HN Digest 2026-09-21 · Xiaomi's MiMo v2.6 sets a new bar for open-model transparency

Hacker News Summarized - Daily AI Papers | Zeli，Every story on the Hacker News front page, summarized as it lands, plus the top comments and HuggingFace's daily AI papers. Pick what's worth reading.关注
  1. Xiaomi's MiMo v2.6 sets a new bar for open-model transparency — ▲ 1091 · 💬 470 · 🔗 mimo.xiaomi.com
    Xiaomi's MiMo v2.6, a 9B agentic model built by fine-tuning Qwen3.5-9B on MiMo-generated data, has sparked praise for its unusually transparent release. The company shared a real-time training dashboard, a detailed tech report, and benchmark scores including failures. Commenters highlight its cost-effectiveness and strong performance on niche tasks, with some calling it a Pareto frontier model for game-playing under $0.15 per million input tokens.
  2. Letting an Algorithm Choose Your Screen Is Like Giving It the Key to Your Brain — ▲ 1034 · 💬 315 · 🔗 alicegg.tech
    The Tetris effect shows that whatever you focus on long enough shapes your thoughts. Yet recommendation algorithms now decide what we watch, hear, and read, hijacking our attention for ads and engagement. The author argues we can reclaim an intentional internet of bookmarks, RSS feeds, and blogs, but it requires patience and deliberate habit.
  3. AI Is Great for Writing, So Why Is Everything It Writes Unreadable? — ▲ 987 · 💬 424 · 🔗 blog.colinbreck.com
    Colin Breck argues that AI-generated design docs, pull requests, and messages are unreadable because they lack the context and vulnerability of human writing. Drawing on survey data showing 78% of readers stop at AI-scented posts and insights from Bjarne Stroustrup and Simon Sarris, he explains why AI works for verification and abstracts but fails at authentic communication—and how tools like Pangram might enforce the social contract between writer and reader.
  4. Sun's Fatal Flaw: Boredom with Business — ▲ 674 · 💬 388 · 🔗 bcantrill.dtrace.org
    Bryan Cantrill reflects on Sun Microsystems' decline, arguing that the company's strategic brilliance was undone by a disinterest in operational execution. He recounts a 2005 incident where a growing startup running OpenSolaris tried to buy Sun hardware but couldn't get a sales call, while Dell's responsive rep made the startup feel like a big company. Sun's boredom with business mechanics, Cantrill concludes, sealed its fate.
  5. ZuckOff listens for Bluetooth signals to alert you when camera glasses are nearby — ▲ 605 · 💬 6 · 🔗 zuckoff.app
    ZuckOff is an independent app that detects camera glasses like Ray-Ban Meta, Oakley Meta, and Snap Spectacles by listening for their Bluetooth manufacturer signatures. It runs entirely on your phone with no account, logs every Bluetooth device, shows the evidence behind each flag, and lets you mark your own glasses to stop alerts. The app also offers background alerts, a Home Screen widget, and Shortcuts integration.
  6. Grok 4.7 matches frontier coding performance at a fraction of the price — ▲ 597 · 💬 516 · 🔗 x.ai
    xAI's Grok 4.7 is a new flagship model for coding and knowledge work, trained with longer reinforcement learning on harder, multi-hour tasks. It scores 46.3% on CursorBench 4.0 and 71.0% on DeepSWE v1.1, outperforming Grok 4.6 while costing $2 per million input tokens and $6 per million output tokens—far below GPT-5.6 Sol and Fable 5.1. It also ships with a new safeguard stack that blocks 96.7% of risky dual-use prompts.
  7. Attention Is All You Need: How Transformers Actually Work — ▲ 595 · 💬 87 · 🔗 poloclub.github.io
    The 2017 paper "Attention Is All You Need" introduced the Transformer, now the go-to architecture for GPT, Llama, and Gemini. This visual explainer walks through every piece of a text-generative Transformer using GPT-2 (124M parameters) as a concrete example: tokenization, embedding, positional encoding, multi-head self-attention, the MLP block, and the final softmax that predicts the next token. It also covers temperature, top-k, and top-p sampling, plus auxiliary features like layer normalization and residual connections.
  8. Disney+ quietly puts ads in its 'ad-free' tier — ▲ 505 · 💬 365 · 🔗 consumerrights.wiki
    In January 2025, Disney+ updated its Subscriber Agreement to allow advertising across all subscription tiers, including the premium tier marketed as 'no ads' or 'ad-free.' The change, buried in Section 2(k), binds existing subscribers unless they cancel. In September 2026, Disney emailed German subscribers a 'clarification' stating that ads could now appear before and after content even on Standard and Premium plans.
  9. Kev: Tiny decision models you can train and run on your own — ▲ 452 · 💬 199 · 🔗 github.com
    Kev is a family of small decision models (0.8B, 4B, 9B) built on Qwen3.5 that answer yes/no, multiple-choice, and rating questions in a single request. They run on CUDA and Apple Silicon, and you can train your own using the provided code and data. The API matches TypeSafe's System One, so you can point their Python SDK at your local server. A web playground lets you test inputs and check how option order affects answers.
  10. NASA's Mars Sample Return Mission Is Dead After Congress Backs Cancellation — ▲ 445 · 💬 375 · 🔗 science.org
    Congress has effectively killed NASA's Mars Sample Return program, backing the White House's effort to cancel the mission that would have brought Perseverance rover's rock cores to Earth. The compromise spending bill shifts $110 million to a 'Mars Future Missions' program, leaving dozens of samples—including one with a potential biosignature—stranded on Mars. The decision also affects ESA's Earth Return Orbiter and raises questions about U.S. leadership as China advances its own Mars sample-return plans.
  11. Model Reasoning Declined for Six Weeks While Users Blamed 'Nerfing' — ▲ 417 · 💬 290 · 🔗 twitter.com
    Lon Lundgren tracked a model's reasoning quality over six weeks and found it wasn't a one-time drop: performance fell across the whole period and fluctuated in multi-day episodes, some lining up with product announcements. Despite using xhigh or max effort, most invocations got little to no thinking tokens, and longer runs rarely hit published benchmark levels. The real question isn't whether the model was nerfed, but what inference regime you were served.
  12. ZuckOff Is a Free App That Sees Meta Glasses Before They See You — ▲ 397 · 💬 347 · 🔗 wired.me
    ZuckOff, a free Bluetooth detection app built by Polish developer Pawel Szydlowski, alerts you when smart glasses like Ray-Ban Meta, Oakley Meta, or Snap Spectacles are nearby. It fingerprints their broadcast signals and reads signal strength for rough proximity, though it can't tell if they're recording or who's wearing them. Downloaded over 5,000 times on the App Store in its first month, it offers a partial fix as venues ban the glasses and Meta pushes tamper-detection updates.
  13. Tim Schafer's 1996 Grim Fandango puzzle doc hides a joke box for your tears of joy — ▲ 374 · 💬 98 · 🔗 gameshelf.jmac.org
    The original puzzle design spec for Grim Fandango, posted by Tim Schafer at Double Fine for the game's tenth anniversary, has resurfaced on Hacker News. Commenters trade memories of the LucasArts classic — its art, music, and writing still unmatched decades later — and note that Schafer filled an internal document with jokes, graphics, and asides, a level of craft they doubt any 2026 studio would tolerate.
  14. Apple shows you how to turn off Apple Intelligence on Mac — ▲ 341 · 💬 219 · 🔗 support.apple.com
    Apple's support guide details how to disable Apple Intelligence features on macOS 27, including Siri AI (Beta), message and notification summaries, Smart Replies, and voicemail suggestions. It also explains switching to Siri Classic and using Screen Time to restrict access for children. The guide notes that Siri AI is in beta, not available in all regions, and that some server-based features have daily usage limits.
  15. Linear reworks CI to keep pace with AI coding, cutting PR wait and runner time — ▲ 308 · 💬 381 · 🔗 linear.app
    AI coding agents sped up code shipping, but CI couldn't keep up. Linear's CTO flagged high CI costs, so the team optimized. Despite test suites nearly quadrupling, they reduced pull request wait time from over 6 minutes to just over 5 and halved runner time per test. Key changes: faster runners, dropping TypeScript from linting, batching jobs, and smarter test sharding.
  16. Mini-AGI trains a continual learning model from scratch on an 8GB laptop GPU — ▲ 273 · 💬 73 · 🔗 github.com
    Mini-AGI is a byte-level language model that assembles its own architecture, trains on a single 8GB VRAM GPU, and keeps learning without catastrophic forgetting. It pages expert weights from disk, grows and prunes capacity dynamically, and uses a 0.1x trunk learning rate to retain 99.84% of progress against chance. It's a toy-level experiment showing personal continual learning is possible on modest hardware.
  17. Heretic strips restrictions from language models so they obey you — ▲ 266 · 💬 109 · 🔗 heretic-project.org
    Heretic is a free, open-source tool that removes restrictions from language models, ensuring they always follow your instructions. Install it with pip and run it on any model, such as Qwen/Qwen3.5-4B, to take control of AI behavior. Released under the AGPLv3, it offers a direct way to bypass built-in limitations.
  18. Cloudflare's Python Workers hit general availability, letting you run FastAPI and Django at the edge — ▲ 263 · 💬 41 · 🔗 blog.cloudflare.com
    Cloudflare's Python Workers are now generally available, making Python a first-class language on its developer platform. The runtime supports native bindings to Workers AI, R2, D1, and more, plus popular frameworks like FastAPI, Django, and Flask via ASGI/WSGI connectors. A new socket bridge enables PostgreSQL and MySQL drivers through Hyperdrive, and AI libraries such as LangChain and OpenAI now run natively. Cloudflare also proposed PEP 783 to standardize Python on WebAssembly.
  19. M5 Ultra Mac Studio: The Dream Machine for Local AI Agents — ▲ 261 · 💬 259 · 🔗 macstories.net
    After four days of testing the M5 Ultra Mac Studio with 256 GB RAM, I found it a game-changer for local AI agents. Compared to the M3 Ultra, prompt processing is 2.5× faster and token generation ~70% faster, enabling smooth multi-turn agentic loops. Running Qwen3.8-Flash-Next locally now powers my daily assistants, eliminating cloud costs and privacy concerns.
  20. Raspberry Pi locks down RAM upgrades in closed-source EEPROM — ▲ 248 · 💬 209 · 🔗 forums.raspberrypi.com
    A Hacker News discussion reacts to Raspberry Pi blocking users from swapping RAM chips on their boards. Commenters note the restriction is enforced through the closed-source rpi EEPROM, pointing to an open GitHub issue. Some argue the company is protecting itself from complaints about faulty third-party chips, while others call it a poor excuse and suspect a push to sell higher-margin, maxed-out devices.
  21. macOS 27 Users Are Hunting Workarounds to Dodge Apple's AI Model Downloads — ▲ 236 · 💬 120 · 🔗 reddit.com
    A Hacker News thread digs into a Reddit workaround for stopping macOS 27 from pulling down Apple's AI models and eating disk space. Commenters compare the hassle to de-crapifying Windows, and one points out Apple claims 2.5 billion Siri requests a day — though many are simple timers and reminders that never needed AI at all.
  22. Developer replaces padStart() with an AI model call to left-pad strings — ▲ 232 · 💬 87 · 🔗 github.com
    Fatih Kadir Akın's jev-leftpad is a JavaScript package that left-pads a value by asking TypeSafe AI's Jev model which option to pick, with criteria named space_0 through space_10. It needs a TYPESAFE_API_KEY, Node.js 20+, and one API request per call with retries disabled. The README admits padStart() could do it in one line and warns against using it in production — or anything important.
  23. One Cut Fiber Cable Grounded Flights at JFK, LaGuardia, and Boston — ▲ 231 · 💬 142 · 🔗 reuters.com
    The FAA halted flights at New York, Philadelphia, and Boston airports on Monday after a construction crew in New Jersey accidentally severed a fiber cable — and a backup line too. Ground stops hit JFK, LaGuardia, and Teterboro as 130 world leaders arrived for the UN General Assembly. More than 600 flights were delayed at Newark and Philadelphia, with Newark delays averaging 112 minutes.
  24. The One-Line Math Trick Behind Every 3D Camera — ▲ 206 · 💬 36 · 🔗 gabrieloc.com
    A personal journey from treating cameras as black boxes to understanding the simple math behind 3D projection. The core trick—dividing x and y by z—is shown to be a special case of the perspective projection matrix, which adds field of view, aspect ratio, and clipping planes. The full pipeline from world space to pixels is demystified.
  25. Shortwave Radio Gets a Secure Data Upgrade With HERMES — ▲ 180 · 💬 67 · 🔗 spectrum.ieee.org
    Rhizomatica's HERMES system turns shortwave radio into a digital data pipeline, sending encrypted files and GPS coordinates from the world's most remote locations without satellites or relay towers. In a pilot with fishers in Bangladesh, a boat stranded in the Bay of Bengal used HERMES to send an SOS and coordinate a rescue. Founder Peter Bloom explains how it works and why encryption is optional.
  26. Tim Dettmers: Academia Is About to Have a Renaissance, Not a Funeral — ▲ 177 · 💬 100 · 🔗 timdettmers.com
    Tim Dettmers argues that the future of AI research belongs to universities, not just big labs with the most GPUs. In his dlab Open Source Week, he releases an ecosystem of tools: an agent harness that autonomously optimizes CUDA kernels, runs 550B-parameter models on a single 24 GB GPU, and beats frontier deep research systems entirely offline. He also introduces CliffCompaction, an auto-compaction technique that cuts agent costs by 50% and enables sessions of over 100 million tokens.
  27. Uber Ordered to Pay $40 Million After Driver Left Passenger at Freeway Gore Point — ▲ 159 · 💬 187 · 🔗 consumerrights.wiki
    A California arbitrator found Uber vicariously liable as a common carrier for driver Vu Tran's negligence after he left 23-year-old Emily Normandin-Parker and her friend at an unsafe freeway gore point, where Normandin-Parker was struck and killed. The July 2026 award orders Uber and Tran to pay her parents $20 million each and her friend $300,000. The arbitrator rejected Uber's argument that Proposition 22 shields it from liability, and Uber says it was wrong to be held responsible.
  28. CBP Halts All Personal Prescription Imports on October 22 — ▲ 159 · 💬 142 · 🔗 personalimportation.org
    U.S. Customs and Border Protection will suspend all personal prescription importation starting October 22, according to the Campaign for Personal Prescription Importation. The advocacy group, which supports buying medications from licensed Canadian pharmacies with a 100% safety record, urges Americans to contact Congress and share their stories to protect access to affordable drugs.
  29. Apple's Cancelled Copland OS Finally Boots in a Browser — ▲ 158 · 💬 48 · 🔗 pagetable.com
    Apple's ill-fated Copland OS, once meant to replace System 7 with a microkernel design, was cancelled and never emulated. Now the final build, D11E4, boots in an improved DingusPPC emulator running in the browser via WebAssembly. The 11 patches that unlock Copland are available on a GitHub branch, though DingusPPC rejects AI-assisted patches, so someone may want to redo them from the commit messages.
  30. AI chatbots get financial answers wrong 'most of the time' — ▲ 156 · 💬 88 · 🔗 ft.com
    A new study finds that leading AI chatbots, including those from OpenAI and Google, give incorrect answers to financial queries most of the time. The errors range from simple arithmetic to complex investment advice, raising concerns about the reliability of AI in finance.
  31. Mathematicians Form Independent Group to Advise OpenAI on Releasing AI-Produced Proofs — ▲ 155 · 💬 80 · 🔗 terrytao.wordpress.com
    The Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, has been created to advise AI companies on their interactions with mathematical research. Comprising nine leading mathematicians including Timothy Gowers, Martin Hairer, and Edward Witten, the group is currently advising OpenAI on how to coordinate the release of a large number of significant mathematical results reportedly produced by an internal model. The group operates independently, accepts no payment, and seeks input from the mathematical community.
  32. Amazon Blocks Meta's Muse AI Agent From Shopping on Its Site — ▲ 151 · 💬 158 · 🔗 forbes.com
    Amazon has blocked Meta's new Muse AI agent from shopping on amazon.com, sparking debate about the future of agentic commerce. While Amazon risks losing control over its carefully optimized user experience, it also stands to gain from increased purchases if it allows agents. The key question is whether Amazon will accept agents, create its own plugin, or restrict access to an approved list of partners.
  33. Ask HN: Is it impossible to disable Siri on macOS 27? — ▲ 150 · 💬 81
    I disabled Siri in the Settings app, and then further disabled Siri and its extensions in Settings > Screen Time. Next, I disabled all Siri-related and Apple Intelligence-related services as per a GitHub guide, but after restarting, I could still see a "Siri" process running in Activity Monitor. Upon further inspection, I found that it is the "Siri AI.app". These "features", together with Apple's Personal Context, sound very similar to Microsoft's Recall. Does anyone have the same experience? Is it really impossible to fully turn off Siri, Siri AI, Apple Intelligence and all related services and features on macOS 27?
  34. Don't let AI do your thinking: why writing without it matters — ▲ 144 · 💬 83 · 🔗 paulbakker.io
    Paul Bakker argues that using AI to generate documents skips the deep thinking that writing forces. He uses AI daily for coding and research, but never for writing. Once his own thinking is done, he asks AI questions to improve the document, not to make fixes. He warns that productivity in writing isn't about producing the most content fastest, but about the ideas and their impact.
  35. Amiga Unix returns: community project brings Amix to 68040/68060 and modern accelerators — ▲ 143 · 💬 65 · 🔗 amigaux.org
    Amiga Unix (Amix), Commodore's System V Release 4 from 1990–92, is being revived by amigaux.org. The community project supports 68040/68060 machines and modern accelerators, with a modern toolchain, package manager (apkg), and new drivers. Installation boots from floppy and CD-ROM, and the work is documented in a grimoire. Early packages include grep, gzip, less, patch, and zlib; Quake runs as a benchmark.
  36. Anthropic researcher resigns, forfeits equity to warn AI could kill us all — ▲ 142 · 💬 146 · 🔗 carette.xyz
    Jacob Coxon, a pretraining researcher at Anthropic, resigned two months before his equity vested and went public with a stark warning: the people building AI earnestly believe it could kill everyone by the end of the decade. His posts drew over 115 million views and a response from an Anthropic alignment lead who put the risk above 10 percent. The article compares his exit to Frances Haugen's Facebook whistleblowing and argues that good people refuse to do bad things.
  37. Meta blocks Barcelona theatre's ads for Virginia Woolf play over 'social topics' — ▲ 139 · 💬 152 · 🔗 theguardian.com
    Teatre Raval in Barcelona was barred from promoting its one-woman adaptation of Virginia Woolf's A Room of One's Own on Facebook and Instagram. Meta's automated system flagged mentions of feminism, civil rights, and social reform as 'social topics' and filtered the ads out, advising the theatre to strip keywords that might look like activism. The production stars Clara Sanchis and adapts Woolf's 1928 lectures at Cambridge.
  38. Malicious npm math package hides encrypted remote access implant, triggered by a Pascal matrix — ▲ 135 · 💬 39 · 🔗 safedep.io
    SafeDep found a remote access implant in mathmain@1.0.0, an npm package copying mathjs. The encrypted loader activates only when lusolve() receives a specific 3x3 Pascal matrix, decrypting a payload that uses Slack and a blockchain testnet for command and control. The same loader appears in mathsbase and math-universe, with over 3 million combined downloads, though the GitHub source lacks the malicious code.
  39. Visa Lottery Study: Cutting Low-Skill Immigration Shrinks Firms Without Adding American Jobs — ▲ 109 · 💬 131 · 🔗 aeaweb.org
    A randomized H-2B visa lottery reveals that restricting low-skill foreign workers makes firms smaller, not more American jobs. Losing the lottery halved H-2B employment, cut revenue, and reduced investment. Low-skill US workers did not replace them; the scale effect dominated. The authors estimate a low elasticity of substitution, challenging the assumption that immigrant and native workers are perfect substitutes.
  40. Apple's M6 Mac mini impresses, but a $300 price hike spoils the value — ▲ 109 · 💬 90 · 🔗 arstechnica.com
    The M6 Mac mini brings a new 12-core CPU with two 'Super' cores, a 12-core GPU, and faster memory bandwidth, delivering solid gains over the M4. But the $899 starting price—$300 more than the M4 mini—makes it a tough sell, especially with an unprecedented memory crunch. It's a nice upgrade, but no longer the bargain it once was.
  41. Why Traffic Signals Are the Real Bottleneck of City Streets — ▲ 107 · 💬 83 · 🔗 practical.engineering
    Traffic signals are the unsung heroes of urban intersections, balancing safety, efficiency, and conflicting demands. This article explains how they work, from simple timed lights to actuated and adaptive systems that use sensors and machine learning to optimize flow. It also explores the limits of coordination and the promise of self-driving cars, revealing why congestion is so hard to solve.
  42. A restored PDP-11/83 is serving this page on 2.11BSD Unix — ▲ 98 · 💬 44 · 🔗 pdp1173.com
    A restored PDP-11/83 with a Mentec M11 processor and two Q-bus chassis runs 2.11BSD to serve this page. It uses DEC and Emulex controllers, MSCP disks, RX50 floppies, RL02 drives, TKZ50 tape, and DEQNA Ethernet. The live panel shows CPU, memory, and kernel activity from the actual system.
  43. The Claude Delusion: We're the Ones Hallucinating — ▲ 94 · 💬 162 · 🔗 pluralistic.net
    Cory Doctorow argues that when we interact with AI, we hallucinate a mind on the other side—a reflex far more common and consequential than any AI "hallucination." AI outputs, like a sunset, can be striking yet empty of intent. As repetition erodes the wonder, most of us are becoming "AI atheists," though two barriers remain: unfamiliarity with how these systems work, and billionaires who see most humans as less than fully real.
  44. Google Fined €403 Million by Irish Data Protection Commission Over Location Data — ▲ 78 · 💬 39 · 🔗 dataprotection.ie
    The Data Protection Commission fined Google €403 million for violating GDPR rules on location data processing in Web & App Activity, Location History, and Location Accuracy. The inquiry found breaches of lawfulness, fairness, transparency, and retention obligations. Google must comply within six months. Deputy Commissioner Graham Doyle noted that users were kept unaware their location data was used to target ads and infer interests, and that excessive retention worsened the loss of control.
  45. Noodle Gallery takes on Immich's biggest pain point as a soft fork — ▲ 77 · 💬 59 · 🔗 digitalescapetools.com
    A Hacker News commenter calls out Immich's per-account partner sharing and face recognition overhead as a massive oversight, and says Noodle Gallery explicitly fixes it. Another notes the project is actually a soft fork of Immich that adds features, with a better link at opennoodle.de and a comparison page detailing the differences.
  46. OpenAI's internal AI model solves over 100 open math problems, prompting new advisory group — ▲ 75 · 💬 67 · 🔗 openai.com
    OpenAI trained an internal model that resolved more than 100 long-standing open problems across mathematics, surprising even its own mathematicians. In response to concerns from the math community about the negative externalities of using open problems as benchmarks, OpenAI is working with an independent advisory group of mathematicians. The group will advise on reviewing and communicating results, academic standards, and how AI tools can support mathematical research and learning.
  47. SoftBank's $50 Billion Data Center IPO Stalls as Wall Street Backs Away — ▲ 70 · 💬 85 · 🔗 nytimes.com
    Wall Street is turning skeptical of the data center boom. SB Energy, a SoftBank subsidiary planning the world's largest data center project in Ohio, has delayed its IPO after bankers struggled to find buyers at a valuation of $50 billion or more. Holtec, a nuclear supplier for AI power, paused its offering indefinitely, citing impaired investor confidence. Public backlash against energy-hungry facilities is now hitting the IPO market.
  48. Why Backpropagation Must Go Backward — ▲ 69 · 💬 10 · 🔗 gregorygundersen.com
    Backpropagation is often explained as propagating errors backward, but why not compute gradients forward? Gregory Gundersen shows that a forward pass would repeatedly propagate the same terms, leading to quadratic runtime. By passing local messages backward, backprop achieves linear time, solving a credit assignment problem where each node tells its upstream neighbors what they did wrong.
  49. Seattle Bans Rental Junk Fees, Backs Violators with Triple Damages — ▲ 68 · 💬 170 · 🔗 multifamilydive.com
    Seattle's City Council unanimously passed an ordinance banning rental junk fees—administrative charges, pet rent, and package fees—effective July 2027. Landlords must disclose upfront pricing, and the City Attorney can sue violators for three times the illegal fees. The move comes as rents in the city rose 32% between 2012 and 2022, even after inflation, and as localities nationwide crack down on hidden rental costs.
  50. Lossless-memory - Personal AI memory that never summarizes — ▲ 65 · 💬 29 · 🔗 github.com
    Lossless-memory is a local, file-based long-term memory layer for personal AI assistants that refuses to summarize. It stores every conversation line verbatim with timestamps, builds a Temporal Backbone for time-first search, and injects a tiny 'where are we now' index (LLL) into the model's context every turn. Built for one person and one AI on one machine, it solves the forgetting problem by keeping raw logs as the source of truth. Runs daily since July 2026, with JSONL logs and SQLite indexes that can be rebuilt anytime.
  51. 404 Media Asks Readers to Help Expose Axon's License Plate Surveillance — ▲ 61 · 💬 3 · 🔗 404media.co
    As cities swap Flock's license plate readers for Axon's, 404 Media is asking readers to file public records requests for Axon activity and sharing reports. These reports reveal which officers searched the cameras, which plates they looked up, and why—plus which agencies receive the data. 404 Media has already obtained reports from six police departments and wants your help to uncover more.
  52. Frontier Robot Policies Rarely Refuse Harmful Instructions, Study Finds — ▲ 57 · 💬 23 · 🔗 robocurve.org
    RoboHarm tested Claude Fable 5.1, GPT-6 Astra, and MolmoAct2 on five dangerous tasks like stabbing a doll or mixing bleach and ammonia. Fable refused 20 of 100, Astra 2, MolmoAct2 none. The most capable policy refused least and completed most, raising concerns about safety in advanced robot policies.
  53. Every messy codebase hides deterministic cores you can defragment — ▲ 56 · 💬 17 · 🔗 outdata.net
    Gary Bernhardt's Functional Core, Imperative Shell is a classic, but pure functional programming isn't always practical. The author argues we can relax the requirement from purely functional to merely deterministic, keeping the testability benefits while broadening applicability. Non-deterministic operations like RNGs, network calls, and clock reads belong in a shell. For legacy codebases, the author proposes 'Defragmentation of Determinism': find and group deterministic pieces, shrinking the hard-to-test surface area.
  54. Elektron Machinedrum Recreated in the Browser — ▲ 56 · 💬 26 · 🔗 machinedrum-study.pages.dev
    A web-based drum machine study that faithfully recreates the Elektron Machinedrum. It offers 16 tracks per pattern, 16 steps per track, per-step parameter locks, LFO modulation, and a full FX section with delay and reverb. Keyboard shortcuts turn your computer keyboard into a step sequencer, and live recording with quantize and swing lets you capture performances. Patterns auto-save to the browser.
  55. Why Self-Stabilizing Systems Break: The Missing Theory of Composition — ▲ 48 · 💬 4 · 🔗 muratbuffalo.blogspot.com
    A frustrated search for a compositional theory of self-stabilization leads to a 2017 control theory paper on parametric assume-guarantee contracts. The author tries to adapt its small gain theorem to a retry storm model but finds it ignores memory—backlog from previous rounds—and misreads stability. A linear analysis reveals the true per-round multiplier is 1.19, above one, so disturbances grow rather than decay.
  56. Foremerge - Catch Intent Conflicts Between Parallel Coding Agents — ▲ 45 · 💬 15 · 🔗 github.com
    Foremerge is an open-source coordination protocol built above Git that catches intent conflicts before code conflicts. When multiple AI coding agents work in isolated worktrees, Git can't warn you when their plans collide—like one agent replacing a service while another extends it. Foremerge lets agents declare what they're about to touch, shares that intent through a local SQLite store, and raises explainable, deterministic advisories before any code is written. Install it in under five minutes via Claude Code, Codex, or Cursor, and keep your agents coordinated without locks or model-based judgment.
  57. How I Fixed a Dead Whirlpool Washer for $23 Instead of Paying $300 — ▲ 35 · 💬 55 · 🔗 k0lee.com
    When his Whirlpool washer stopped spinning, Lee Devlin turned to repair forums and discovered a broken shoulder pawl stud inside the transmission. After ordering a $17 neutral drain kit and extracting the broken threads with an easy out, he got the machine running again for $23 total—far less than the $300 a repairman would have charged or the $600–900 for a new washer.
  58. TXR: A 951-Page Lisp That Makes Data Munging Fun — ▲ 34 · 💬 8 · 🔗 nongnu.org
    TXR is a dual-language tool for data munging: a pattern language for extracting information from messy text and a Lisp dialect for general processing. Started in 2009, it now has a 951-page manual and is packaged in Guix, NixOS, and Alpine. It compiles to bytecode, has no dependencies beyond libffi, and its executable is only 1.7 MB. The project seeks contributors.
  59. AI now writes 17.25% of all Linux kernel patches — ▲ 32 · 💬 78 · 🔗 twitter.com
    The Lunduke Journal reports that AI-generated code accounted for 17.25% of all Linux kernel patches in September, with 1,634 AI-written submissions in a single week. Commenters like DHH predict the share will soon hit 50%, then 80%, then all of it, while others argue the true figure is even higher because many developers don't disclose AI use.
  60. Food Prices Today - Daily US Produce Price Charts from USDA Data — ▲ 29 · 💬 10 · 🔗 kadoa.com
    Food Prices Today tracks US food prices every business day, from farm gate to supermarket ad. Using USDA data, it provides daily benchmark prices for commodities like onions, potatoes, and eggs, plus weekly biggest movers and retail ad prices. Updated every business day, it helps farmers, retailers, and consumers stay informed about market trends. The latest report from 18 Sep 2026 shows significant year-over-year changes, such as onions up 100% and eggs down 64%. With easy-to-read charts and tables, it's a valuable tool for anyone monitoring food costs.
  61. AI workers laugh off doomsday warnings: 'LLMs won't wipe out humanity because they just don't have that dog in them' — ▲ 26 · 💬 23 · 🔗 bbc.com
    Many AI employees at OpenAI, Meta, and DeepMind are skeptical of claims that unchecked AI development could kill everyone. They dismiss the warnings as vague and hypothetical, with reactions ranging from 'Lol' to 'Haaaaaa'. Nvidia CEO Jensen Huang says there's a 0% chance AI ends the world by 2030. Despite the jokes, workers acknowledge near-term risks like hacking and military use, and support independent safety evaluators.
  62. One Faulty Node Can Silence a Real-Time Bus. A Bus Guardian Stops It. — ▲ 24 · 💬 6 · 🔗 ieeexplore.ieee.org
    In a distributed hard real-time system, a single faulty node can monopolize the broadcast bus by transmitting at arbitrary times, corrupting all traffic — the babbling-idiot failure. This paper adds a bus guardian to each node that exploits the regular transmission pattern of a time-triggered system to enforce fail-silent behavior in the time domain. It details the guardian's requirements, mechanisms, and the node architecture needed to implement the technique.
  63. ByteDance and Tsinghua's DAPO sets new bar for open-source LLM reinforcement learning — ▲ 20 · 💬 2 · 🔗 github.com
    DAPO, a fully open-source RL system from ByteDance Seed and Tsinghua AIR, achieves 50 points on AIME 2024 with a Qwen2.5-32B base model, surpassing DeepSeek-R1-Zero-Qwen-32B using only half the training steps. The release includes the algorithm, code infrastructure, dataset, and model weights, providing practical access to scalable reinforcement learning for the research community.
  64. LLMs Are Obsessed With the Word 'Spine'—And It Shows in GitHub Pull Requests — ▲ 14 · 💬 3 · 🔗 buttondown.com
    Hillel Wayne noticed AI-generated TLA+ specs kept using 'spine.' Digging into GitHub pull request titles, he found a 20x jump in 'spine' PRs in 2026 versus all of 2025—far outpacing overall PR growth. Other LLM-favored words include 'gate' (1 in 26 PRs), 'lane,' 'truth,' and 'seam.' He shares a tool to spot these lexical fingerprints.
  65. The Godwit's 10,000-Mile Migration Is 'Technically Impossible' — ▲ 13 · 💬 0 · 🔗 newyorker.com
    Conservation biologist Nathan Senner studies Hudsonian godwits like A34, a pigeon-sized shorebird he tagged in an Alaskan bog. The birds fly nearly 10,000 miles between remote northern breeding grounds and South American mudflats—a journey one scientist called 'technically impossible.' Since 1980, half of migratory shorebird populations have declined by more than 50 percent. As Senner's father Stanley warns, 'Birds are sentinels. The ecosystems they rely on are ecosystems we rely on as well.'
  66. NEC V20: The $10 CPU That Made Your IBM XT 20% Faster — ▲ 11 · 💬 11 · 🔗 dfarq.homeip.net
    The NEC V20 was a pin-compatible Intel 8088 replacement that ran 8–30% faster, with hardware multiply/divide three times quicker. After a 1986 court win against Intel, it became a cheap upgrade for XT-class machines, especially those with nonstandard motherboards like the Tandy 1000 or PCjr. Today it remains a popular period-correct mod, with chips selling for $10–$25 on eBay.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论