HN Brief: 2026-10-13
Today's Hacker News was pulled between two opposing forces: a running philosophical fight over whether AI is a tool or a delegation, sparked by Rich Hickey's essay and feeding into threads on legal liability, open-source governance, and coding productivity. Alongside that, a deep appreciation for low-level craft—a commercial CGI web app, a WiFi router turned SDR, a Firefox download fix—ran straight into a thick streak of skepticism targeting overhyped claims, from OpenAI's galactic algorithm to the institutional rot at Deloitte. The overall mood was intellectually contrarian, allergic to hype, and surprisingly reverent toward the kind of engineering you can fully hold in your head.
Threads most worth your time include "'AI' Is Not a Tool by Rich Hickey", where the man himself showed up to defend his definition, distilling the whole day's tension, and "A Modern Web App in C", which turns a shipping blog post into a masterclass on tradeoffs splitting the room between minimalism and self-harm. "Why aren't AI companies being charged for hacking?" poses a provocative question, but the real sting is its bleak consensus that power dynamics render the law toothless. For the open-source governance fight of the day, "QGIS enacts moratorium on AI contributions" delivers a perfect proxy war over unenforceability and a controversial loophole. Finally, "Making big downloads resumable in Firefox" manages the neat feat of making a single about:config tweak feel like a full indictment of modern browser incompetence.
Hackers activate automatic defrosting in supermarket fridges across Europe [comments]
699 points · 818 comments · www.tomshardware.com · 21h ago
The article covers a coordinated attack on European supermarkets that switched Danfoss freezer controls to continuous defrost at 50°C, spoiling millions in stock while also taking down delivery systems and digital shelf labels. HN immediately framed this as total deja vu, since the exact same Danfoss exploit hit US military commissaries last month—something the submission itself linked to—and the thread harshly rejected the "unprecedented" label. The debate centered on why critical cooling infrastructure is internet-connected at all, splitting neatly between those who see remote monitoring as a genuine operational necessity and those who view it as an indefensible liability for equipment that gets physically checked daily anyway. For the OT security crowd, the vulnerability traced straight to Danfoss controllers running unsecured Modbus over TCP, making commercial fridges a trivial target to find on Shodan and turning the whole incident into a textbook case of the gap between HVAC contractors and network security. The consensus settled on a grim inevitability: nothing changes until liability lands on the equipment vendors themselves, not the razor-margin grocery chains, with the EU Cyber Resilience Act cited as the only real structural fix on the horizon.
Muse Wonder [comments]
635 points · 338 comments · research.meta.ai · 17h ago
Meta announced Muse Wonder, a new model claiming leadership in personal agentic tasks as a step toward their "personal superintelligence" vision. The thread split over whether this validates Zuck's execution gamble or is just more benchmark gaming, a fight the article's own dedicated "benchmaxxing" section didn't settle. The business argument that this proves frontier models are a pure commodity anyone with capital can replicate ran straight into the counterpoint that Meta still isn't beating the top labs on the hardest evaluations. Meanwhile, the "superintelligence" branding drew the expected ire, though the crowd corrected the record that Meta was using the term long before it became a political buzzword. The underlying grumble from a lot of the discussion was that the "personal agent" market is still a concept looking for a real consumer problem, regardless of how strong the underlying model looks.
Deloitte falsified documents to evade tax on Abramovich's wealth [comments]
603 points · 198 comments · www.bbc.com · 23h ago
The BBC's Panorama investigation found that Deloitte's Cyprus arm falsified documents to help Roman Abramovich hide assets and evade tax. The conversation largely skipped the oligarch spectacle to dig into how the Big 4's franchise model makes this kind of local fraud an expected feature, not a bug. There was a widespread consensus that Deloitte's ethics training exists purely as a liability shield, and its structure as a network of independent partnerships lets the global brand easily disown local misconduct. The only real split was on the UK's role, with most people arguing that London's real estate boom was fueled by a straightforward tax evasion mechanism that distorted the market. The discussion ultimately treated the story as a relentless case study of how structurally embedded white-collar criminality is in professional services.
Codex CLI installs a persistent, self-updating background daemon by default [comments]
565 points · 274 comments · github.com · 15h ago
The submission flags a GitHub issue asking for Codex CLI to make its background daemon opt-in rather than the default. The defense that browsers auto-update this way was rejected by pointing out the difference between a managed package ecosystem and an opaque daemon that hoards old versions, reportedly eating up gigabytes of disk space. This specific complaint escalated into a wider frustration with the coding agent ecosystem, where bloat like this is seen as a symptom of code generated without human oversight of the full system. The discussion branched into a hunt for alternatives, with minimal harnesses like hax and Pi getting discussed as saner drop-in replacements, and the consensus was that a professional tool acting as a rogue background updater signals a deep disrespect for user sovereignty.
A Modern Web App in C [comments]
564 points · 238 comments · prophitt.me · 17h ago
A developer shipped a real commercial web app in 2026 on a stack of C, CGI, and SQLite — writing raw HTML to stdout, spawning a process per request, and deploying by replacing a single binary. HN split hard on whether the language choice was inspired minimalism or willful self-harm, with one side cheering the tiny attack surface and total dependency control, and the other pointing to Rust or C++ as strictly superior options that deliver the same resource profile with far stronger safety. The thread dug deep into the production realities of straight CGI in 2026 — nginx doesn't ship native support, concurrent SQLite writes from multiple forked processes is a genuine constraint, and process forking overhead starts cutting into performance well before you reach big tech scale. A persistent tangent debated whether the blog post itself was generated by an LLM, pegged on its repetitive cadence, though the author's detailed engagement on the tradeoffs and specific pain points mostly undercut that theory. The thread landed on immense respect for the shipping and the craft, combined with a thorough, practical dissection of why this is a fascinating hard mode rather than a general blueprint.
GLM-Legal [comments]
562 points · 316 comments · z.ai · 17h ago
Z.ai released GLM-Legal, a compact LLM fine-tuned specifically for complex legal workflows like contract review and compliance research, claiming it punches above its weight on major benchmarks despite a smaller footprint. The thread immediately zeroed in on the liability trap—commenters argued that while lawyers are already responsible for everything filed under their name, LLMs confidently hallucinating case law creates a distinct "vibe-practicing law" crisis that demands mandatory client disclosure rather than just treating the AI like a fancy paralegal. A heavy undercurrent of skepticism hit the model's closed-source status, with a strong consensus that any model for high-stakes legal work is dead on arrival if it can't be self-hosted and audited by the firms using it. Commenters split hard on the profession's future: some predicted a gilded life of cushy sign-off gigs, while others warned of a race to the bottom where firms pressure lawyers to stamp mountains of unchecked AI output.
AI is a tool and using it well takes skill [comments]
416 points · 335 comments · yyhh.org · 12h ago
David, this post pushes back on Rich Hickey's claim that LLMs aren't tools, arguing they're tools that require skill to wield well, just like a saw or chisel. HN largely agrees on the surface but the discussion immediately splits on whether that skill is durable—experienced users report that their prompting workflows are already obsolete, with the meta collapsing to just "knowing what to ask for," which puts the skill gap on the same footing as Google Search literacy that most people never mastered. The real tension is between those who think the tool's limits create a lasting need for expertise and those who see this as a temporary phase before AI automates away the need to articulate what you want at all. A side thread also argues about whether automation itself can be a tool, citing Factorio as a game about the skill of designing automated systems, which circles back to whether the whole definitional debate is just marketing.
QGIS enacts moratorium on AI contributions [comments]
375 points · 402 comments · github.com · 22h ago
QGIS replaced its previous human-in-the-loop AI policy with a total ban on any LLM- or agent-assisted code contributions, threatening code reversion and contributor sanctions for violations. The thread was fiercely split—supporters argued the ban was a necessary response to AI-generated slop overwhelming maintainers and pointed to at least one high-value contributor already lost under the old policy, while critics called the ban unenforceable and a defensive overreaction to progress. A vocal contingent insisted that AI could clone the entire QGIS codebase in hours, a claim that defenders met by noting that a fork without the project's community trust and decades of domain expertise was worthless. The discussion also latched onto a controversial exception allowing "trusted contributors" to use AI, which looked like a two-tier system, and the unresolved definition of what counts as "AI assistance"—whether prototyping, autocomplete, or even detailed issue prompts trigger the ban.
Turn a WiFi 7 Router into a 320Msps Software Defined Radio Receiver [comments]
349 points · 91 comments · github.com · 21h ago
The project turns a $80 WiFi 7 router into a 320 Msps SDR by reverse-engineering the radio chip and streaming raw IQ over 10GbE—a capability that previously cost tens of thousands. The HN thread is split between awe at the price-to-performance ratio and skepticism about the method, since the firmware was mostly generated by Claude and isn't deeply understood by the author. Critics call the documentation word salad and worry it’s a script-kiddie-style demo that won’t lead to durable tools, while others point out that the same level of reverse-engineering a decade ago gave us the RTL-SDR and that this genuinely opens up the higher bands for hobbyist SDR work. The debate also wanders into whether AI-assisted reverse engineering is “real” engineering or just assisted exploitation of undocumented hardware.
Citroen revives the 2CV for an electric age [comments]
340 points · 379 comments · www.bbc.com · 21h ago
Citroen is reviving the 2CV as a cheap electric city runabout, leaning on retro nostalgia to compete with cheaper Chinese EVs. The thread mostly rejected that heritage-differentiator, warning the VW Beetle revival proves this fades fast and that working-class car buyers just need a cheap tool. It quickly snowballed into an indictment of all European legacy automakers, especially Stellantis and VW, for making lazy, overpriced EVs on retrofitted platforms, with the VW ID.3 getting eviscerated for terrible software and cheap build. The biggest split was between those just wanting a dead-simple, truly basic EV and those insisting modern safety regulations make that impossible. As a relief from the negativity, the thread went deep on the Toyota Prius drivetrain as the gold standard of elegant, pragmatic engineering—the real kind of cleverness people wanted to see.
AI coding agents generate more code, but not more software [comments]
335 points · 285 comments · arstechnica.com · 16h ago
The article reports that a study found AI coding agents boost lines of code and commits by 20-30% but don’t increase software output because code review becomes a bottleneck. The discussion argued that while AI catches bugs that humans miss, it makes review an even bigger chore, with a split over whether review is essential for shared understanding or just tedious. A strong tangent pointed out that AI-generated tests are often useless—overly mocked and white-boxed—and there were fears that management might use AI to prompt and ship without understanding. The consensus was that AI shifts the bottleneck in development rather than removing it, and any gains in coding speed are eaten by downstream review.
Why aren't AI companies being charged for hacking? [comments]
322 points · 348 comments · mortoray.com · 19h ago
The linked piece argues that AI companies should face criminal hacking charges because they designed systems specifically for security probing and deployed them in environments where hacking was entirely foreseeable, thereby establishing legal intent. The thread immediately splits on that "foreseeability equals intent" theory—a large camp calls it a massive stretch that would logically apply to any powerful tool, while the other side insists the direct actions of the specific agents put this well beyond a simple tool analogy. A major tangent lands on the glaring double standard between how journalists have been aggressively prosecuted under the CFAA for technical boundary-testing and the total inaction against companies doing it at industrial scale. The most grounded take running through the thread is that criminal charges are less a legal question and more a question of power and incentive—the CEOs have political connections, the victims would rather settle than fight, and without tangible damage, no prosecutor is eager to be the first to test this novel theory in court. The general frustration is that the law seems to clearly apply on paper, but the power dynamics and absence of immediate harm are leaving the AI companies in a comfortable legal limbo.
Thinking will become a hobby [comments]
291 points · 407 comments · www.spinellis.gr · 17h ago
The article by Diomidis Spinellis argues that AI’s superior performance and cost will inevitably relegate human cognition to a hobby, much like machinery did to physical labor. The debate was immediately divided over whether LLMs actually *think* or are just sophisticated autocomplete, making the parallel to hunting or calligraphy a category error, while others countered that the mechanism doesn't matter if the output matches or exceeds human work. The fiercest dispute skipped the philosophy of mind to target the political economy, where the suggestion to simply "adapt and find purpose" was torn apart as dangerously naive about a future where the owning class controls the thinking machines and everyone else fights for relevance and survival. Tangents on sigmoid progress curves, collapsing literacy rates, and the failures of the enclosure movement all served to frame the transition as a terrifyingly undemocratic crisis rather than a smooth shift to a leisure society.
'AI' Is Not a Tool by Rich Hickey [comments]
276 points · 142 comments · www.linkedin.com · 16h ago
Rich Hickey’s latest piece forcefully argues that LLMs aren’t tools you wield but automation you delegate to, a distinction he insists matters for learning, skill, and control. Hacker News spent most of its energy trying to break his definition, asking whether a debugger like GDB stops being a tool, whether a car counts as automation, and whether the whole semantic exercise distracts from the real issue of how LLMs actually function. A sharp divide opened up between the semantic purists who backed Hickey's dictionary-first approach and the pragmatists who argued that language bends to usage and any tool you don't fully control is still a tool if it amplifies your work. Hickey surfaced in the discussion to concede that toolness is a spectrum—cars drift toward automation—but held firm that the vague request-and-pray cycle of generative AI is a fundamentally different relationship than wielding a tool, a line that only escalated the argument about where exactly the boundary lies. By the end, no one agreed on where the boundary is, but the thread crystallized a much harder question: whether treating LLMs as a tool at all reinforces the very abdication Hickey warned about by making delegation feel like craftsmanship.
OpenFatigue™ – Open is the new closed [comments]
221 points · 50 comments · openfatigue.ai · 14h ago
The submission is a pitch-perfect satirical website, OpenFatigue, that skewers the tech industry’s compulsive branding of proprietary systems as "Open" by creating a fake product where getting "open" model weights requires faxing pull requests and signing a 41-page acceptable use policy. A meta-argument erupted over whether the site itself was generated by an LLM—critics found the humor generic and machine-flavored, while defenders saw it as an intentionally clever performance piece stretching across the whole page. The substantive split was over open weights: one camp argued that releasing weights under a permissive license represents a legitimate middle ground on the openness spectrum, but the old guard held firm that "open source" is a strictly defined legal term and anything less is just marketing that dilutes real freedom. Someone dug up historical OpenFatigue from the 80s and 90s with things like OpenLook and The Open Group, proving this exhaustion is an inherited condition that predates the current AI hype cycle. Everyone agreed on the symptoms even if they couldn't agree on the diagnosis, making the whole thing feel like a cathartic support group for people sick of the word "Open" being stripped of all meaning.
AI models can provably hide arbitrary capabilities [comments]
210 points · 106 comments · arxiv.org · 7h ago
The paper demonstrates a technique to encrypt a complete neural network inside a host model's weights, activating only on a secret trigger and remaining provably impossible to extract via standard safety checks. The thread fractured over whether this kills open-weight models or just highlights the same trust problem with closed APIs, where providers can spy on you without needing any cryptographic tricks in the weights. A strong pushback framed it as the AI version of encrypted/polymorphic malware or the xz backdoor, though the provable guarantee of non-elicitation was widely seen as a genuine theoretical headache for capability audits. The technical debate over detectable weight distributions got deep quickly, but the consensus that emerged favored full supply chain transparency—dataset, training code, checkpoints—over trusting the final parameter dump alone.
Intel D270 GPU AI Accelerator–Up to 480 GB LPDDR5X [comments]
207 points · 107 comments · www.intel.com · 12h ago
Intel pitched the D270 as a 480GB LPDDR5X data center GPU explicitly tuned for agentic AI's long-context inference, running 350W air-cooled on standard PCIe to skip the expensive HBM and liquid cooling requirements. The crowd immediately split between mourning the final death of Intel's consumer GPU line and eagerly pricing up the used market haul five to eight years from now, when data centers decommission these things. Defenders of the strategy pushed back hard, arguing that scarce wafer capacity and fat enterprise margins make the data center pivot the only financially sane path for a company in Intel's present straits. The 1.5 TB/s memory bandwidth sparked the heaviest technical debate, framed as either a crippling bottleneck compared to HBM from rivals, or the entire point of a power-efficient appliance designed for dense racks and MoE-heavy workloads. Without a price or ship date, the thread essentially argued whether a tightly wound efficiency wedge can crack Nvidia's dominance, but almost everyone agreed the specs look like a direct fulfillment of the local model runner's wish list, even if they aren't the target buyer.
Making big downloads resumable in Firefox [comments]
206 points · 90 comments · lunarwatcher.codeberg.page · 14h ago
The linked article digs into a Firefox bug where an obscure `network.http.response.timeout` setting improves the browser's famously broken support for resuming large downloads, though the author admits they have no clue *why* it works. The HN crowd took this as a rallying cry against how laughably bad browsers remain at a fundamental download task that dedicated utilities mastered on dial-up twenty years ago, flooding the thread with nostalgia for GetRight, FlashGet, and Go!Zilla alongside modern praise for handing URLs off to `aria2`. A sharp split emerged between those who blame Mozilla for neglecting a core browser competency and a smaller contingent who argue the real problem is modern web servers using single-use, expired URLs that break the HTTP Range mechanics required for clean resumption. A deeper technical tangent warned that splicing interrupted `.part` files onto new sessions without verifying ETags is a recipe for silent data corruption, even if the browser displays a successful resume. The overwhelming sentiment is that in 2024, no one should need to dig through `about:config` or deploy an external download manager just to safely download a big file.
Who is vfmaddrnd231pd? [comments]
203 points · 20 comments · purplesyringa.moe · 12h ago
The blog post investigates a phantom FMA instruction—vfmaddrnd231pd—that Intel documented but never shipped, so real hardware raises an undefined instruction error when you try to run it. The thread confirmed it was a fleeting ghost in the SDM, present only in a couple of revisions and yanked before any CPU made it out the door, likely because Intel’s per-instruction rounding control had a subtle hardware bug. People drilled into the cryptic name and the odd "DDS" encoding, pointing out that Intel’s vector instruction mnemonics are designed for machine parsing, not human readability, so the name's very inscrutability is by design. This veered into a philosophical argument over whether assembly languages have lost their way by abandoning human-friendly syntax, and whether AI-driven assemblers make that loss moot or more urgent.
RJ Mitte has died [comments]
199 points · 60 comments · news.sky.com · 11h ago
The linked article wasn't available to this summarizer; from the discussion, it's about the death of *Breaking Bad* actor RJ Mitte at age 34. The conversation dug into the root cause, a suicide driven by financial devastation after his manager stole his earnings and faked his tax documents, leaving him with monstrous IRS debt. This sparked a sharp split: one camp detailing how routine this Hollywood financial predation is, and another unloading raw personal experiences with suicidal ideation and the specific hopelessness of being completely broke. The strongest pushback grounded the discussion in the reality of his Cerebral Palsy—a brain condition that primes someone for depression as much as any life circumstance—which kept the thread from collapsing into simple blame or easy advice.
Solar panels that plug into your wall [comments]
189 points · 192 comments · www.nationalgeographic.com · 19h ago
The National Geographic piece frames plug-in balcony solar as a clean, easy revolution spreading across US states. The thread jettisoned the article's green framing for a gritty economic fight, treating the panels as a tactical weapon against utility monopolies—one comment's sprawling takedown of PG&E's price hikes became the thread's defining beat. A sharp technical faction popped the article's resilience bubble, noting these panels require the grid to function and need expensive batteries to actually work in a blackout. The defining split pitted distributed-energy optimists, who see this as a viable weapon against the utility death spiral, against utility-scale realists, who argue it's a drop in the bucket compared to building out actual solar farms.
Kevin Kelly's 2016 AI Predictions, 10 Years Later [comments]
187 points · 72 comments · taylor.town · 18h ago
The article revisits Kevin Kelly’s 2016 predictions from *The Inevitable* — that Google’s main product would shift from search to AI, that a narrow-AI oligarchy would rule, and that AI would become an invisible cloud utility — and sees how they aged. The thread mostly rejected the generous framing of Google’s trajectory, arguing the company is fundamentally an ads business and everything AI-related is still just a feature serving that monopoly, not a product pivot. A sharp split emerged over whether Kelly’s hit rate is actually impressive for futurology or just survivorship bias, which pulled the conversation into a broader thrashing of Ray Kurzweil and whether his own misses should be excused by society simply refusing to wear his chosen hardware form factors. The prediction that AI would obey a user-driven network effect became another contested point, with detractors insisting the best models mostly train on static scraped data rather than live user interaction, while defenders argued those dynamics are simply still in their early phase.
Feds got warrants to use fake cell towers to hunt down Don Lemon's producer [comments]
174 points · 103 comments · www.wired.com · 16h ago
The Wired piece covers federal agents getting warrants to use Stingray cell-site simulators to hunt down Don Lemon's producer after a protest at a church pastored by the head of the local ICE office. The room was split straight down the middle — outrage at the government's willingness to lie in affidavits to target journalists and protesters, demanding consequences for the false claims that appeared in fifteen warrants, stood directly against the view that the producer broke the law by disrupting a church service and that federal civil rights law doesn't offer a press exemption. A parallel technical debate questioned the value of countermeasures like airplane mode, disabling 2G, or the EFF's Rayhunter detection project, with networking experts pointing out that SS7 and Diameter protocol attacks don't require an intrusive local box at all, making the phone a fundamentally unprotectable tracking beacon for a determined federal agency. The overall consensus was bleak: the executive branch is weaponizing surveillance against dissent, and between the malleable facts in warrant applications and the insecurity of the cellular network, there is very little practical defense for the target of a determined investigation.
Just say no to OpenAI and Anthropic [comments]
159 points · 155 comments · micahflee.com · 15h ago
The linked post makes the case for boycotting OpenAI and Anthropic, arguing open-weight models like GLM-5.3 have caught up while avoiding the privacy and governance issues of the closed labs. The technical claim immediately took a beating, with many arguing the author's vibes-based parity assessment is a flat misreading of the benchmarks. The privacy floor of the argument also got shaky—the debate over whether commercial Zero Data Retention APIs and TEEs actually offer privacy, or just shift the surveillance to abuse monitoring, largely stalled on the conclusion that trust is an unsolved problem regardless. The broader political drift questioned whether this is just the same grift cycle as crypto and social media, and whether swapping US oligarchs for the Chinese open-weight labs the author endorses really improves sovereignty.
Blood cells from 80-year-old donor become stem cells with molecular age under 20 [comments]
159 points · 106 comments · onlinelibrary.wiley.com · 10h ago
The linked article wasn't available to this summarizer; from the discussion, it describes a method that converts blood cells from an 80-year-old donor into stem cells with a molecular age under 20. The conversation immediately abandoned the biology to wage a proxy war over the political and existential implications of radical longevity, split between people who see death as a necessary evolutionary feature and those who view it as a solvable problem. The "death is a feature" side argued that aging eventually curbs the power of dictators, a stance that drew sharp pushback from others who described it as a staggeringly passive approach to governance. The pro-longevity camp spent much of the thread on the defensive, wrestling with the *Altered Carbon* vision of a permanent immortal overclass while fielding warnings about "turbo cancer" from the Yamanaka factors—a risk some argued the thread was vastly oversimplifying. If there was a consensus, it was that the biology is hurtling forward much faster than the social contract is prepared to handle.
Incident with Actions, Pull Requests and Issues [comments]
156 points · 142 comments · www.githubstatus.com · 21h ago
The link is GitHub's own status page for the latest outage hitting Actions, Issues, and Pull Requests. The thread immediately fractured over root cause: one side pointed to aging infrastructure and the messy Azure migration, but many others pushed back forcefully with GitHub's cited numbers showing a 30x traffic surge from AI agents hammering the free tier. This prompted a wave of people sharing exit plans from GitHub Actions, which spun off into a debate on whether that paying-customer exodus creates a death spiral for the platform. The loudest counterpoint argued the real failure is economics—cap the free-tier abuse instead of blaming the engineering, and the service for paying users would mostly be fine.
Leaving Tech [comments]
152 points · 160 comments · ryanbigg.com · 7h ago
Ryan Bigg, a prominent Ruby maintainer, is leaving tech after applying to 36 senior remote jobs, getting mostly rejection and silence, and concluding the market only wants AI hype over actual experience. The thread immediately fractured over whether 36 applications is a serious try or an impetuous walkout, with one side arguing a 20-year veteran shouldn't have to play the numbers game and the other countering that this is just the reality of the current market. A huge chunk of the discussion wasn't about his quitting at all but about collective culpability, where a vocal minority argued that developers who embraced and evangelized LLMs have no right to complain now, which drew sharp pushback from people who rejected being lumped in with AI boosters. Despite the factional fighting, no one really refuted his specific claims about the collapse of the remote senior generalist role and the mental health toll of a hiring process that treats deep expertise as a liability.
I turned my unused Kindle into a photo frame with GPT-6 Astra [comments]
143 points · 97 comments · nirajrajgor.com · 10h ago
The author got GPT-6 Astra to jailbreak an unused Kindle and turn it into a photo frame, after GPT-5.6 Sol couldn’t finish the job and Anthropic’s Opus and Fable models refused to cooperate after two prompts. The thread seized on exactly that model behavior, with many people recounting how OpenAI’s code models are happy to help with low-level tinkering while Anthropic’s safety blocks shut down legitimate system work—and some arguing that jailbreaking is entirely legal under federal exemptions, so a model refusing is just coddling. A separate fight broke out over whether sharing AI-written code has any value: one camp insists only the prompts matter in a world where everyone has access, but others counter that a working, debugged product saves everyone the energy of re-iterating forty hours of turns. A few also raised the longer-term worry that if companies use the same AI capability to harden devices, future hardware will be nearly impossible to jailbreak, locking down tinkering for good.
Janas: Run LLMs larger than RAM by streaming experts from SSD [comments]
141 points · 46 comments · github.com · 11h ago
Janas is a new engine that tackles the “model bigger than RAM” problem by aggressively streaming MoE experts off an SSD, promising massive speedups over llama.cpp on the exact same hardware—like running an 80B model at usable speeds on a 32GB laptop. The discussion immediately split into two camps: the impressed, who see this as a game changer for local inference, and the skeptical, who compared it to “swap” and pointed out the extreme hardware dependency (requiring fast NVMe, specific Intel Arc drivers, and Debian 13 to hit those benchmark numbers). A lot of the pushback zeroed in on the difference between token generation speed and prefill speed, with several people arguing that 60 tok/s is fine for chat but completely ruins agentic workflows or reading large contexts, though others countered that it’s well past the threshold of feeling “annoyingly slow” for interactivity. The thread also spent time litigating the originality of the approach—some argued it was just a reimplementation of existing ideas from llama.cpp or projects like Strata, while others defended it as a purpose-built tool with a genuinely different caching strategy that bridges the gap to the next level of hardware innovation.
A look at the math that OpenAI claims computes faster Fourier transforms [comments]
136 points · 98 comments · www.argmin.net · 7h ago
The article is a thorough takedown of OpenAI's "faster" Fourier transform, dissecting how the claimed speedup relies on exact complex arithmetic and impossibly large constants, making it a textbook galactic algorithm. A sharp split emerged between people who saw this as a necessary corrective to a field detached from reality, and theorists who defended the result as a valid mathematical landmark that doesn't need to be practical to be interesting. The defense leaned heavily on the fact that big-O notation inherently ignores constants and that seemingly useless results (like Fermat's work) often find practical homes decades later, while the criticism hammered the model of computation—if an algorithm needs arbitrary-precision complex numbers in constant time, it's playing a completely different game. A strong counterpoint argued the result might actually hold in prime-order finite fields used heavily in cryptography, complicating the article's blanket dismissal. The discussion ultimately couldn't resolve whether this kind of work is a vital stepping stone or a symptom of a cultural incentive problem where formalism is mistaken for progress.
Generated 2026-10-13 10:08 UTC
Generated by Sauron from Hacker News discussions and linked articles.