← Nov 2Tuesday, November 3, 2026 · all days
1.Astra can read and write weights to improve Qwen(rikiyat.github.io)
2790 points by dbreunig 18 hours ago | 853 comments | permalink
tl;dr: Researchers show that frontier LLMs (specifically "GPT-6 Astra") can directly read and write neural network weights in natural language, without gradient descent—identifying what games small policy networks were trained to play, locating copying heads in transformers, and writing playable Go networks from scratch. Most notably, Astra edited just 0.016% of Qwen3.5-0.8B's weights to fix a reasoning-loop failure, boosting MATH-500 from 7.9% to 48.1% and tripling IFEval. The authors call this "Weight-Space Programming" and argue it could enable auditing, unlearning, and model merging without retraining, though it couldn't add genuinely missing capabilities (e.g., to Gemma 3 270M).
HN Discussion:
  • ↓Results are overstated; weight editing and generation have existed in the subfield for years
  • ↑Genuinely shocked and amazed; this represents a stunning emergent capability of frontier LLMs
  • ~Impressive but unsurprising given LLMs' demonstrated ability to rapidly grasp complex systems
  • ↓Not emergent—likely result of deliberate training on weights data by OpenAI for recursive self-improvement
  • ↓The article conflates separate results and misleadingly frames the capability claims
2.Show HN: Hacker Times, a New York Times-inspired version of Hacker News(hackertimes.ca)
331 points by polpenn 10 hours ago | 104 comments | permalink
tl;dr: Hacker Times reformats Hacker News front-page stories as New York Times-style headlines and blurbs, grouping items like an MIT CSAIL report on GPT-6 Astra editing neural network weights, Karpathy's HackerTube video library, and a PostgreSQL release delay tied to AI-generated bug reports. Other featured threads cover privacy-respecting TVs, website archiving tools, a voxel model of Warsaw, and browser-based Vim practice exercises.
HN Discussion:
  • ~Similar projects already exist, suggesting this isn't novel
  • ↑Appreciates the design details like the Manufacturing Consent font choice
  • ↓Questions the value since NYT's worth is journalism, not UI; a browser extension would suffice
  • ↓Dislikes mimicking the clickbait online NYT rather than the classic print edition
  • ↑Enjoys the improved readability and story separation compared to HN's dense layout
3.DiscImageStudio: burn visible artwork onto CDs, DVDs and Blu-ray discs(github.com)
272 points by bpierre 11 hours ago | 54 comments | permalink
tl;dr: DiscImageStudio is an open-source (GPL-3.0) tool that encodes images into CD-DA, DVD, and single-layer BD-R data streams so the burned disc surface shows a visible picture, using physical-layer tricks like controlled byte patterns (0D/AA for CD, A5/92 for DVD scrambling). It offers a Windows WPF GUI plus experimental Linux/macOS CLI, supports ISO/RAW/WAV+CUE export or direct IMAPI2 burning, and includes calibration tooling that improves track-pitch accuracy from ~10⁻⁶ to 10⁻¹⁴. The project builds on the earlier CDImage work and warns that live burning is experimental—use expendable media.
HN Discussion:
  • ↑Shares related personal project using the same technique for a CTF challenge
  • ↑Nostalgic comparisons to prior disc-labeling tech like LightScribe and Neato
  • ↑Appreciates the tactile, artsy nature of the project as a refreshing throwback
  • •Points to prior art (CDImage) that the project builds on
  • •Questions practicality, such as whether discs still play or if drives are made
4.Show HN: Subvocal – Talk to your computer without making noise
733 points by aarushyadav 19 hours ago | 348 comments | permalink
tl;dr: Two Georgia Tech students (YC F26) built smart glasses that decode silently mouthed (or even closed-mouth) speech into text by emitting RF waves that pass through the face and get modulated by the tongue, lips, and soft palate. After abandoning EMG electrodes due to noise and usability issues, they trained foundation models on paid data collection to handle cross-user anatomical variation, achieving usable performance with just a 20-sentence calibration. Preorders are open for a 500-unit first batch, with shipping targeted for Q1 2027.
HN Discussion:
  • ↑Enthusiastic early adopters requesting hardware/fit considerations and preordering the product
  • ↑Sees major accessibility potential for deaf or speech-impaired users
  • •Curious about technical implementation details like on-device inference, offline capability, and noisy environment use
  • ~Raises privacy and enterprise security concerns about the technology
  • •Questions how this compares to existing silent-speech projects and alternatives
5.US boy names that became girl names and vice versa(namedaisy.com)
424 points by xoobdev 20 hours ago | 410 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Confirms article's finding that girls' names drive naming dynamics
  • ↓Article conflates distinct names that share spelling across languages
  • ↑Shares personal/cultural examples of gender-flipping names supporting the premise
  • ↓Questions the geographic scope and generalizability of the data
  • ~Suggests gender flips are often caused by single famous namesakes
6.If you love a website, back it up(blog.mistys-internet.website)
592 points by ingve 23 hours ago | 135 comments | permalink
tl;dr: A web archivist laments losing a favorite fansite to domain squatters and urges readers to personally back up sites they care about rather than relying on the Internet Archive. The post walks through three tools (all from the author's employer) for creating WARC/WACZ archives: the ArchiveWeb.page browser extension, its standalone app, and Browsertrix for larger crawls. Archived sites can be replayed locally, embedded in blog posts via ReplayWeb.page, or shared as files with others.
HN Discussion:
  • ↑Agrees with bitrot concerns and recommends additional tools like SingleFile for one-off captures
  • ↑Shares personal regret about losing their own websites/content, reinforcing the article's urgency
  • ↑Warns about broader threats like LLMs replacing/rewriting the open internet, extending the article's concerns
  • ↑Describes personal archival workflows and tools (bookmarklets, wget, IA extensions) as practical implementations
  • ~Suggests going beyond backup to re-implementing/improving beloved sites
7.Self-Employed? Solopreneur? NIST Has a Cybersecurity Guide Just for You(nist.gov)
252 points by tomrod 18 hours ago | 74 comments | permalink
tl;dr: NIST released Cybersecurity White Paper 50, a plain-language guide aimed at the 28 million U.S. "solopreneurs" who lack dedicated IT support. It's a scoped-down update to NISTIR 7621, drawing from the NIST Cybersecurity Framework 2.0 but assuming zero prior knowledge, with actionable basics like changing default device passwords, regular backups, and software updates. The document includes worksheets and three hypothetical use cases (a lawyer, artisan, and consultant) to help solo operators build a baseline security program.
HN Discussion:
  • ↓The guide's title and jargon undermine its goal of being accessible to laypeople
  • ↓The simplification attempt falls short and doesn't truly start from the basics needed for non-technical users
  • ↓The guide's backup advice is insufficient without emphasizing immutable, offsite, and tested backups
  • •Questioning whether withdrawing the broader small business guide leaves a gap in coverage
  • •Target readers asking experts to validate the guide's accuracy and actionability
8.Scaling Git at OpenAI(openai.com)
241 points by makeramen 12 hours ago | 166 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • •Notes this is part of a trend of scaling-git articles from AI companies
  • ↓Skeptical about why codebase would grow so massively and what engineers are doing
  • ↑Praises the post for being substantive, technical, and free of AI/marketing hype
  • ↓Critiques the lack of concrete benchmarks and performance numbers
  • ↑Appreciates the clever use and open-source contribution to packfile-uri
9.Show HN: Warsaw in voxels – An interactive 3D map of the city centre(p.migdal.pl)
289 points by stared 20 hours ago | 46 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Enthusiastic praise with personal memories of visiting Warsaw
  • ↓Project has severe performance problems on various hardware/browsers
  • ↓Accuracy issues with specific buildings and areas like Old Town
  • •Comparisons to similar voxel/3D city map projects
  • ~Usability feedback on controls and default settings
10.Ask HN: Who is hiring? (November 2026)
224 points by whoishiring 20 hours ago | 187 comments | permalink
tl;dr: Monthly Hacker News "Who is hiring?" thread for November 2026, where companies post job openings directly (no recruiters or job boards, one post per company). Posters must include location and specify REMOTE or ONSITE, and only list actively open positions they'll respond to. Links to search tools like gethnjobs.com and hnwork.app are provided, along with a pointer to the companion "Who wants to be hired?" thread.
HN Discussion:
  • ↑European companies posting onsite/hybrid roles following the thread's format conventions
  • ↑Remote-first companies advertising distributed engineering positions with salary transparency
  • ↑AI/ML focused startups recruiting for specialized GenAI and ML engineering roles
  • ↓Individual job seeker posting in the wrong thread (belongs in 'Who wants to be hired?')
  • ↑Deep-tech and mission-driven companies pitching ambitious technical challenges to attract talent
11.The SEC Should Block Anthropic's IPO to Protect the Public(techpolicy.press)
369 points by cdrnsf 14 hours ago | 275 comments | permalink
tl;dr: Anthropic's upcoming IPO shows ~$96B in cash and strong headline revenue, but footnotes reveal nearly $900B in non-cancelable compute commitments to Amazon, Google, Microsoft, AMD, and Broadcom, with unaudited financials showing revenue growth stalling in Q3 2026. The author argues demand weakness, fierce price competition, and soaring infrastructure costs make these obligations unsustainable, and that going public would shift catastrophic downside risk onto retail investors and retirees. He urges the SEC to block the IPO until Anthropic provides two quarters of audited statements with full disclosure of its spending commitments.
HN Discussion:
  • ↓SEC's role is disclosure, not deciding investment merit; blocking IPOs invites corruption
  • ↓If SpaceX and similar unprofitable companies were allowed to IPO, Anthropic should be too
  • ~Index fund inclusion creates real systemic risk by forcing must-buy demand regardless of fundamentals
  • ↓Article's claims about declining AI usage contradict anecdotal evidence and Anthropic's strong margins
  • ↓Transformational tech productivity is hard to measure early; exploratory use pays off later
12.Ubuntu still has not released patch for Linux 0day RCE in almost 3 weeks(freenode.net)
549 points by majorchord 18 hours ago | 164 comments | permalink
tl;dr: Ubuntu has gone nearly three weeks without shipping a patched kernel for CVE-2026-98432, a CVSS 9.8 RCE affecting 22.04, 24.04, and 26.04 LTS releases. No ETA has been provided, leaving internet-facing and multi-tenant hosts exposed with no official fix. Admins must choose between running the vulnerable stock kernel, applying unofficial mitigations, or migrating workloads to already-patched kernels.
HN Discussion:
  • ↑Ubuntu's failure to patch is unacceptable and reason to abandon the distro
  • ~Mitigations via iptables/nftables or proposed pocket exist and should be shared
  • •Questioning why Ubuntu is delaying without jumping to conclusions
  • ↓Risk is overblown since no exploits are seen in the wild
  • •Broader concern about unpatched Linux kernels in consumer devices
13.WTF Ubuntu?(chalmovsky.com)
1552 points by chalmovsky 21 hours ago | 647 comments | permalink
tl;dr: Ubuntu has failed to patch CVE-2026-98432, a critical (CVSS 9.8) remote root exploit in the Linux kernel's TCP code, more than two weeks after Debian shipped a fix and six days after a working exploit became known. All supported Ubuntu releases, including LTS, remain listed as vulnerable, despite Canonical's marketing promises of "timely patches" and paid security subscriptions. The author is migrating all servers off Ubuntu and notes that volunteer-run distros like Debian, Arch, and Alpine have already patched the flaw.
HN Discussion:
  • ↑Ubuntu's handling and communication around the vulnerability is poor and contradictory
  • •Fixes exist in the proposed/testing pocket and Canonical is working on stability testing
  • ↓The post is unconstructive outrage-bait lacking technical details
  • ↓The CVSS 9.8 rating overstates real-world risk; it's likely just a DoS, not RCE
  • ↑Competitors and other distros patching faster validates moving away from Ubuntu
14.HackerTube: YouTube built from videos recommended by HN(karpathy.bearblog.dev)
2841 points by karpathy 2 days ago | 632 comments | permalink
tl;dr: Karpathy, frustrated with YouTube's algorithm, used Claude Opus 5.5 to build HackerTube.org—a YouTube-style site populated with ~6,454 videos extracted from 40K HN comments recommending YouTube links from 2007-2026, categorized across 150 topics. The entire project took roughly 2 hours, with Claude handling comment scraping, filtering via Sonnet, UI, deployment, and hardening; Karpathy's only manual steps were getting a YouTube API key, provisioning a VPS, and buying the domain.
HN Discussion:
  • ↑Feed disaggregation offers opportunity to escape engagement-maximizing algorithms using AI filtering
  • ↑The unpersonalized HN-sourced feed surprisingly outperforms YouTube's personalized recommendations
  • ↑Appreciates the pure nerd content free from influencer and clickbait material
  • ~Concern that popularizing such scrapers will burden HN moderation and invite gaming
  • ↑Extending the project with derivatives like book versions and Apple TV apps
15.Show HN: I've updated my Linux System Call website and API(syscall.sh)
225 points by guitmz 1 day ago | 52 comments | permalink
tl;dr: An updated reference site cataloging Linux system calls across 27 architectures (x64, arm64, RISC-V, LoongArch, etc.), showing syscall numbers, arguments, registers, and calling conventions for kernels from 3.3 to 7.3. Data is extracted directly from kernel source via SYSCALL_DEFINE macros, tracks when syscalls were added or removed, and is available as JSON downloads or via a REST API at api.syscall.sh.
HN Discussion:
  • ↑Appreciation for the resource as useful reference material
  • ~Suggestion to verify syscall numbers locally rather than trusting any single source
  • ↓Technical correction about register usage (eax vs rax) on x86_64
  • •Questions about implementation details, build process, and why syscalls vary by architecture
  • ↑Author/related creator sharing complementary tooling like a CLI companion
16.Show HN: HackerBooks – HackerTube, but for books mentioned on HN(hackerbooks.org)
1828 points by roansh 1 day ago | 303 comments | permalink
tl;dr: Summary not available.
HN Discussion:
  • ↑Requests feature additions like fiction-only filters to improve the site
  • ~Points out missing books or classification errors needing refinement
  • •Shares interesting data/observations about book mentions on HN trends
  • ↑Hopes the site has longevity unlike similar past projects
  • ↓Jokingly mocks the proliferation of 'HackerX' derivative projects
17.Intel stock collapse worsens as $3.5T chip rout spreads(finance.yahoo.com)
213 points by entryleveljobsc 12 hours ago | 189 comments | permalink
tl;dr: Chip stocks shed over $400 billion Monday, with Intel falling another 9% following its 29% crash on Oct. 30 — leaving it worth less than half its late-September value. Memory and storage names (Micron, Sandisk, Seagate, WDC) dropped 6-8%, and Marvell slid after announcing a $115 billion deal to buy Lumentum, while Nvidia held up relatively well. Overall, chip and memory stocks have lost roughly $3.7 trillion in market cap since Oct. 6, though some analysts believe last week's sell-off marked a bottom.
HN Discussion:
  • ↓The 'collapse' framing is overblown; this is minor volatility compared to yearly gains
  • ↓Intel is too strategically important to fail; government or reality will backstop it
  • ↑Intel's decline is real and deserved due to lost engineering culture and leadership in all segments
  • ~The sell-off is temporary, driven by earnings timing and election-related market jitters
  • •Questions why Nvidia isn't falling as hard, suggesting Intel-specific issues beyond the broader rout
18.Postgres 19 is late because so many more bugs are being found now(rhaas.blogspot.com)
829 points by soheilpro 3 days ago | 497 comments | permalink
tl;dr: PostgreSQL 19's release is delayed from its usual September/October window to November, with several major features (SQL/PGQ, UPDATE/DELETE FOR PORTION OF, partition MERGE/SPLIT) reverted late in beta. The cause: LLMs are finding bugs at an unprecedented rate, with back-patched commits up 60-74% year-over-year, overwhelming human committers who must still review and apply fixes. The author worries about the long-term implications of 40 years of latent bugs now being discoverable, requiring developers to significantly raise their standards.
HN Discussion:
  • ↑Finding and fixing latent bugs is a net positive worth the release delay
  • ↑Personal experience confirms LLMs find massive numbers of real bugs in mature codebases
  • ~Verification and maintainer review, not discovery, is now the real bottleneck
  • ↓Author's dismissal of committer error as a cause is intellectually lazy reasoning
  • ~Shift toward security hardening comes at the cost of feature development
19.Supermarket freezers across Europe hit by second suspected cyber-attack(theguardian.com)
398 points by fergie 1 day ago | 440 comments | permalink
tl;dr: Supermarkets across at least seven European countries (Belgium, Estonia, Ireland, Finland, Luxembourg, the Netherlands, and Sweden) had refrigeration systems remotely hijacked overnight, with attackers cranking freezer temperatures up to 50°C and spoiling stock at chains including Coop, ICA, Ahold Delhaize, Carrefour, Ikea, and SuperValu. The attack mirrors a similar incident last month in France that hit ~700 stores via vulnerable Danfoss controllers, causing tens of millions in losses. With no ransom demands, analysts suspect sabotage targeting internet-exposed, unpatched operational technology controllers that remain widely reachable.
HN Discussion:
  • ~Technical defense explaining why freezer controllers are more complex than a simple knob
  • ↑Sarcastic appreciation that this attack serves as a wake-up call for securing critical infrastructure
  • •Speculation about the specific attack vector involving 4G-connected rural stores and exposed ICS systems
  • ↑Blame on store operators for ignoring vendor warnings to disconnect after the French attack
  • ↑Argument that OT devices shouldn't be internet-connected or should be read-only if they must be
20.Anthropic's internal AI use, at retail prices, now exceeds its employee pay(businessinsider.com)
334 points by entryleveljobsc 20 hours ago | 255 comments | permalink
tl;dr: Anthropic's IPO filing reveals that in April, the retail value of its employees' internal Claude usage surpassed its monthly employee compensation (excluding stock), with usage growing 250x from January to September. The figure reflects notional retail pricing rather than actual cash outlays, but underscores AI's growing role—over 90% of code added to Anthropic's software in September was written by Claude.
HN Discussion:
  • ↓The retail-pricing framing is misleading accounting; actual cash cost is what matters
  • ↓Heavy internal usage is expected and natural for employees dogfooding their own product
  • ↓Headcount growth contradicts the implied narrative that AI is replacing engineering labor
  • ~Unlimited internal token use distorts product design away from real user constraints
  • ↓The stat says little about Anthropic's actual moat or competitive advantage