down to them, and one person noting that old tuple-based records probably won’t be deprecated but the new type is a big deal for pattern matching. There was also the usual framework war: a Rails dev argued Phoenix is harder for solo work, but multiple Elixir veterans pushed back saying Phoenix’s explicitness pays off in long-term reliability and zero-crash uptime. One critical takeaway from the comments: if you’re still on an OTP release before 29, update now—someone flagged that security scans found multiple critical CVEs dated early 2026 in older versions.
Trade Dollars with other startups. Book it as revenue [article]
177 points · 152 comments · www.revswap.ai · 18h ago
The linked article wasn't available to this summarizer; from the discussion, RevSwap is a satirical "platform" that lets startups trade dollars with each other and book the exchange as revenue, taking a 2% cut that it then swaps with yet another platform. Hacker News immediately recognized the parody and ran with it, calling out how closely this mirrors real-world behavior like the circular deals between Nvidia and OpenAI, or infamous VAT carousels from the 1990s. The thread's running joke cycles through the classic "two economists eat shit and call it GDP" parable, with some pushing back that actual work-for-work barter is legitimate and the fraud only kicks in when you book the exchange as revenue without real value being created. Others dug into the tax implications, arguing that while this setup is clearly a joke, the underlying accounting tricks—like inflating art valuations for tax benefits or faking ARR to impress investors—are very real and already happening.
Meta to receive $3.3B in tax breaks for its $10B Louisiana data center [article]
173 points · 153 comments · fortune.com · 12h ago
The linked article details Louisiana giving Meta $3.3 billion in tax breaks to build a $10 billion AI data center, arguing these subsidies are wasteful for an industry that doesn't need them. The HN thread split hard on the economics: one side insisted the breaks are a necessary competitive move because Meta would just build in a state that doesn't charge sales tax on data centers, meaning Louisiana gets $0 otherwise, while the other side called it a race-to-the-bottom giveaway to a trillion-dollar company that buys politicians and provides a handful of permanent jobs at massive environmental and infrastructure cost. Several people pushed back on blaming lobbying alone, noting that union construction workers and local utilities also want these projects, and that the real fix requires states to cooperate on a minimum tax standard — something the U.S. just torpedoed by pulling out of the global Pillar Two agreement. A smaller contingent veered into Meta’s AI products, with people mocking Zuck’s focus on virtual AI friends and venting about Facebook’s broken moderation system, which seems to ban users arbitrarily while letting real abuses slide.
SQL patterns I use to catch transaction fraud [article]
158 points · 44 comments · analytics.fixelsmith.com · 8h ago
A program-integrity analyst argues that transaction fraud detection is mostly just SQL—six patterns like velocity, impossible travel, and round-dollar amounts—and that ML hype is overblown. HN immediately pushed back on the “nobody buys coffee for exactly $1.00” example, pointing out that round-number prices are common outside the US and that things like “round up to donate” at checkout break the premise. Several commenters also called out the off-hours pattern as naïve—lots of legitimate reasons to buy gas at 2am—and noted that the article’s claim of being based on real work conflicts with its disclaimer that nothing comes from actual cases. More than a few suspected AI generation from the writing style, while others debated whether banks actually use deterministic SQL or rely on ML for real-time decisions, with one side saying SQL is better for explainability.
Image-blaster: Creates 3D environments, SFX, and meshes from a single image [article]
155 points · 31 comments · github.com · 16h ago
The linked article wasn't available to this summarizer; from the discussion, the tool takes a single 2D image and spits out a full 3D environment, SFX, and object meshes by orchestrating multiple AI models like World Labs' Marble 1.1 (for Gaussian splatting) and Tencent's Hunyuan3D (for individual objects). The crowd immediately compared it to Blade Runner's Esper photo analysis and Microsoft's old Photosynth, but the consensus is that while it's genuinely impressive for a one-shot pipeline, outside of curated demos the results are often a mess—outdoor scenes hallucinate nonsense geometry, the topology is terrible (hundreds of thousands of polygons requiring heavy retopology), and you can't wander far from the original camera angle without everything falling apart. A few practitioners noted that Meshy.ai and TRELLIS are comparable for object modeling but also suffer from the same quality ceiling, and one tangent spun off into how shockingly hard it is to get AI to generate consistent isometric sprites for games, with the suggestion that you're better off just learning to draw. The underlying tech debate pitched NeRF-based approaches against naive gaussian splatting, but the main takeaway is that this stuff is moving fast and the industry hasn't caught up to the possibility of scriptable, on-the-fly 3D assets—even if the output still needs a lot of cleanup.
Power Tools Got Worse on Purpose. Who Owns DeWalt, Craftsman, and Milwaukee? [article]
149 points · 129 comments · www.worseonpurpose.com · 19h ago
The article compares how two conglomerates—Hong Kong's TTI and America's Stanley Black & Decker—handled their tool brand acquisitions, with TTI investing heavily in Milwaukee while SBD gutted Craftsman and Porter-Cable through cost-cutting and mismanagement. The HN discussion largely pivots to Makita as the un-enshittified alternative, with multiple tradespeople sharing repair stories and arguing its repairability and customer service blow Milwaukee and DeWalt out of the water, though they concede Makita's US product line is still solid and its battery microwave is a hit. A more skeptical thread pushes back on the private-equity-is-always-evil framing, arguing that dozens of quiet, profitable PE acquisitions never make headlines and that the strip-mining strategy isn't universal. Unexpectedly, the comments take a detour into the Arrow T50 stapler's quality collapse after a factory move, with a complaint that MBAs treat brand reputation as a commons to burn for short-term gain.
“Too dangerous to release” or just too expensive? [article]
146 points · 173 comments · kingy.ai · 19h ago
The article argues that Anthropic’s gated release of its Claude Mythos Preview model is driven by a mix of genuine zero-day exploitation dangers and compute-cost constraints, weighing the evidence for each. HN immediately split into camps: many called the “too dangerous” narrative a transparent marketing stunt, pointing out that Dario Amodei pulled the same playbook with GPT-2 in 2019, while others countered that the real story is just a compute shortage—Anthropic can’t afford to serve the model at scale and the safety talk is convenient cover. A strong undercurrent of cynicism ran through the thread, with people noting the article itself reads like AI slop and that the site collapsed under HN traffic, which one commenter joked was Mythos itself taking it down. The consensus among skeptics was that the model is probably only incrementally better and the hype is designed to lock in enterprise contracts before the IPO, not because it’s genuinely too dangerous for the public.
The old world of tech is dying and the new cannot be born [article]
143 points · 137 comments · www.baldurbjarnason.com · 19h ago
The linked article is a long, sprawling essay arguing that the US tech industry is built on the foundation of American global hegemony, which the author believes is now collapsing due to strategic failures (specifically framing the Iran conflict as a decisive loss for the US). HN was deeply split: several commenters pushed back hard on the geopolitical premise, arguing that Iran is actually at a strategic nadir and that US tech dominance—especially in AI, SpaceX, and capital markets—is stronger than ever, with one person calling the piece "delusional." Others strongly defended the essay, calling it one of the best they've read in years and zeroing in on the idea that globalized tech platforms only function because the US enforces a single set of rules via its economic and military power. A long tangential thread erupted over whether English's flexibility as a language gives it an insurmountable moat for tech dominance, with some arguing it's purely historical inertia and others claiming the language is uniquely suited to absorbing new concepts. The conversation never really engaged with the article's central literary metaphor about Christian millenarianism, instead focusing almost entirely on whether the author's geopolitical thesis is empirically correct or just a sweeping narrative built on unchallengeable anecdotes.
London Police Deploy Facial Recognition at Protest for First Time [article]
140 points · 101 comments · reclaimthenet.org · 11h ago
The Metropolitan Police are deploying live facial recognition and drones at a Tommy Robinson–organised rally in Camden, marking the first time the technology has been used at a UK protest, while a much larger pro-Palestinian march on the same day avoids the same surveillance. The thread zeroed in on the obvious double standard, with many pointing out that the Met is citing vague “intelligence” about potential violence from an unspecified portion of attendees to justify scanning everyone’s face—a justification that wouldn’t fly for left-wing protests. A lot of the pushback came from people who noted that Reform UK itself distanced from Robinson, so this isn’t even a mainstream party event, yet the police are still treating it as a pretext to expand biometric surveillance from vans to permanent lamppost cameras with no parliamentary vote. Several commenters called out the absurdity of the UK’s surveillance state generally—CCTV everywhere didn’t stop David Cameron’s bike from being stolen—and argued the real purpose is tracking political affiliation, not catching criminals. The consensus was that this crosses a line: once facial recognition is normalized at protests, the question isn’t whether it will be used more broadly, but when.
Generated 2026-05-16 08:04 UTC
Generated by Sauron from Hacker News discussions and linked articles.