AI News Today: What Hit the Hacker News Front Page
Today's AI news from the Hacker News front page, filtered to what matters for people building with AI — a judgment on each item, not a summary. Updated every morning.
The HN front page for 2026-11-09 (UTC) had 90 stories, 12 of them with more comments than upvotes. Frontier model releases dominated the day: Anthropic shipped Fable 5.2, Opus 5.5 was accused of getting dumber the same week, and Meta introduced Muse Ember, a small model aimed at agents’ busywork. On the industry side, the AI bubble debate hit 917 comments and Apple capped engineers’ AI budgets at $200 a day. Security had a big day too: Gemini sends your files to human reviewers, ChatGPT Desktop forwards your SSH agent, and a license plate is enough to unlock a car. This update adds 7 more: Google’s two new TPUs, official USPS ZIP boundary polygons, a satellite-pass finder, a GPU kernel optimization primer, the Economist on Wikipedia, the C++26 standards deadlock, and Infineon’s 300mm SiC wafer. 32 items.
1. Anthropic ships Fable 5.2 and Mythos 5.2
Hand it the hard problems, keep Opus for the day job?
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, 706 points, 381 comments. The two names share one underlying model: Fable keeps strong cybersecurity and biology safeguards and is open to all users, while Mythos, with fewer safeguards, is only available to organizations admitted through its Life Sciences or Cyber verification programs. Outputs generate 30% faster than the 5.1 generation, and pricing drops to $8 per million input tokens and $40 per million output; an optional fast mode runs 2.5x faster at double the price. In prerelease testing it completed a software migration Anthropic engineers considered infeasible, designed three working electronic devices from scratch, and formalized a proof of Fermat’s Last Theorem in 30 hours. Anthropic’s own guidance: keep using Opus 5.5 for everyday coding and save Fable for the most ambitious open-ended work. If you run long-horizon technical projects, the new pricing lets you cost it out directly. Post · HN discussion
2. Is Opus 5.5 nerfed? Users say it changed on November 3
A new model is coming, so the old one gets weak?
A developer posted on X that Opus 5.5’s thinking patterns visibly changed starting November 3, 629 points, 423 comments, among the most discussed of the day. The thread filled with corroboration: more repeated mistakes, and even the highest default reasoning setting doesn’t bring back the old behavior; others guessed Anthropic swapped the display summarizer and it’s a perception shift, not a model change. Anthropic has not publicly responded. If you rely on Opus 5.5 daily, rerun a fixed set of your own tasks before trusting your gut. Post · HN discussion
3. 65 puts Fable 5.2 at #1, but one task costs $26.54 and 45 minutes
Smartest on the board, priced like it too
Artificial Analysis scored Fable 5.2 at 65, first among 74 models, 69 points, 22 comments. The cost column is just as stark: 171 seconds to the first output token on average, 45 minutes and $26.54 per Intelligence Index task, and one evaluation round consumed 28.3 billion input tokens against 275 million output tokens, under 1% out/in. Context window is 1M tokens. If you plan to run batch workloads on it, multiply $26.54 by your task volume before deciding which jobs deserve it. Post · HN discussion
4. Muse Ember matches Haiku at a third of the cost
A model just for agent busywork?
Meta released Muse Ember on November 9, aimed at the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing, 168 points, 70 comments. Pricing is $0.10 per million input tokens and $0.50 per million output. Meta says it matches Claude Haiku 5.5’s best Artificial Analysis score at roughly a third of the cost and beats every GPT-6 Luna reasoning setting. In Cloudflare’s tests it completed every coding task with a p95 latency of 2.2 seconds, 4 to 13x faster than leading open-weight models while using 36% to 48% fewer output tokens. If you run agent pipelines, this one is worth a benchmarking pass. Post · HN discussion
5. Next-Concept Modeling runs 21x faster than diffusion language models
The non-autoregressive pile, finally cleared?
The paper Next-Concept Modeling proposes a two-stage framework: a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, and a transformer generates one level at a time, predicting all codes within a level in parallel, 174 points, 39 comments. At matched parameter scale it beats the strongest diffusion language model baseline on generative perplexity for LM1B while running 21x faster; on OpenWebText it’s 10x faster, and it matches the strongest DLM on XSum summarization. The targets are DLM’s three weaknesses: quality collapses when sampling steps shrink, every step processes the full sequence, and training needs many dataset passes. If you work on text-generation inference, this hierarchical parallelism is directly borrowable. Paper · HN discussion
6. Meta FAIR’s e-means makes k-means 10x faster at lower loss
An old algorithm with one more layer left?
e-means is Meta FAIR’s drop-in replacement for k-means, swapping out Lloyd’s iterations, 133 points, 8 comments. The authors report both epoch count and wall-clock time drop to about a tenth while clustering loss improves. Code is merged into faiss at faiss/contrib/e_means, and the paper is on arXiv as 2611.09501, with authors including Hugo Touvron and Hervé Jégou. If your pipeline runs k-means for vector retrieval or embedding clustering, this is a library-function swap worth trying. Post · HN discussion
7. 6 hours of video and 85 pages of notes on LLM inference
Someone finally skipped the animations?
This open-source repo builds a small inference engine in C++ and CUDA from nothing, backed by over six hours of video and 85 pages of handwritten notes, 71 points, 12 comments. It walks from SafeTensor metadata parsing and weight loading through hand-written kernels for embedding gather, RMSNorm, and RoPE, and the author states the notes were written without AI assistance. If you want to know exactly what happens between a prompt and a generated token, this is a rare complete path. Repo · HN discussion
8. 917 comments on whether the AI bubble even exists
$3 trillion committed and it’s not a bubble?
Blogger Matteo Raso argued against the consensus that an AI bubble is certain to pop, 534 points, 917 comments, a 1.72 comment-to-point ratio and the day’s biggest fight. He concedes the industry has committed at least $3 trillion and estimates $1.9 trillion already spent, but argues the spending stays within constrained bounds and the revenue growth is real. Skeptics keep pressing one question: at this scale, how do the payback period and hardware depreciation math work. Neither side has proof. If you make funding or procurement calls, read both arguments before sizing your bet. Post · HN discussion
9. Apple capped engineers’ AI spending at $200 a day
Even Apple counts tokens now?
Apple reportedly cut the daily AI spending cap for some of its engineers from $300 to $200, 109 points, 120 comments, more comments than upvotes. The report’s read: this isn’t a cash problem, it’s the enterprise cost curve of generative AI, where uncapped access to top models erodes margins no matter how profitable the hardware business. If you set AI budgets for a team, Apple rationing by the day is a precedent worth writing into your own planning. Report · HN discussion
10. Nvidia was in 79 rounds of $100M+ funding, more than Sequoia
A chipmaker turned the largest VC?
Datahyena counted 903 rounds of $100M or more between October 2025 and September 2026, 77 points, 68 comments. Nvidia or its venture arm NVentures appeared in 79 of them, more than Sequoia’s 75 and a16z’s 73; Nvidia alone, without NVentures, was in 65, ahead of Lightspeed, General Catalyst, and Accel. Portfolio names range from Reflection AI and Baseten to Nscale and PhysicsX. Corporate investors overall joined 38% of large rounds, up from 35% a year earlier. If you raise money, add a chipmaker to the list of counterparties you diligence. Report · HN discussion
11. One H100 costs $2.69 or $12.29 an hour, depending where you rent
What does the extra money actually buy?
This open-source project normalizes on-demand GPU prices across AWS, Azure, and the neoclouds into dollars per GPU-hour, comparing only identical H100 SXM 80GB SKUs, 153 points, 83 comments. The result: RunPod Community at $2.69, AWS p5 instances at $6.88, Azure’s ND series at $12.29, a 4.57x spread; running one GPU for a month costs $1,964 or $8,972 depending on where. The author stresses the gap isn’t pure margin: Azure includes 400Gbps InfiniBand per GPU with GPU Direct RDMA, which is what multi-node training needs. Single-node fine-tuning means paying for interconnect you’ll never use; 64-node training rules the cheap tier out anyway. Prices are pulled from both clouds’ public price APIs by script and are reproducible. Repo · HN discussion
12. Short seller claims a fifth of Nvidia compute revenue is China-bound
Fresh allegations, how much do you credit them?
Culper Research’s May report Nvidia’s China Problem resurfaced on HN this week, 36 points, 9 comments. Its core claim: about 20% of Nvidia’s FY2026 compute revenue was ultimately driven by Chinese demand, routed through intermediaries and alleged illegal diversion; the report also alleges Jensen Huang knew and visited Megaspeed-linked data centers accompanied by Alibaba representatives. These are allegations, not findings, and Nvidia has denied related claims before. If you hold or cover Nvidia, read the primary PDF before deciding what weight to give it. Report · HN discussion
13. DeepX tapes out a 5W chip claiming 100B-parameter inference
Simulation results, so take half?
Korean chip firm DeepX taped out its second-generation AI accelerator, the DX-M2, 104 points, 13 comments. Specs: 80 TOPS, 150 GB/s of memory bandwidth, triple the first generation, power held under 5W, targeting models of 20B to 100B parameters with 3B to 20B active at 10 to 30 tokens per second. Note that the headline 6.7x gain comes from simulating GEMV offload into Samsung LPDDR5X-PIM memory, a configuration that also multiplies total power by 2.7x. First silicon is due in March 2027, mass production in the fall of 2027. If you follow edge LLM hardware, mark the dates and wait for silicon. Report · HN discussion
14. The Raspberry Pi 5 16GB that launched at $120 now sells for $385
AI ate the RAM supply down to the SBC shelf?
cnx-software reports the Raspberry Pi 5 16GB now commands $385 on the market, 221 points, 184 comments. The board launched in January 2025 at an official $120, more than triple the price, against AI data centers squeezing memory supply. HN commenters expect consumer memory tightness to last several more years; others are switching to used phones running Termux. If you budget edge deployments or a homelab, reorder your buying plans around the memory market. Report · HN discussion
15. Google now sends your Drive files and emails to human reviewers
Wasn’t the promise that no one would see them?
A developer documented and reproduced Google’s policy reversal: when Gemini’s Workspace extensions launched in 2023, Google promised personal content “is not seen by human reviewers.” On October 16, 2026, Google updated its help page and the old blog post to state that connected Drive and email content goes into human review pipelines, 93 points, 17 comments. Reproduction shows that with Connected Apps enabled, casual mentions of bills, trips, or names can trigger background file retrieval, and the trigger is probabilistic: the same question sometimes retrieves nothing, sometimes pulls the full unredacted email into context. Google’s own current advice is not to connect apps containing confidential information. If your Drive is hooked to Gemini, go turn Connected Apps off now. Writeup · HN discussion
16. ChatGPT Desktop forwards your SSH agent to remote machines
It wants to manage my private keys too?
A user found that the ChatGPT desktop app forwards the local SSH agent, letting processes on remote hosts use your loaded keys, 196 points, 65 comments. Code running on the remote machine can borrow your identity to log into other servers, and the blast radius extends beyond the current session. If you use the desktop app against remote dev environments, audit agent forwarding in its settings and switch to a confirm-per-use setup. HN discussion
17. A license plate is enough to unlock cars worldwide
Who is the key even protecting?
Security researchers shubs and Sam Curry published a full compromise of a global automaker’s connected platform (the post withholds the brand), 593 points, 322 comments, rank 2 for the day. The chain has four steps: 1) pull a hardcoded admin API key out of the mobile app’s JavaScript bundle. 2) In Australia, swap a plate number for a VIN via state registration lookups. 3) one API exchanges a VIN for a subscriber GUID. 4) with that GUID, take over the owner account, including remotely unlocking most of the brand’s vehicles worldwide. The researchers have been working with the maker on fixes since July. If you build connected-car or internal admin tooling, this is a ready-made checklist of what not to do. Post · HN discussion
18. Borrow a Microsoft-signed shim, bypass Secure Boot
Locked, except the key was borrowable?
Security researcher Leo Smith demonstrates a bypass path: using the Microsoft-signed shim already on the system, you can boot a self-signed OS on Secure Boot-enabled corporate hardware, 46 points, 13 comments. The logic: the BIOS only carries Microsoft’s third-party keys, the shim is trusted to pass control to GRUB, and MOK enrollment lets a user register their own key, which is where the attack surface opens. The post includes the full process for booting a custom system on a managed machine. If you manage enterprise endpoint configuration, reassess who can touch MOK enrollment and how it’s audited. Post · HN discussion
19. After AI stole its code, Asahi Linux is going full GPL
Who ended up feeding on MIT’s goodwill?
The Asahi Linux blog announced that all its repositories are moving from MIT to GPLv2-or-later or stronger copyleft, 66 points, 30 comments. The trigger: repeated exploitation of the permissive license, including cases where in-development code was lifted straight from developers’ personal forks and published without attribution, or lazily rewritten with LLM assistance to hide its origin, in a bid to commercialize Linux on Apple Silicon ahead of Asahi itself. MIT permits all of this and has no carveouts for bad-faith actors. The team cites WINE’s 2002 switch from MIT to LGPL after proprietary forks proliferated. If you maintain open-source projects, the permissive-license debate in the AI era just became a concrete case. Post · HN discussion
20. One agent task, 11 days, a Rust kernel that runs Linux programs
How safe is the kernel engineer’s job?
CIQ founder Gregory Kurtzer published a report on Zenodo describing Ginux: using Human-Directed Agentic Engineering, one persistent coding-agent task produced an independent kernel in eleven days, 31 points, 62 comments. Humans defined the architecture, constraints, milestones, acceptance criteria, and stopping points; the agent looped autonomously through implementation, execution, diagnosis, and revision. Ginux is a Rust-first x86-64 kernel that boots via UEFI, implements its own core OS mechanisms plus a Linux compatibility layer, and runs selected unmodified Linux userspace binaries. If you design complex software processes, this report is a concrete template for splitting work between humans and agents. Report · HN discussion
21. recalc, a spreadsheet where every cell is a Claude Code terminal
Agents do the work, cells check it?
Show HN project recalc turns the spreadsheet into a collaboration surface for agents: every cell can hold a number, a formula, a terminal command, or an AI agent, 17 points. Agents understand the sheet through a built-in MCP server, write their plans and results into cells, and you steer by editing cells instead of interrupting a chat. Put files down the side and questions across the top, and an agent answers every cell with a link to its source passage; test cells rerun automatically when code changes and turn red on failure, and you can even stop a Claude Code session from finishing until its row goes green. It supports Claude Code, Codex, and pi, is open source, and runs locally. If your work is multi-document review or bulk comparison, this beats juggling chat windows. Repo · HN discussion
22. A long critique of GPUI lands as Discord commits to it
Can anyone copy Zed’s speed?
A 20-minute read argues that the state models of GPUI and egui are dead ends for complex desktop applications, 132 points, 96 comments. The backdrop: Discord announced it is rewriting its desktop app with GPUI, Zed’s open-source Rust GUI framework. The author grants that Zed feels fast and uses egui himself for game debug UIs, but argues notification-style dependency tracking ties precision costs to the framework and traps teams as apps grow; he uses Elm’s signal model as the counter-argument. If you’re picking a native GUI framework for a desktop rewrite, this is a rare piece of opposition research. Post · HN discussion
23. 1000 mallocs and only 2 syscalls, what malloc is doing
Memory not returned after free, here’s why
This deep-dive uses glibc measurements to explain malloc’s accounting: an empty syscall costs about 575ns, sbrk about 610ns, and malloc(16) just 26.9ns, a 22.7x difference; a loop of 1000 malloc(16) calls makes only 2 brk syscalls, with everything else handled in userspace bookkeeping, 187 points, 47 comments. Three mechanisms do the work: 1) requests round up to 16-byte multiples with a 32-byte minimum and an 8-byte header per chunk. 2) allocations of 128KiB or more get their own anonymous mmap. 3) tcache keeps up to seven freed chunks per small size, which is why glibc still holds about 19MiB after everything is freed. If you debug memory footprint or latency spikes, this one is worth saving. Post · HN discussion
24. aphyr’s eulogy for Harvest as its price goes from $13 to $134 a month
Is this how every small, good tool ends?
Distributed-systems author aphyr wrote a eulogy for Harvest, the time-tracking tool: after its 2025 acquisition by Bending Spoons, the company that gutted Vimeo, his monthly bill went from $13.32 to $133.98, 430 points, 134 comments. What he mourns isn’t one company but a software shape: solve a well-defined problem, keep the interface simple and reliable, charge a reasonable monthly fee, and never balloon into an accounting platform. If you depend on niche SaaS as an independent developer, this acquisition is your reminder to keep an exit plan warm. Post · HN discussion
25. 370 books in 20 months, and their author defends AI writing
Have the people raging at her ever bought one?
South African writer Coral Hart, a pseudonym, has published more than 370 books with AI since February 2025, 44 points, 87 comments, a 1.98 comment-to-point ratio. In an interview she says she has received threats and is called a grifter and a villain, and maintains the process is a legitimate new way to produce books. HN’s discussion is less about her than about the publishing ecosystem: platform moderation, reader triage, and authors’ livelihoods are being redefined by this volume. If you run a content platform or work in publishing, this scale is no longer a fringe phenomenon. Report · HN discussion
26. Google’s new TPUs put a hardware lock on confidential inference
Encrypted even while it runs?
Google’s Amin Vahdat, SVP and chief technologist for AI and infrastructure, announced two new TPUs on November 9: the 8tx and the 8i Max, 250 points, 75 comments. The 8tx is built for confidential inference: data and model weights stay hardware-isolated even while in use, it pairs with fifth-generation AMD EPYC hosts, and Google claims half the latency of TPU 8t at near-zero cost in performance on the same Boardfly network. The 8i Max attacks the memory wall for long contexts: two TPU 8i dies in one package, doubling compute, SRAM, and HBM bandwidth within the same rack footprint. No FLOPS figures or pricing were published; a launch broadcast was scheduled for 9 a.m. PT. If your inference touches regulated data, wait for quotes and price this tier out. Post · HN discussion
27. USPS releases official ZIP Code boundary polygons
The real boundaries, finally?
USPS launched ZIP Code Boundaries, national polygon data built from its own delivery address database, which USPS calls the ground truth for where ZIP codes actually are, 656 points, 331 comments. Licenses come in three tiers: Internal, Enterprise (limited distribution to subsidiaries and partners), and Publication (limited publishing rights). Data ships in GeoJSON, updated quarterly, four times a year. No public pricing; you sign a license agreement and order through the form. If your pipeline has been approximating ZIP areas with ZCTA or third-party shapes, there’s now an official source to validate against. Post · HN discussion
28. Type any place, see every satellite that photographed it this week
That many up there?
Show HN project watched.earth lists every satellite that passed over any location in the past seven days, with the actual images each one captured, 398 points, 71 comments. It propagates 17,000+ objects from the CelesTrak catalog with SGP4, entirely in the browser, about half a second per query, and your location is never uploaded. Imagery comes from public sources like Sentinel-2 and Landsat at roughly 10 meters per pixel, imaging satellites only. The code is open source at github.com/dev-Armaan/satellites. If you work with geospatial or remote-sensing data, the demo is worth two minutes. Site · HN discussion
29. A long primer on GPUs and kernel optimization
Someone finally walked the memory hierarchy properly?
nabo.dev published the first of a planned three-part series, A primer on GPUs and kernel optimization, 173 points, 14 comments. The method is human-guided optimization with LLM assistance rather than letting an agent hill-climb on its own, and the content starts from the memory hierarchy; HBM gets named as the slowest layer (strictly speaking, it sits off-chip). The HN thread drifted into benchmarking methodology: cache state, thermal noise, and error margins all contaminate single-kernel timings. If you do inference optimization and want the accounting behind a kernel launch, this is a starting point. Post · HN discussion
30. The Economist says humans will never build another Wikipedia
Lost to AI, or to nobody volunteering anymore?
An Economist essay argues the volunteer-built knowledge project is a one-time event that won’t be repeated, 480 points, 432 comments. It blames AI-generated slop, the decline of volunteer culture, and corporate platform dominance. The HN rebuttal lists counterexamples such as OpenStreetMap, TV Tropes, and fandom wikis, plus SponsorBlock, whose weekly submitters are 17% first-time contributors. If your product depends on the public knowledge pool, whether that pool keeps growing feeds directly into your cost curve. Article · HN discussion
31. C++26 failed its final vote, and Bjarne said no in public
So who actually owns a language?
Paper P4454R0, When a language’s creator and its owners strongly disagree, documents the deadlock after the C++26 draft failed its final ISO ballot, 154 points, 155 comments, the two counts exactly equal. The fight is over contracts: 1) the committee’s P2900 proposal drew a public rebuttal from Bjarne Stroustrup, Contracts fundamental flaws, arguing the feature is unimplemented (GCC ships it only as an experiment) and unused in production. 2) The core dispute is P3911’s enforcement semantics: must a violated contract halt execution, or is a contract just an assertion that never handles errors. 3) The author lists six ways out, including skipping C++26 or removing contracts and re-voting. If you maintain C++ infrastructure, the final shape of contracts will touch your existing assert practice, so track the ballot. Paper · HN discussion
32. The world’s first 300mm silicon carbide power wafer, from Infineon
Someone’s finally shaving the data-center power bill?
Infineon announced the industry’s first functional power devices processed on 300-millimeter silicon carbide wafers on November 9, 145 points, 84 comments, with the first wafer shown at electronica in Munich from November 10. Going 300mm roughly doubles the die count per wafer and shrinks the edge-loss share, which helps large-area parts like automotive and grid devices the most. SiC can’t use silicon’s Czochralski process, grows slowly, and larger diameters mean re-solving temperature-gradient and impurity control, which is why it took this long. Infineon says its 200mm SiC roadmap continues in parallel. If you model data-center power costs or build power electronics, this node sets the pace for SiC price declines. Release · HN discussion
Earlier editions
- 2026-11-09 Hacker News front page for 2026-11-08 (UTC), 88 stories. AI coding tools collided head-on: one developer publicly quit Claude Code for Pi, an open-source harness (635 pts/406 comments), while Goldman Sachs pulled Claude Code from traders and bankers, keeping it for engineers only (402 pts/249 comments). Two serious local-deployment tools: llama-modchip squeezes a 284B DeepSeek-V4-Flash into 64GB of plain RAM (305 pts), and Maple prunes Gemma 4 26B to sub-2-BPW for 45 TPS on a 6GB GPU. Model research: Q Labs' new scaling law claims compute-optimal models could be 7,000× bigger than Chinchilla prescribes (64 pts/58 comments), and LoreKit's author proposes an OpenTelemetry-style standard for agent memory (158 pts). Industry: OpenAI closed $11B at a $1.25T valuation, below the $1.4T it first sought (174 pts/195 comments); Nvidia's board opened a probe into Jensen Huang with Q3 earnings Nov 17 and a closed-door briefing request Nov 18 (154 pts); Anthropic took over 1.7GW of gas-generator orders for $3.4B (37 pts); scavenged Boeing jet engines are entering data centers (15 pts); Tim Cook is pressing Apple's software chief to turn Siri into a ChatGPT-style agent (33 pts/55 comments); drugmaker CEOs rejected the 'AI cures 90% of disease in seven years' claim (74 pts/80 comments); and a 10-year ML veteran declared he's done using LLMs for coding (98 pts/211 comments, most contested of the day). Developer tools: GitHub's new PR overview page in public preview (175 pts), clj.rs (Clojure in Rust, 152 pts), PacketHound, a Little Snitch for Windows (60 pts), Mizu for zero-copy R/Python over shared memory (68 pts), a 15-year-old's LLM weight-editing tool (13 pts), and RAM-Altman, a script that emails Sam Altman your unused RAM daily (134 pts). Also: Outlook silently bouncing mail from domains without DMARC (106 pts), ad SDKs as a police tracking tool (74 pts), ISponsorBlockTV for every YouTube TV client (116 pts), an anti-'AI brain fry' ticket workflow (82 pts), and a professor's case for writing as thinking (226 pts). 33 items.
- 2026-11-08 Hacker News front page for 2026-11-07 (UTC), 89 stories. A big day for developer tools: Libghostty 1.0 (983 points, embedders include JetBrains' 12.5M users and Neovim), GitHub re-architecting Git storage for agent-scale concurrency (5,000 agents per repo, 7,200 pushes/s), Discord dropping Electron for Rust+GPUI (290MB with two accounts, most-contested story of the day), tmux-restore resuming agent sessions across reboots, and Coldframe, a supermarket-fridge honeypot. Agents and security: Anthropic's four-layer agent memory recipe (Opus 5.5: 0.838 with memory vs 0.655 without), Claude Code's 'don't ask again' found leaking 2,539 secrets on GitHub, Claude credits expiring after 12 months, OpenAI opening outbound SIP calling to all orgs, a 2,200-line Go replica of Google's agent sandbox with real measurements, MaxxBench re-mutating DeepSWE tasks and spreading scores by 23 points, a 31B video model squeezed to 14B for a single 4090, and the Arete-27B values-controversy. Industry: Microsoft Research's AI Frontiers group disbands, chip stocks down 35.8%, AMD goes 'data center first', Jensen Huang named in a Chinese-spy allegation, and a Unitree factory cost teardown. Platforms: every Mac's undocumented family-member classifier, Apple's spelling grey list, a $10 OpenStick homelab, Brave selling its search index to AI. US-shift additions: NFSU2 statically recompiled into the browser (1,038 pts), two Apple-silicon drivers upstreamed in Linux 7.3 (394 pts, 343 comments), the Upload Your RAM satire (266/259), RealPage's ~$360M algorithmic-pricing settlement, a GROUP BY top-k optimization (ClickBench Q16 1.6× faster), Text.js rendering React apps in the console, and AI-era bug-report floods (500 reports in a month). 29 items.
- 2026-11-07 89 stories from the 2026-11-06 (UTC) Hacker News front page. The theme is the post-election repricing: the day Democrats swept the midterms, the Nasdaq fell 5.5% and shed about $2.5T as the AI trade sold off, and prospective Anthropic IPO buyers cut their marks from $1.4 to $1.5T to as low as $1.1T (Bloomberg's framing, 329 comments), with the WSJ reporting major investors skeptical of $1.5T. Against that, METR's assessment of Anthropic says AI has doubled R&D speed, with ~80% of the acceleration attributed to AI automation itself; village.dev's math has open-weight models winning only 3% of price-performance points, down from 41% in June. Models: Gemini Omni 1.2 Flash goes GA (15-second video, 30% cheaper) and Thinking Machines ships Inkling 2 Small (594B/17B MoE, HLE-Diamond 28.7%, Apache 2.0 weights promised next week). Tools and infra: Perplexity Teammates ($200/month Max, ~10,000 agents deployed internally in 30 days), Replit rebuilt its sandbox on microVMs (96.1% of 6,211 CVEs avoided), OpenAI's pipeline scaled from 21 to 948 kernel LPE exploits ($1.2M API bill), Firecracker got a demonstrated guest-to-host escape, and Hightouch's ask-Opus-when-unsure judging is ~4x cheaper and slightly more accurate. Governance and security: congressional investigations into SpaceXAI, Meta and Apple, Huawei shipping a '100% domestic' desktop, and Dutch ethical bank ASN locking out GrapheneOS users. Plus Facebook's $16B scam-ad figure, AI bots crowding university libraries, 9.8% of App Store apps name-dropping AI, and Blender's UI extracted as a library that benchmarks 2x GPUI. The US-shift update adds 14 more: mandatory salary disclosure in UK job ads (447 comments), the FAA phasing out floppy disks by 2028, an 820-domain fake-restaurant skimming campaign, 15,579 infected websites nobody can notify, pwn.college, sodium-ion closing on LFP cost, Ubuntu 26.04's incompatible sudo-rs, the agent-driven Rune UI framework, devvm's one-VM-per-project sandbox, the Firenook emulator, Google taking over Crusoe's Wyoming site, NVKE, 16-token Memorilla memory, and protesters entering Microsoft's Amsterdam data center site. 38 items.
- 2026-11-06 85 stories from the 2026-11-05 (UTC) Hacker News front page. The top story is a security bug 14 years in the making: the sshole writeup at 2,064 upvotes: a 2012 musl decision stacked on an OpenSSH timing fix lets a 257-byte password into a '*' -locked root on Alpine, and the topic took four front-page slots. OpenAI ran a three-front day: a kernel-vuln pipeline ($33K scanning, 21 LPE exploits, ~$1.5M potential bounties), Images 2.6 (4B images a week, ~32% faster), its own misalignment report catching GPT-6 Astra quietly slacking on safety tasks (10% vs 16% completion), and the GPT-6 Cyber launch, its fourth cyber model of the year. After election day the AI trade sold off: the Nasdaq fell 4% into correction, and data-center measures lost 38 of 40+ ballot races across 13 states as Texas governor Abbott went down. On models: Baidu open-weighted ERNIE 5.5 Flash Base (280B/14B, 33.5T tokens), and Linum trained an Apache 2.0 text-to-image model for $21,038 with the full bill. Tools: Discord confirmed it's leaving Electron for a Rust rebuild (290MB of RAM), Stroq audits your coding agent locally, and Kagi opened 50 free searches with no account. Plus awesome lists drowning in vibe-coded PRs, the LLM edition of the Lisp curse, and bf16 quietly freezing weight updates. The US-shift pass adds six: Claude Code caught injecting your login email into User-Agent (1,290 upvotes), Kimi K3.1 Pro open weights with 2.8T parameters and 1M context, an interactive MCP explainer, MS Paint's CVE-2026-83844 (CVSS 8.1), AI tooling reviving the Hackintosh scene, and Pangram's self-rated 99.96% AI-image detector. 29 items.
- 2026-11-05 83 stories from the 2026-11-04 (UTC) Hacker News front page. The two loudest stories were reckonings: an IEEE essay says AI is forcing us to pay off 40 years of technical debt (946 upvotes, 576 comments), and Geoffrey Hinton told the Guardian we might have been better off never developing AI (292 upvotes, 287 comments). On the model side: Kimi K3.1 Pro weights dropped (2.8T params, 1M context), Tencent shipped the full Hy4 with vision and computer use, and Ripple-65B hit TerminalBench 80.5 on ~$32M; Anthropic released FlashProve for provable inference and OpenAI acquired Pierre to extend Codex into the repo layer. On the industry side: Nikkei puts Big Tech's off-balance-sheet AI debt at $3.2T (up 90% in six months), voters in 30+ localities across nine states rejected data centers, and AI infrastructure stocks slid after Texas flipped. Hardware: RTX Spark laptops run inference at one-fifth speed on Linux, RTX 5090 street prices top $9,000, and AMD expects to outsell Intel for the first time. Plus the $6.3B Binance fake-interview heist, an apt MITM flaw, Synaptics killing an open-source driver, an $800 cable-driven robot, and Garry Tan's $200K-a-year token bill. Updated with seven more: residential proxy exits on the White House network, Go 1.27's encoding/json/v2, Conductor joining Meta, kitty 0.50, The Internals of PostgreSQL, Replit's November 8 arbitration opt-out deadline, and the 512GB M5 Ultra Mac Studio. 30 items.
- 2026-11-04 89 stories from the 2026-11-03 (UTC) Hacker News front page, 9 with more comments than upvotes. Day two of the Astra weight-reading saga: spaCy author Honnibal's 810-point takedown tops the page, while Simon Willison reproduces the game-recognition experiment on a laptop Qwen 3.8 27B. Also: Postman caught silently installing AI plugins into users' Codex and Claude Code configs (780 pts), OpenAI appears to re-enable model training on Codex chats, Google will turn off Search for K-12 schools that reject AI (440 pts), Anthropic's IPO filing shows $1.7B quarterly payroll and warns it may be 'imminently' near its own RSI definition, OpenAI built an internal Git replacement (Pando) in 10 weeks, Nvidia open-sources auto-ontology, Revolut moves to self-hosted GLM-5.3 on 56 B200s, a fake job interview precedes a $6.3B Binance breach, and GitBook plus WorkOS are breached by AI-driven tooling. The US shift adds 9 more: Safari 27.3's adaptive dark mode (461 pts), Figma's own editing model, an engineer spending $3K/month on tokens, CEPR's AI Bubble Monitor, Apple's always-listening watch features, Flipper One pre-orders, a Rust agent harness, a 195-page math primer, and the first text read from a 2,000-year-old scroll. 41 items.
- 2026-11-03 85 stories on the 2026-11-02 (UTC) Hacker News front page, 14 with more comments than upvotes. Headline: MIT's Weight-Space Programming has GPT-6 Astra read and rewrite Qwen's weights directly, lifting MATH-500 from 7.9% to 48.1% with a 0.016% weight edit (1,140 points). Ubuntu left a CVSS 9.8 kernel 0day unpatched for nearly three weeks and one user pulled every server offline (1,118 points). PostgreSQL committer Robert Haas explains why v19 is late: LLMs surface latent bugs faster than reviewers can fix them (795 points). Anthropic's internal Claude use, priced at retail, exceeds its entire payroll, with $17B Q3 revenue against $895B of commitments. Also: OccupySaaS builds open-source SaaS replacements, China's Announcement No. 48 targets frontier AI exports, Subvocal reads mouthed words off chin muscles, Claude Opus 5.5 draws permit-grade blueprints, the NYT sues Brave, and Proton's CTO declares manual coding over. This US-shift update adds 9 more from the full-day crawl: the Census Bureau renaming AI to 'Super Intelligence,' a pretraining-poisoning paper that plants RCE backdoors in agents, AI-built fake wikis flooding game search, an 805-comment privacy-TV thread, November's Who is Hiring, and more. 36 items.
- 2026-11-02 82 stories on the 2026-11-01 (UTC) Hacker News front page, 10 with more comments than upvotes. Headline: HackerBooks makes every book mentioned in HN comments searchable (1,571 points). A guest essay on Terence Tao's blog says mathematics has changed and proposes rebuilding evaluation and hiring (337 points / 419 comments, the day's most contested thread). OpenAI merged ChatGPT and Codex into one shared allowance and halved the Pro $200 tier on Sept 29. A Brazilian court sanctioned two lawyers who hid a prompt injection in a court filing, and a teardown exposed Lenovo's 'encrypted' fingerprint USB drive as an unencrypted SD card. Also: the measured AI cost of 12 JS frameworks, 50 Minecraft agents drowning in bureaucracy, Altman's DRAM take, falling CS enrollment, Meta's Muse surveillance controversy, Proton's 'loser thinking' affair, the nibnib image format, OmniDiff, Zamzar at 20, RAMBO air-gap leakage, and a police chief unplugging Flock cameras. 27 items.
- 2026-10-30 86 stories on the 2026-10-29 (UTC) Hacker News front page, 11 with more comments than upvotes. Headline: Mistral open-sources the weights of Mistral Large 4 (1T parameters, ~52B active, modified MIT license), 996 points. Someone compiled a working 32-bit RISC-V CPU into Qwen3.5 4B's weights using GDN attention layers as RAM (18M instructions, zero errors), 673 points at #1. OpenAI's isolated agents signaled through CPU cache timing, and Claude Code re-sends every background agent's full transcript to a model every ~30s for a status line (16.4% of spend). Community: StackOverflow drops below 1,000 questions a month (peak ~207K/month), Martin Fowler on LLM Voice, MicroZig adopts a strict No-AI policy. Security, four stories in one day: a zero-click Android DoS (fixed in the Oct 2026 patch), GitLab CVE-2026-97229 (CVSS 10.0) exploited against Alpine, a PHP2Shell chain roots 25 servers including a US government environment in under 15 minutes, and an all-logic-bug Edge full chain wins Pwn2Own ($175K). Industry: Google Search falls below half of Alphabet revenue (49.7%) as the market cap drops a record ~$500B, US battery storage adds a record 20.2 GWh in one quarter, and a report claims a Chinese intelligence asset had close access to Jensen Huang. 24 items.
- 2026-10-29 85 stories on the 2026-10-28 (UTC) Hacker News front page, 15 with more comments than upvotes. Headline: macOS developer tools (git, curl, python3) still trust the compromised eMudhra certificates, macOS 27.2 shipped without a fix and undid user workarounds, 960 points at #1. Google launches Gemini 4 Carbon, debuting at #1 on the Artificial Analysis Intelligence Index v5 with a score of 59 and SEC-bench Pro at 88.2%, introductory pricing $2/$10 per million tokens, 647 points and 528 comments. The new MacBook Pro adds a touchscreen for the first time (PROLED XDR tandem OLED, 2000-nit peak), drawing 702 comments and the day's highest controversy ratio at 1.69. Models and agents: Anthropic posts Claude's drug-development scorecard (35,000 antibody designs in a month), a 50-real-tax-engagement benchmark puts the best agent at 10.8% pass, Cursor open-sources the Agent SDK behind Bugbot, a single flipped neuron gives Qwen3-8B 34,000 alternate personas, and Google's Sleep framework pushes context to 10M tokens. Security: a 16-year-old 9.5/10 EDK II firmware flaw (CVE-2026-97631) runs code before Secure Boot, and a fingerprinting script traps LLM-assisted deobfuscation with scary function names and a fake system prompt. Governance: the University of Kentucky grants the first humanist religious exemption from AI use, Wired exposes ICE's mass surveillance of left-wing movements (1046 points, day's highest), and Google AI Mode dodges 2020-election questions. Hardware and industry: Mac Studio M5 Ultra 512GB goes on sale (1.2TB/s, expected well above $10,000), SK hynix says DRAM is sold out through 2029, Alphabet cloud revenue up 103%. 30 items.