HN Brief: 2026-08-17
Today's HN was dominated by three threads poking at the power dynamics of infrastructure: Mozilla's new iOS adblocker sparked a fight over whether it's hypocritical revenue-protection (it exempts its own ads), Cloudflare was caught silently injecting JavaScript into users' sites after a nameserver switch, and Stripe's $7B+ acquisition of OpenRouter was debated as a smart tollbooth play or a thin-proxy valuation. A separate throughline emerged around model behavior and control—users discovered Qwen 3.8 over-thinks everything by default (17,000 reasoning tokens for a simple tool), Anthropic's published system prompts have ballooned to 3,000 words of "nonsense," and a piece argued models are being deliberately dumbed down on factual recall in exchange for reasoning ability.
The threads most worth clicking: "Claude: System Prompts" for the shock of seeing how much real-world context (export controls, codenamed models) gets shoved into a prompt to patch training cutoffs; "Firefox for iOS now has a native adblocker" for the sharp split between those who see the ad exemptions as hypocrisy and those who note no default adblocker blocks them; "A third world engineer responds to 'RISC-V: They should have known better'" for a rare meta-debate about whose perspective counts in architectural critiques; "Tell HN: Cloudflare silently injects its analytics when you switch nameservers" because silent HTML injection by a TLS-terminating proxy is a man-in-the-middle move, full stop; and "Models Are Getting Dumber on Purpose" for the growing tension between coding/agentic users who love the reasoning shift and everyone else who just wants a reliable fact-check.
Firefox for iOS now has a native adblocker [comments]
645 points · 255 comments · support.mozilla.org · 19h ago
Mozilla announced that Firefox for iOS now includes a native adblocker, built directly into the browser rather than requiring a separate app like the older Firefox Focus. The HN crowd immediately noticed the fine print: the blocker explicitly exempts ads on search engine result pages and on Firefox's own New Tab page, which led to a sharp split between people calling it hypocritical revenue-protection and others pointing out that no adblocker (even uBlock Origin on desktop) blocks those categories by default. The bigger debate quickly shifted to why Firefox can't support real extensions on iOS at all, with a long, heated argument about whether Apple's WebKit-only policy is security-conscious or anti-competitive—some argued Mozilla is right to refuse a hacky extension implementation like Orion's, while others said the real enemy is Apple's malicious compliance with EU regulations, which has so far prevented any third-party browser engine from actually launching. A secondary thread reminisced about Firefox Focus being a system-wide content blocker for Safari, with several people surprised to learn they could have been using it that way all along.
Claude: System Prompts [comments]
635 points · 252 comments · platform.claude.com · 19h ago
The article is Anthropic publishing the actual system prompts that power Claude's web and mobile interfaces, showing how they've ballooned from ~300 words to over 3,000 as models have scaled up. Simon Willison helpfully turned them into a git history, and the big finding is that these prompts now contain elaborate explanations of real-world events like U.S. export controls temporarily blocking access to newer models "Fable 5" and "Mythos 5" — Claude has to be told about these because they happened after its training cutoff. There's a sharp split on whether these bloated prompts actually help or hurt: one side says they're "useless nonsense" eating precious context window and degrading performance, while the other argues they're necessary behavioral guardrails that get incrementally added like building codes written in response to actual failures. A recurring frustration is that Anthropic publishes the chat prompts but deliberately omits the tool definitions and Claude Code prompts, which people point out are trivially extractable with a logging proxy anyway, making the omission feel performative rather than protective.
A third world engineer responds to “RISC-V: They should have known better” [comments]
482 points · 250 comments · rvembedded.com · 15h ago
The linked article is a response to Dmitry Grinberg’s earlier critique of RISC-V, written by an embedded engineer in Trinidad and Tobago who argues that Grinberg’s architectural complaints miss the point for the global majority. The author concedes Grinberg’s technical gripes—like weird compressed store offsets and Zicsr being a separate extension—are real, but says the critic’s own first-principles derivation of a cheap microcontroller core (RV32EC) actually proves RISC-V will win the low end on cost and availability, not elegance. The HN thread quickly got sidetracked by a meta argument about whether the site itself is still Bay Area-centric or has shifted toward a DACH/European userbase, with several people pushing back hard on the shipping-cost claims by pointing out that shipping small chips from Asia to places like Nigeria and Bangladesh isn’t nearly as expensive as the article makes it sound. A few commenters also noted the irony of Grinberg’s privilege-tossing language coming from a position of assumed geographic and economic centrality, while others dug into the real architectural debate about whether one ISA can genuinely serve both ten-cent parts and MMU-equipped chips running Linux—and largely landed on the side that RISC-V’s fragmentation _is_ its scalability.
Tell HN: Cloudflare silently injects its analytics when you switch nameservers
460 points · 117 comments · news.ycombinator.com · 14h ago
A Hacker News user reported that after switching nameservers to Cloudflare to serve an R2 bucket under their subdomain, they discovered Cloudflare had silently injected a JavaScript analytics snippet into their otherwise JS-free HTML site—they had to enable analytics in the dashboard just to access the control to disable it. The thread quickly split between people who dismissed this as well-known behavior when using Cloudflare's reverse proxy (the orange-cloud proxied records) and those who were genuinely alarmed, noting they'd never intentionally enabled analytics and found the same injection on multiple sites. Several people pointed out that if you're using Cloudflare as a pure DNS provider (grey-cloud records) this can't happen, but the submitter and others argued the proxy itself was enabled by default without clear opt-in, and that injecting scripts without consent is a man-in-the-middle move regardless. A deeper concern emerged: since Cloudflare terminates TLS and can rewrite your HTML for analytics, they can also snoop on or modify traffic in other ways, and the fact that an obscure blog post about enabling web analytics by default got almost no attention on HN 11 months ago suggests the enshittification pattern is already in motion.
Research papers using "kidney disappointment" instead of "kidney failure" [comments]
392 points · 136 comments · scholar.google.com · 19h ago
The HN thread is about a Google Scholar search revealing research papers that use the bizarre phrase "kidney disappointment" instead of the standard medical term "kidney failure." The consensus quickly formed that this is a classic case of "tortured phrases"—clumsy synonym substitution used to evade plagiarism detectors, likely from old-school article spinners or poor machine translation, not from modern LLMs. People dug up related gems like "renal disappointment," "lactose bigotry," and a chemistry paper where "the final solution" got replaced with "the mass killing of an ethnic group," turning the discussion into a morbidly hilarious catalog of thesaurus abuse. The pushback was that this pattern predates ChatGPT by years, pointing to tools like SpinBot and Google Translate as the real culprits, and the thread veered into sharing anecdotes of similar mangled language from peer-reviewed research.
Qwen 3.8 27B is excellent, but it defaults to overthinking things [comments]
383 points · 171 comments · simonwillison.net · 8h ago
Simon Willison published a hands-on review of Alibaba’s new Qwen 3.8 27B, a vision-capable open-weight model that fits in 17GB and runs on a laptop — and he’s genuinely impressed by its coding, image bounding boxes, and agentic tool calling, but he’s frustrated it ships defaulting to an “xhigh” reasoning effort that burns tens of thousands of tokens overthinking even a request for an SVG of a circle. The HN crowd immediately validated that default is a problem: one person reported a task that took eleven hours on dual GPUs with xhigh, while the same task clocks twenty minutes on GPT-5.5, and someone else ran a side-by-side against the new Muse Glimmer 30B and found Qwen used 17,576 reasoning tokens to produce a beautiful, over-engineered HTML tool, while Glimmer used just 1,021 tokens for something functional but uglier. The obvious fix is to turn reasoning down to “low” or off — and when you do, the model is nearly as good as the previous Qwen 3.6 35B at medium thinking — but some argued that even the verbose thinking is useful: Simon himself noted that with no reasoning, the model placed bounding boxes in the wrong spot, while xhigh nailed it. A few people have hacked their own workarounds — a proxy that cuts off thinking after 2K tokens and injects “we’ve thought enough” — and others are playing with a community LoRA called ThinkingCap that claims to halve thinking tokens while keeping quality, but the consensus is that if you control the reasoning effort, this is the best self-hostable model yet, held back only by raw token speed (15–30 tok/s on consumer hardware).
Stripe will reportedly acquire OpenRouter for $7B+ [comments]
320 points · 204 comments · techcrunch.com · 11h ago
Stripe is reportedly buying OpenRouter for over $7 billion, a startup that acts as a unified API gateway letting customers pick and route between different AI models without managing multiple vendor accounts. The thread zeroes in on whether this valuation makes any sense, with several people arguing OpenRouter is just a thin proxy with no real moat, especially since most providers already use a near-common OpenAI-style API and open-source alternatives like LiteLLM exist. Others push back hard, saying the real value is centralized billing and procurement—OpenRouter gives you one bill and one approval process across 400 models, which is huge for enterprise compliance, and Stripe buying them locks up that tollbooth on AI token flow the same way they dominate payment flow. A secondary debate erupts over whether you can trust OpenRouter to keep data away from Chinese providers, with skeptics calling it a fly-by-night Aliexpress for models while defenders counter that with $7 billion on the line and transparent logging, the risk is no worse than trusting any cloud provider.
Models Are Getting Dumber on Purpose [comments]
310 points · 172 comments · w4g1.dev · 12h ago
The article argues that model labs are deliberately trading factual knowledge for reasoning ability, producing models that ace math benchmarks but hallucinate on simple factual questions at alarming rates. HN latched onto the hallucination data hard—the thread notes those SimpleQA numbers are already outdated, and points to a newer "SimpleQA Verified" benchmark for current readings. A major split emerged between coding/agentic users, who see this shift as perfectly fine since retrieval can supply facts at runtime, and everyone else who uses LLMs for learning, curiosity, or trivia and finds 80% hallucination rates unacceptable. The most active debate was around a commenter's dream of pluggable, laser-focused models that snap together Swift or electronics knowledge on demand, with others pushing back hard that you can't just swap knowledge modules like Lego bricks—coding patterns transfer across languages, and a model that knows only Swift is provably worse at Swift than one trained on multiple languages.
The AI Credit Resale Economy [comments]
277 points · 113 comments · vectoral.com · 17h ago
The article maps out a growing grey market where people buy up unused AI inference credits—often from startups that got them for free through accelerators like YC or provider grants—and resell them at steep discounts, sometimes up to 98% off. The HN thread immediately got skeptical, with people pointing out that those deep discounts almost certainly come from stolen API keys, fraudulently obtained trial accounts, or bait-and-switch routing where the broker sends you DeepSeek or a Kimi distillation instead of the Anthropic model you paid for. A big chunk of the discussion argues that the real money isn’t just arbitrage—brokers sell the same tokens twice, once to the buyer and once as training data to labs doing distillation, turning the transaction into a data-stealing operation. Others note this is just the latest iteration of decades-old abuse patterns (loyalty points, hotel credits) and that the AI companies could shut it down by tracing relay IPs back to the source accounts, but they might not bother because even fraudulent traffic pads their usage KPIs. There’s also a side thread about how absurdly easy it is to spin up a relay using open-source proxy gateways, and a warning that anyone buying these ultra-cheap tokens should assume their prompts are being logged or that the proxy will inject malicious tool calls.
The federal keyword lists that canceled billions in research funding [comments]
213 points · 100 comments · www.highereddive.com · 7h ago
The linked article wasn't available to this summarizer; from the discussion, it's about how the current administration used specific keyword lists to cancel billions in research funding, with grants pulled for mentioning things like "inequalities" or "diversity" in a technical context. The thread spiraled hard into a broader argument about whether this is raw incompetence or deliberate malice mixed with grift, with most people landing somewhere on the "maliciously incompetent" scale and one camp insisting it's a feature, not a bug—chaos engineered so insiders can cash out while the next administration gets blamed. Several people shared firsthand accounts of grant applications getting axed for absurd keyword hits (math professors told to avoid "inequalities," geology grants flagged for "mineral inclusion"), and the conversation devolved into comparing this to Canada's earlier grant-writing DEI requirements and whether the current U.S. situation is worse or just differently stupid. A recurring argument was that trying to explain this as foreign agent infiltration is a cop-out, and that the real rot is homegrown American greed and cruelty, with one person quoting Russell Vought saying they want bureaucrats "traumatically affected" to make the point that the damage is intentional.
Nvidia dramatically reduces amount of OpenAI infra financing it may guarantee [comments]
199 points · 94 comments · www.reuters.com · 10h ago
Reuters reports that Nvidia has pulled back on its guarantee to finance OpenAI's massive data center construction, scaling down from earlier commitments to a $500 billion project. The thread quickly zeroed in on the circular financing scheme underpinning the whole thing—Nvidia selling chips at huge margins while backstopping the loans that pay for those same chips, essentially turning itself into a bank that happens to design GPUs. Several people ran the math and concluded that even if the backstop gets fully written off, Nvidia’s 75% gross margins still leave it profitable, so the real bagholders are SoftBank, pension funds, and sovereign wealth funds. A long subthread debated whether frontier models will eventually run on consumer GPUs, with one camp arguing inference is always the cheap part and the real cost is training, while the other side pointed to history—what’s frontier today runs on a potato in a few years—and noted that SoftBank’s bizarre “400-year plan” investor presentation suggests the people funding this don’t have a sober grasp of the economics.
The weekend is 100 years old [comments]
194 points · 140 comments · www.theguardian.com · 16h ago
The Guardian piece traces the 100-year history of the two-day weekend, from Henry Ford’s 1926 five-day workweek to the Soviet Union’s disastrous random-day-off experiment, arguing that collective leisure time is now fraying due to hybrid work, smartphone intrusion, and the rise of the four-day week. Hacker News seized on the Soviet angle, with several people noting that the policy failed precisely because a day off is worthless without everyone sharing it—you can’t have family gatherings or social life if your spouse and kids are on different schedules. The comments quickly spiraled into a broader debate about how recent the seven-day week itself is, with references to the French Revolutionary ten-day week and pre-industrial agriculture where "weekends" didn’t exist because the work depended on weather and seasons, not clocks. A few pushed back hard on the article’s framing that capitalism or workers’ rights are younger than 100 years, citing the Dutch East India Company and Roman fiat currency, while others wondered aloud whether AI will finally give us a true three-day weekend or just let bosses extract more output.
Anthropic's 'Watermark' Text Adulteration in Claude Is a Perversion of Writing [comments]
186 points · 182 comments · daringfireball.net · 10h ago
John Gruber’s piece on Daring Fireball argues that Anthropic’s plan to embed semantic watermarks into all Claude text output—biasing word choices toward “green” and “red” lists to leave a detectable fingerprint—is a fundamental perversion of writing, because any synonym substitution sacrifices precision and clarity for the sake of surveillance. The HN thread immediately split: a large camp of practicing engineers and researchers pushed back hard, pointing out that LLMs already introduce randomness through sampling temperature and top-k/p strategies, so Gruber’s demand for “the best word every time” misunderstands how models work—many commenters argued the watermark merely swaps one equally-good token from the same distribution, making his outrage technically illiterate. A more pragmatic contingent focused on the implications for proofreading and code: Anthropic’s own documentation admits watermarking is negligible for math and structured code, but the thread zeroed in on the eerie consequence that if you use Claude to lightly edit your own prose, the watermark could later flag the entire document as AI-generated, effectively destroying the reliability of collaborative human-AI writing. Several commenters also flagged the EU regulation’s terms-of-service requirement as a landmine—Anthropic must forbid users from stripping watermarks, which raises the specter that nobody truly owns or can freely use the text Claude produces, and the scheme remains a secret-key black box only Anthropic can verify, making it useless as a public authenticity tool and ripe for abuse.
St Lucie Nuclear Reactor Unit 1 manually shutdown, 3 control rods drop into core [comments]
186 points · 138 comments · www.wptv.com · 16h ago
The St. Lucie nuclear plant manually shut down Unit 1 after three control rods unexpectedly dropped into the core, with the NRC classifying it as a non-emergency and the plant now back at full power. HN's nuclear-knowledgeable crowd immediately dug into the NRC filing and explained that this is a boring, routine fail-safe event — the rods are designed to drop on power loss or electrical fault, and operators correctly scrammed because running with unexpected rod positions falls outside the approved operating envelope. The real debate swerved hard into whether AI should ever control critical infrastructure, sparked by someone praising nuclear's boring safety record as a model for future AI-governed systems. Commenters split sharply: one side argued that LLMs are inherently unpredictable risk-multipliers that shouldn't touch life-or-death systems, while others pointed out that people already trust AI for medical and legal work, and that the real question is architecting failsafe "trip" mechanisms for complex autonomous systems rather than pretending they won't be deployed.
The Life and Death of Direct File [pdf] [comments]
183 points · 90 comments · www.ischool.berkeley.edu · 7h ago
The linked article wasn't available to this summarizer; from the discussion, it’s a post-mortem on the IRS Direct File program, a government-built free tax filing system that was shut down. The thread is split between people who see it as a successful proof-of-concept that was killed by political meddling—specifically Intuit’s lobbying and a single executive with personal animus—and those arguing it broke the rules by building a full product when Congress only authorized a $15 million study, spending $41 million before anyone had permission to launch. A significant chunk of the debate pivots to the broader ideological question of whether government should build digital infrastructure at all, with one side pointing to the interstate highway system and the Apollo program as evidence it can work, and the other insisting contracting out is always cleaner, despite allegations that contractors routinely bilk taxpayers on billable hours. Several people make the pointed observation that the government already accepts electronic filing from corporations without requiring a private intermediary, and that killing Direct File forces millions of individuals to hand their most sensitive financial data to TurboTax for no structural reason.
NIH is ending a key grant for budding clinical researchers [comments]
150 points · 76 comments · www.science.org · 15h ago
The linked article wasn't available to this summarizer; from the discussion, the NIH is ending a key grant program that funnels money and support to early-career clinical researchers. The thread is largely united in seeing this as deliberate malice, not bureaucratic incompetence, with several people pointing to the Trump administration's open hostility to federal scientific institutions as the obvious motive. A major thread explores why someone like Sean Eddy, a foundational figure whose lab just got defunded, can't simply move his curiosity-driven research into big pharma—the consensus is that pharma won't fund public-good tool-building and foundational science, sparking a broader debate about whether the golden age of industrial research (Bell Labs, Xerox, IBM) is a myth or a lost model. There's also a tense split between those who think the public supports NIH funding once they know what it is, and those arguing that a significant slice of high-IQ tech voters actively chose this outcome, either because they didn't believe the campaign promises or because they prioritize culture war and immigration stances over science funding.
Protobuf has LSP support [comments]
147 points · 98 comments · buf.build · 13h ago
The article announces Buf's release of a new Language Server Protocol server for Protobuf, bringing features like go-to-definition and code completion to editors that never had proper IDE support for the schema language. The HN thread immediately splits into two big fights: one camp argues Protobuf's strict schemas and versioning ceremony are becoming obsolete in the LLM era, while the other fires back that LLMs make structured contracts *more* important, pointing out that major LLM platforms themselves use ConnectRPC and Protobuf internally. A separate thread tears into Protobuf's performance from a Python perspective, claiming dropping gRPC for pure JSON results in faster and more reliable Google Cloud clients, though that gets met with pushback about unfair comparisons and the inherent encoding speed advantages of protobuf. Someone also raises the good point that renaming fields in protobuf is actually fine as long as the field ID stays the same, correcting the article's implication that many common LSP operations conflict with protobuf's compatibility guarantees.
ICE Shot a Journalist and Threw Him in Detention. He's Approaching 300 Days [comments]
125 points · 53 comments · theintercept.com · 11h ago
The Intercept reports on Ricardo Parias, a journalist shot by ICE agents during an arrest, then held in immigration detention for nearly 300 days in atrocious conditions where he continues reporting from his cell. The thread immediately split over credibility — one side cited the Justice Department's press release painting Parias as a violent fugitive, while others pointed to the mounting evidence of government misconduct, including agents deleting surveillance footage and a judge dismissing all criminal charges. Several commenters pivoted to broader claims about ICE force-feeding detainees, framing this as part of a systematic pattern rather than an isolated case. The political argument got ugly fast: some argued the judiciary's tolerance of executive abuse will only correct itself after Trump dies, while others shot back that there's no pendulum swing coming, and that hoping for one is naive when the entire system has been corrupted from the top down.
A U.S. Strategy to Prevent the Creation of Mirror Life [comments]
124 points · 87 comments · www.rand.org · 18h ago
A RAND Corporation report outlines a proactive strategy to prevent the development of "mirror life"—organisms built from reversed-chirality molecules that could evade immune systems and cause ecological collapse if released. The Hacker News thread immediately wrestled with the core feasibility question, with several domain experts pushing back hard against the assumption that such life couldn't find food, pointing to achiral nutrients like CO2 and water that both normal and mirror organisms can consume, and citing Stanford research showing engineered E. coli can grow on them. A major split emerged between those who praised the report's rare call for international cooperation and mutual restraint rather than surveillance and control, and skeptics who argued this naively assumes good faith from powers like China, where much of the cutting-edge work is already happening at Westlake University. Some commenters dismissed the whole idea as a non-threat, arguing that if mirror life were viable it would have already emerged through billions of years of evolution, while others countered that the precautionary principle is warranted since a single successful mirror cell could spread before we have defenses. A meta-thread also broke out over the use of ALL CAPS for emphasis, with some calling it a Zoomer affectation and others noting it's just digital shouting that's been annoying people since the 90s.
Banned Words List [comments]
118 points · 50 comments · pen.org · 19h ago
The linked article wasn't available to this summarizer; from the discussion, it's a list of hundreds of words flagged by the current administration for removal from federal agency communications, grants, and websites. The thread immediately split into a heated, definitional debate over whether this counts as "censorship" — one side argues that the government simply choosing what it will and won't say is the government's right under the First Amendment, citing a Supreme Court case, while the other side fires back that this is a de facto ban with chilling effects on researchers who have to self-censor grant proposals or risk losing funding. A few people fact-checked the literal list, noting that about 200 of the ~350 terms come from an HHS lawsuit and that some words like "opioids" are still clearly in use on CDC pages, which undercut the more alarmist claims. The conversation also drifted toward broader comparisons, with one participant arguing Australia's post-1973 approach of simply dropping racial classifications from the census worked better than the US's obsession with race, while others insisted the administration's real intent is fostering white supremacy and destroying public data, calling this just one piece of a larger, unconstitutional data-erasure campaign.
Young People Hate AI CEOs So Passionately That It's Almost Hard to Believe [comments]
113 points · 116 comments · futurism.com · 10h ago
The article reports on a CNBC poll showing that U.S. adults aged 18-34 overwhelmingly distrust nine major AI CEOs, with 81% distrusting Palantir’s Alex Karp and even the “best” performer, Microsoft’s Satya Nadella, only earning trust from 35% of respondents. The thread split sharply between people who think this is obviously deserved—pointing to decades of enshittification, data privacy violations, and CEOs openly saying AI will eliminate jobs—and those who argue the media manufactured the hostility by relentlessly dunking on AI since its early days. A long, heated subthread devolved into a debate about whether Uber and food delivery actually made life worse, with older commenters insisting the pre-Uber taxi experience was so terrible that any nostalgia is delusional. Another major tangent saw someone arguing that the real enemy isn't AI but the capitalist system that values the robot over the worker, drawing accusations of naive Marxian wish-casting from people who think AI is just a resource-hungry bubble that will never deliver on its promises.
Reticulum – Decentralized Mesh Network [comments]
107 points · 23 comments · reticulum.network · 8h ago
The linked article presents Reticulum as a cryptography-based networking stack designed to build sovereign, decentralized mesh networks that can operate on minimal bandwidth and under adverse conditions, positioning itself as an alternative to mainstream internet infrastructure. The HN crowd was split between genuine curiosity about practical use and sharp skepticism about the project's viability, with several people pointing out that Reticulum appears to be a near-burned-out one-person project taking on an enormous task, making Meshcore a more realistic near-term choice for mesh networking. A deeper technical pushback came from someone who noted Reticulum's claim of not using source addresses is misleading, since observer nodes can still triangulate message origin from entry points into the network—reminiscent of MeshCore's same vulnerability. Others dug into the implementation situation: while the project claims no published spec, someone countered that they built a working clean-room Rust implementation from the existing spec, though the maintainer's stance that "the implementation is the specification" has led to a proliferation of broken vibe-coded forks. The thread also wandered into appreciation for Neal Stephenson's *Anathem* sharing the name, a brief comparison to Netbird as a Tailscale alternative, and a mention of ratspeak.org as a modern Rust fork with an active community.
A True Telnet BBS on a Casio Calculator [comments]
103 points · 11 comments · ei3lh.eu · 19h ago
Someone got a Casio VX-4 pocket computer (essentially a calculator with BASIC and an RS232 port) and turned it into a fully operational telnet Bulletin Board System, hosted entirely on 8kb of RAM with a Raspberry Pi Zero W acting only as a network bridge and message archive. The thread was short on pushback—almost everyone loved the hack for its sheer, joyful absurdity, with multiple people reminiscing about owning similar Casio pocket computers as kids or calling out specific models like the FX-730P and the IR-7000 communicator. A couple of commenters noted the blog’s font and color contrast made it genuinely hard to read, but they explicitly separated that critique from the project itself, which they praised. The only substantive correction came from someone pointing out that these are pocket computers, not calculators, and that the two markets had different hardware and fates.
Ask HN: Do you know of any company that went back to hand-written code?
99 points · 118 comments · news.ycombinator.com · 20h ago
From the discussion, it's an Ask HN post asking whether any companies have actually abandoned LLM-generated code and gone back to writing software entirely by hand. The thread is split sharply between defenders who argue AI's productivity gains are undeniable and critics who counter those gains are a mirage—that AI lets five engineers build a legacy-code nightmare in six months, not a real asset. A startup founder shares the most concrete example: his team agreed that AI-accelerated development created such messy, unstable code that they're considering a partial rewrite without it, but his developers are resistant to losing the speed. The core debate isn't really about the technology's capability—it's about whether the ease of churning out code destroys discipline, with one side insisting that's a leadership failure and the other claiming the tools themselves erode the incentive to build anything solid.
The Case Against Formal Verification, 50 Years Later [comments]
90 points · 101 comments · ivan-gavran.github.io · 11h ago
The article revisits a famous 1979 paper that argued formal verification would fail, responding to each of its six arguments with modern developments in mind—especially the rise of AI coding agents that both create a need for verification and make it easier. The HN thread largely agreed with the article’s core point that specifications, not verification itself, are the weak link, but many pushed back hard on the feasibility of formally specifying anything as messy as “Facebook” or anything with a GUI, arguing that for many programs the shortest description is the code itself. A few seasoned engineers with decades of experience waded in to note that the 1979 paper’s authors were actually pushing mutation testing (a test-for-the-test-suite idea) and that the real obstacle was always bad notation and not enough compute—both problems now solvable with SAT solvers, LLMs, and cheap hardware. The thread split cleanly between those who see verification as a practical extension of testing (proving specific properties like memory safety or API idempotence) and those who think the whole enterprise is doomed because human-generated specs will always be wrong or outdated, with one commenter arguing that if ordinary warranty rules applied to software, companies would already be out of business.
MathCode, Mathematical Coding Agent [comments]
86 points · 28 comments · math-ai-org.github.io · 13h ago
MathCode is a new terminal-based coding assistant that takes plain-English math problems, converts them into Lean 4 theorems, and attempts to generate formal proofs using a persistent Lean REPL, reusable theorem libraries, and parallel proof strategies. The HN discussion quickly zeroed in on whether the tricky part is actually getting the natural-language-to-formal translation right—some users with Lean experience pushed back, arguing that if you design primitives well and use mathlib’s solid API, verifying the formalized statement isn't that hard. Others questioned what real use cases look like, with one person noting the example "prove the square of an even number is even" is just a pre-existing theorem, not a novel problem. A licensing concern came up because the project has no visible terms, though a few commenters argued AI-generated code can't be copyrighted anyway, making the point moot for most practical purposes. The thread also saw a playful exchange over whether hooking "slop to slop" can produce anything valuable, with one user jokingly offering a modal-logic proof that it can't.
AI Coding Without the Vibes [comments]
82 points · 50 comments · peterbloem.nl · 21h ago
The post argues that instead of either banning AI outright or fully surrendering to "vibe-coding," students and professionals should adopt "craft coding"—writing code yourself and using AI only as a reviewer to catch bugs and suggest improvements. The thread split into two main camps: one group wholeheartedly agreed, arguing that using AI as a reviewer preserves understanding and prevents the deskilling seen in junior developers who never learn to debug. A counter-camp pushed back, insisting the article underestimates AI's real power as an *explanation* tool—using it to ask focused questions ("what does this bit do?" or "is there a data leak here?") to deepen understanding, not just as a proofreader. A recurring analogy war erupted over calculators and slide rules: some argued calculators made people worse at mental math and the same will happen with coding, while others countered that teaching augmented skills is more relevant to real-world work. The most pointed criticism came from people who said trusting an LLM to review code you wrote is naive, because the same model that is "good at focused questions" is also terrible at catching its own blind spots, and that treating code review as a service an LLM can perform just kicks the can down the road on genuine comprehension.
GIMP Development Update [comments]
78 points · 24 comments · www.gimp.org · 4h ago
The GIMP team posted a detailed update on the upcoming 3.4 release, highlighting a new zipped-XML project file format to replace the decades-old XCF, non-destructive editing improvements, better PSD import support, and native file chooser dialogs. The big technical debate on HN immediately zeroed in on the zipped-XML format choice, with several commenters arguing that SQLite would be a far better modern alternative—it handles in-place modification, crash recovery, and data integrity in ways zip files simply don't. A few people pushed back, noting that zipped XML has mature tooling and compresses well, but the SQLite proponents had the stronger technical case. There was also a brief aside about the long-dead Glimpse fork, and a pedantic detour over whether "XCF" takes "a" or "an" based on pronunciation. The overall takeaway: the GIMP team is making real progress on long-requested features, but the format decision rubbed some deep expertise the wrong way.
GPS and the Lost Art of Getting Lost [comments]
77 points · 64 comments · www.newyorker.com · 17h ago
The New Yorker piece examines how GPS reliance is reshaping our relationship with physical space, using anecdotes about Israeli GPS spoofing, studies on spatial memory decline, and a journalist’s attempt at orienteering to argue we’re losing the ability to get lost. The thread pushed back hard on the premise that this is a simple loss—several people pointed out that pre-GPS navigation wasn’t some golden age of skill but often meant driving with a map on your lap in unfamiliar cities, which was dangerous, or relying on landmark-based strategies that fail in dense urban grids like Boston. One commenter directly accused the article of starting with a nostalgic premise and cherry-picking evidence, noting the New Yorker wouldn’t run this kind of speculative innuendo on something like vaccine hesitancy. Others shared concrete horror stories—a bus driver blindly following a GPS into a neighborhood with no turnaround, a backpacker whose phone showed a trail that had been obliterated by fire—that undercut the idea that the problem is just about lost art, framing it instead as brittle infrastructure and the need for fallback systems. There was also a split between people who actively cultivate getting lost as a practice (taking random buses, walking toward landmarks without a phone) and those who argued that “getting lost” has always been mostly annoying and is better replaced by autonomy.
US oil reserves are so low, the caverns holding them could be damaged [comments]
75 points · 62 comments · www.independent.co.uk · 19h ago
The article warns that America's Strategic Petroleum Reserve has dropped below 300 million barrels for the first time since the 1980s, and experts say the repeated rapid drawdowns to fund the war with Iran are permanently destabilizing the underground salt caverns that hold the crude, risking structural collapse and loss of pumping capability. The thread immediately split over whether the problem is physics or politics: a bunch of people argued the reserve was designed for five major emergencies over its lifetime, not dozens of shallow political cycles to soothe gas prices in election years, while others pointed out that "dozens of cycles" is a misleading framing since only a few were full drawdowns—the real issue is the rate and pattern of partial extraction and reinjection wearing out the cavern walls. A persistent counterpoint pushed back by linking to the actual reserve history graph, noting the largest drawdowns happened from 2020 onward under both parties, and that a cheap opportunity to refill at $24/barrel during COVID was blocked by Democrats who called it a "bailout for big oil." The Energy Department's statement that "the caverns are always full, just the ratio of oil to water changes" was treated with open contempt—readers translated that as an admission they're running the infrastructure into the ground with no plan for the century-long rebuild. A handful of comments dismissed the whole panic as irrelevant because the US can't refill the reserve anyway, since global tanker supply lines from the region have been severed by the war and there's no domestic refining capacity to turn local oil into usable fuel.
Generated 2026-08-17 08:10 UTC
Generated by Sauron from Hacker News discussions and linked articles.