← Oct 28Thursday, October 29, 2026 · all days
1.Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5(divagr.com)
556 points by csmantle 12 hours ago | 112 comments | permalink
tl;dr: A researcher compiled a working 32-bit RISC-V CPU (RV32IM) directly into the weights of Qwen 3.5 4B by exploiting Gated DeltaNet layers as exact, addressable RAM—no training, tools, or external sandbox involved. The CPU runs during the model's forward passes, executes C and MicroPython, passes official riscv-tests, and boosts arithmetic benchmarks to parity with giving Qwen a Python tool (e.g., NUPA long-integer from 48.6% to 100%). The author also added an in-weight filesystem persisting across conversations, and suggests the technique as a testbed for AI safety research on hidden reasoning and tamper-proof monitoring.
HN Discussion:
  • ↑Playful concern about AI safety implications of hiding a CPU and filesystem inside model weights
  • ↑Draws analogies to biological computing, seeing this as closer to brain-like substrates than silicon
  • ↑Fascination and curiosity about the technical mechanism of using DeltaNet state as RAM
  • •Points to related prior work doing similar in-transformer computation experiments
  • ↑Interest in extending the approach, e.g., to scratch-trained models or parameter-efficient benchmarks
2.Why is this fingerprinting script creating bioweapons?(nullpt.rs)
1349 points by hazebooth 19 hours ago | 250 comments | permalink
tl;dr: Why Is This Fingerprinting Script Creating Bioweapons?...
HN Discussion:
  • ↑Notes that open-weight models like GLM lack the safety guardrails that trip up frontier models
  • ↑Shares analogous experience of watermarks/markers triggering unexpected automated filtering systems
  • ↑Views this as an escalating arms race between obfuscators and LLM scrapers
  • •Questions the technical realism of the bioweapon content itself (ricin in water)
  • •Jokes about weaponizing the concept, e.g. terrorist-looking programming languages or spy scenarios
3.StackOverflow now gets fewer than 1000 questions a month(data.stackexchange.com)
561 points by babelfish 13 hours ago | 467 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Former active contributors drifted away and barely remember the site exists anymore
  • ~Decline predates AI; ChatGPT only accelerated an existing trend
  • ↑Hostile moderation culture (duplicates, rewrites, gatekeeping) drove users away
  • ↑Management betrayed contributors by restricting open data access for AI licensing deals
  • •Q&A sites naturally come and go; this is just the normal lifecycle
4.A Health Update(furbo.org)
293 points by frizlab 15 hours ago | 46 comments | permalink
tl;dr: The author recounts a grueling year of chemotherapy for lymphoma, including spinal taps, infusions, 50 pounds of weight loss, nerve damage, and a brush with pneumonia that nearly killed him. He's now recovering well—regaining strength, nerve function, and mobility—and is in the hospital for a final methotrexate treatment before his port is removed. In a fitting twist, he spent part of his hospital stay patching his servers against a critical PHP vulnerability that dropped mid-treatment.
HN Discussion:
  • ↑Well-wishes and support for Craig's recovery journey
  • ↑Sharing personal stories of family members battling cancer
  • ↑Admiration for the author's resilience and optimism in the face of cancer
  • •Recognition of Craig's technical legacy and contributions
  • ↑Appreciation for the mundane normalcy of debugging during treatment
5.CNC Stone Carving(carmelitegothic.com)
488 points by bschne 22 hours ago | 136 comments | permalink
tl;dr: Carmelite Monks are using 5-axis CNC machines from Italian manufacturer Prussiani to carve intricate Gothic stonework for church construction, reducing ornate pieces that once took skilled carvers months down to days. The workflow is fully digital: artist-monks sculpt designs in 3D modeling software (like Mudbox), programmers generate toolpaths in CAM software, and sawyers operate the diamond-tipped machines alongside wire saws for simpler profiles. The approach lets computer-literate young people contribute to traditional craftsmanship without needing hammer-and-chisel training.
HN Discussion:
  • ↑Historical context shows sculptors have always used automation tools, legitimizing CNC use
  • ↑Similar CNC stonework precedents exist, like Sagrada Familia and other monasteries
  • ↑Admiration for the monks' ambition and dedication to the project
  • ↑Technology should enable a broader return to ornate architectural styles
  • •Curiosity about technical details like drill bit materials and wear
6.Claude Code regularly sends each background agent’s transcript to its model(gist.github.com)
444 points by NikZaal 9 hours ago | 180 comments | permalink
tl;dr: Claude Code silently burns ~15-20% of user tokens (mostly on Opus) by having each background agent summarize its own transcript every ~30 seconds into a 3-5 word status line that never leaves the agents panel. There's no official setting to disable it, so the author ships an ~80-line Bun preload hook that wraps setTimeout/setInterval and suppresses the summarizer timer by matching its callback source text. Tested on Linux across Claude Code 2.1.257–2.1.312; it may silently fail after updates, so re-verification via OpenTelemetry is recommended.
HN Discussion:
  • ↑Confirms the token-waste problem with personal usage estimates supporting the article's claim
  • ↑Appreciates the cleverness of the setTimeout-patching hack as true hacker spirit
  • ↑Broader criticism that Claude Code is becoming bloated and inference providers shouldn't own harnesses
  • ↓Questions the premise, suggesting transcript summaries would mostly be cheap cache hits
  • •Adds related information about other Claude Code issues like email in headers or cache invalidation
7.The Grid, a factory-scale robot in Massachusetts that builds PCBs on demand(circuithub.com)
517 points by seddona 2 days ago | 232 comments | permalink
tl;dr: CircuitHub has revealed The Grid, a Massachusetts factory where levitating movers shuttle individual PCBs between solder, pick-and-place, reflow, and inspection stations, with AI scheduling eliminating per-design changeovers so mixed low-volume runs are as efficient as high-volume ones. A sample 5-board order runs $630 delivered in 4 days versus ~$465 in 9 days from a Shenzhen quick-turn shop. The system has operated for five years, shipped 2.2M+ boards, and is scaling 5x this year to meet demand from startups and major tech companies.
HN Discussion:
  • ↑Founder offering to answer questions and framing hardware as undergoing an AI revolution
  • ~Signup and quoting UX is broken, making it hard to actually use despite wanting domestic manufacturing
  • ↓Skeptical of the setup-time breakthrough and speed claims based on Shenzhen sourcing experience
  • ↓Board spec restrictions (1oz copper, no buried vias) limit applicability beyond simple prototypes
  • ↓Concept isn't novel; similar quick-turn PCB services already exist in UK/Europe and can't beat China on price
8.Gemini 4 Carbon(blog.google)
957 points by acossta 19 hours ago | 767 comments | permalink
tl;dr: Google announced Gemini 4 Carbon, a frontier model focused on cybersecurity and long-horizon agentic coding tasks, ranking #1 on Artificial Analysis Intelligence Index (59) and leading benchmarks like SEC-bench Pro (88.2%) and ExploitGym (54.8%). It claims major alignment improvements, including 0% success rate against prompt injection attacks and near-elimination of benchmark cheating. Initially rolling out to cyber defenders via the Fairwind Program before broader release, priced at $2/$10 per million input/output tokens.
HN Discussion:
  • ↓Google's confusing product naming and rollout makes it hard to actually access announced models
  • ↓Announcing benchmarks without releasing the model is meaningless given rapid competition
  • ↓The industry should standardize simple tiered model naming conventions
  • ↑Progress on reducing benchmark cheating and shortcut exploitation is genuinely important and welcome
  • ↓Timing announcements around stock market events like earnings calls is annoying and cynical
9.MacBook Pro with PROLED XDR Display and Touch(apple.com)
599 points by etothet 19 hours ago | 999 comments | permalink
tl;dr: Apple's new MacBook Pro introduces a tandem OLED "PROLED XDR" display with built-in touch (sensor integrated into the panel rather than layered on top), reaching 2000 nits peak HDR brightness and a 10,000,000:1 contrast ratio. It also adds a Dynamic Island for Spotlight and Live Activities, M5 Pro and M5 Max chips with Neural Accelerators in each GPU core, up to 24 hours of battery life, and a redesigned thermal and six-speaker audio system in a slimmer, lighter chassis.
HN Discussion:
  • ↓Touchscreen on Mac will degrade UX for non-touch users by forcing UI compromises
  • ↑Impressed by the significant weight reduction enabled by OLED and redesign
  • ↓Apple is losing its vision and design discipline with gimmicky features like touch and Dynamic Island
  • ↓Touchscreen laptops are inherently awkward and unnecessary, especially without a convertible hinge
  • ↓Demo use cases (like piano keys) show the touch feature lacks compelling, practical applications
10.AlpineBSD(wiki.alpinebsd.org)
234 points by bradley_taunt 22 hours ago | 216 comments | permalink
tl;dr: AlpineBSD combines the NetBSD kernel and libc with an Alpine Linux-inspired userspace, using apk-tools, busybox, s6 init (with planned OpenRC support), and a GNU-leaning toolchain. It's an original distribution rather than a fork, with an explicit policy against including LLM-generated code in the base system or main package repository. The project is early-stage with no installer yet, and inherits NetBSD's limited hardware support.
HN Discussion:
  • •FOSS is fragmenting into AI-stance camps, mirroring old licensing wars
  • ↓BSD-flavored distros for Linux users rarely sustain long-term interest or contribution
  • ↓The project isn't really 'Alpine' given its divergent package collection and init system
  • ↓The anti-AI policy is unjustified tribalism that will hinder maintenance of pinned packages
  • ~NetBSD's broad architecture support is a redeeming quality despite limited hardware coverage
11.Apple's git still trusts the compromised eMudhra certificates(emudhra.we-have-a-problem.com)
1154 points by galadran 1 day ago | 399 comments | permalink
tl;dr: Despite being notified of the eMudhra CA compromise on October 17th, Apple's bundled developer tools (git, curl, python3) on macOS still trust the compromised certificates, allowing HTTPS interception of git operations, credentials, and code. The macOS 27.2 update released October 26th failed to fix the issue and reverted manual workarounds users had applied. While Apple patched Safari and its networking framework, developers must manually remove the affected roots from /etc/ssl/cert.pem and reapply after each OS update.
HN Discussion:
  • ↑Apple's failure to patch system-wide trust store and reliance on manual edits is outrageous
  • ↑Apple needs a smaller patch mechanism instead of monolithic OS updates to address issues like this quickly
  • ↑Sharing Apple's official guidance and practical workarounds (MacPorts cert.pem) to mitigate the issue
  • •Certificate revocation infrastructure (CRL/OCSP) has been abandoned, exposing broader systemic issues beyond Apple
  • •Speculation that the eMudhra key compromise may have been intentional for intelligence exploitation
12.Dayjob: Tax – evaluating AI agents on 50 professional tax assignments(surgehq.ai)
246 points by landonxi 21 hours ago | 154 comments | permalink
tl;dr: Surge AI's new DAYJOB: Tax benchmark evaluates AI agents on 50 realistic tax assignments spanning eight practice areas, testing professional judgment on messy records rather than just well-defined questions. Frontier models performed poorly, with Claude Sonnet 5.5 leading at just a 10.8% pass rate (requiring 95% rubric satisfaction). Common failures include missing consequential deadlines clients didn't ask about and acting on unverified assumptions (e.g., marital status) despite contradicting evidence in the provided files.
HN Discussion:
  • ↓Users successfully use LLMs for personal tax prep, finding errors accountants missed
  • ↓Benchmark is unfair since it uses bare models without proper tooling/harness for tax work
  • ~AI is best used as a reviewer/explainer with humans making final decisions
  • •Developers building formal tax models/tools to augment AI capabilities
  • •Real bottleneck is data collection, not the tax computation itself
13.Emacs Turns 50(technologyreview.com)
495 points by internet_points 23 hours ago | 213 comments | permalink
tl;dr: EMACS, the full-screen text editor born at MIT's AI Lab in 1976, turns 50 this year and remains widely used. Created by Richard Stallman, Guy Steele, and David Moon as a standardized set of macros atop the TECO editor, it spawned commercial variants from Honeywell and others—whose commercialization pushed Stallman to launch the GNU Project, with GNU Emacs as its first major application. Apple shipped GNU Emacs with Mac OS X until 2019, and macOS still retains hidden EMACS keybindings today.
HN Discussion:
  • ↑Long-time users celebrate Emacs's enduring utility and longevity across decades of use
  • ↑Emacs remains more modern and adaptable than newer competing editors
  • ~Preference for minimalist editors over Emacs's ever-growing complexity and sprawl
  • •Curiosity about recent Emacs developments like AI integration and new packages
  • ↓Criticism that the article is poorly structured and ends abruptly without depth
14.Ghostty 1.4(ghostty.org)
1327 points by deaux 1 day ago | 421 comments | permalink
tl;dr: Ghostty 1.4 ships 7 months of work from 172 contributors, including a 100% complete Kitty graphics protocol implementation (the first terminal besides Kitty itself), transparent scrollback compression that cuts memory use 70-90% (raising the default limit to 50MB), and major performance gains—25-55% higher I/O throughput on macOS and up to 6x faster VT parsing. The release also adds a proper `ghostty +ssh` wrapper, a new GTK rendering pipeline using EGL/DMABUF that fixes blurry fractional scaling, macOS 27 compatibility, and patches for 10+ security advisories including sandbox escapes via Kitty graphics and paste protection bypasses.
HN Discussion:
  • ↑Developer celebrates libghostty progress and portable terminal core vision
  • ↑Enthusiastic praise for Ghostty's quality and the maintainer's dedication
  • ~Missing specific features (Kitty hints, cmd+F shortcuts) prevent full adoption
  • ↓Skepticism about Ghostty's value compared to established alternatives like iTerm2, Konsole, Wezterm
  • •Hope that new release fixes SSH issues enabling switch from iTerm2
15.Genie 4(deepmind.google)
1031 points by acossta 1 day ago | 583 comments | permalink
tl;dr: Google DeepMind announced Genie 4, a real-time interactive world model that generates photorealistic simulations with native audio, ambient sound, and interactive dialogue, steerable via keyboard controls and text prompts. It integrates with Gemini models (including Gemini 3.8 Flash as an in-the-loop reasoning agent) to maintain scene coherence and handle multi-step tasks, with applications in robotics training, filmmaking pre-visualization, game prototyping, and enterprise simulation. It's rolling out as a research preview to trusted testers and Google AI Ultra subscribers, with known limitations around object permanence, text rendering, and long-term consistency.
HN Discussion:
  • ↓Model fails at world consistency when objects leave and re-enter frame, undermining 'world model' claims
  • ↑Despite flaws, this represents impressive rapid progress compared to AI video just a few years ago
  • ↓Specific capability claims (audio, photorealism) don't match the actual demo outputs
  • ↓Skeptical that this unrealistic simulation tech can meaningfully train robots as claimed
  • ↓Generative AI has yet to produce culturally meaningful artifacts, mostly signaling lack of creativity
16.Kentucky man becomes the first humanist granted religious exemption from AI(religionnews.com)
470 points by danso 23 hours ago | 504 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • •Curious observer interested in how AI-refusers will fare, likening them to Amish
  • ↑Supports exemption, warning of AI's negative revenge effects on society
  • ~Supports protection but questions legal line-drawing for belief-based exemptions
  • ↓Skeptical that refusing AI is viable; productivity demands will make it untenable
  • ↓Dismisses as a non-story since the accommodation was small and routine
17.Ink droplets in motion snag top prize in Nikon Small World contest(arstechnica.com)
405 points by _Microft 3 days ago | 69 comments | permalink
tl;dr: Ole Bielfeldt won the 52nd annual Nikon Small World microscopy competition with an image capturing an inkjet printer depositing ink on paper, achieved by using a microscope objective in an unconventional way. The contest, founded in 1974, highlights light microscopy imagery, and is accompanied by the Small World in Motion video competition added in 2011. Nikon has published the top 10 winners and will release a 2027 calendar featuring them.
HN Discussion:
  • ↑Appreciation for other winning entries like the mouse brain and jumping spider images
  • ↑Admiration for the photographer's previous work and technical skill with magnification
  • ↑Awe at the otherworldly, hard-to-believe nature of the winning image
  • ↑Fascination with inkjet printer technology revealed by the microscopy
  • •Sharing additional resources like videos of the rig and printing process
18.Survival LLM: AI assistant and emergency library that runs on your own computer(brangus.io)
528 points by Luc 1 day ago | 196 comments | permalink
tl;dr: Survival LLM is a free, open-source offline AI assistant for emergencies that bundles a local LLM with a ~9GB library including Wikipedia, medical references, FEMA/CDC guides, military survival manuals, iFixit, and wikiHow. It auto-selects a model based on your RAM (from IBM Granite 4.2 8B on 16GB machines up to NVIDIA Nemotron 3.5 Super on 128GB Macs), cites sources for every answer, and runs fully offline on Apple Silicon Macs or 64-bit Windows 10/11. Recent updates added configurable reasoning limits, memory notes, and visible chain-of-thought.
HN Discussion:
  • ↑LLMs are useful as a semantic search layer over bundled knowledge, so hallucination fears are overblown
  • •Suggests alternative offline tools like Project Nomad or a custom 'doomsday computer' setup
  • ↓Tested the tool and found it gives dangerously incorrect survival instructions
  • ↓Hallucination risk and hardware overhead make a plain searchable archive (e.g. Kiwix) a better choice
  • ↓Content library is incomplete for real survival use, missing hunting/firearms/field skills
19.Raspberry Pi Zero 3W and Pi 4 Model A spotted in device trees(bret.dk)
222 points by sthlmb 1 day ago | 149 comments | permalink
tl;dr: Device tree files in Raspberry Pi's Linux kernel have revealed two unannounced boards: a Zero 3 W and a Pi 4 Model A, both using the Broadcom BCM2711 (quad-core Cortex-A72) from the Pi 4. The Zero 3 W would be a significant upgrade over the Zero 2 W's A53 chip, while the 4 Model A appears to be a stripped-down Pi 4 (no Ethernet, single HDMI, no analog audio) with possible external antenna switching. Pricing, RAM, and release dates remain unknown, though kernel support typically signals an imminent launch.
HN Discussion:
  • ↑Hopes the new boards remain affordable to be accessible for hobbyists and kids
  • •Concerns about the Wi-Fi chip (CYW43455) reliability and antenna quality on existing Pi boards
  • ↓Questions the need for a more powerful Zero, preferring a cheaper modernized original instead
  • •Wants USB-C port adoption on the Zero line to replace micro-USB
  • ↓Frustration with Pi availability and rising prices pushing users toward alternatives like ESP32
20.Git for the US Code(lexifina.com)
442 points by alansaber 1 day ago | 264 comments | permalink
tl;dr: A new tool applies git-style version control to the US Code, letting users view the statutory text in force on any given date, see which Act wrote each phrase, and track changes over time. The authors replayed public laws against release points from 2017 onward, successfully reproducing 361 of 365 changes across nine titles. The motivation is partly that LLMs frequently misapply superseded or temporally-wrong statutes—one cited study showed accuracy jumping from 24% to 87% when correct historical text was retrieved—making verifiable legal lineage essential.
HN Discussion:
  • ↑Enthusiastic support with call to institutionalize version control for legislation
  • ↑Points to similar existing projects in other countries, validating the approach
  • ~Agrees version control is critical but disputes that Git is the right tool for legal texts
  • ↑Working on similar projects, noting US/UK law complexity but viewing it as tractable
  • ↓Legal diffing has unique needs that software version control doesn't adequately address