AI Frontier Daily Briefing: 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.
89 stories on the 2026-11-06 (UTC) Hacker News front page; 11 had more comments than upvotes. 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), the day’s most-commented story. Against the selloff, METR’s assessment of Anthropic says AI has roughly doubled the pace of R&D, with about 80% of the acceleration attributed to AI automation itself; village.dev’s math adds a colder number for open weights: they win the price-performance crown at only 3% of budget points, down from 41% in June. On models, Google took Gemini Omni 1.2 Flash to general availability (15-second video, 30% cheaper) while Thinking Machines shipped Inkling 2 Small, a 594B-parameter MoE with only 17B active, promising Apache 2.0 weights next week. Tools and infra had a busy day: Perplexity put always-on Teammates agents behind its $200-per-month Max plan, Replit replaced its container sandbox with microVMs, OpenAI’s exploit pipeline scaled from 21 kernel LPEs to 948, Firecracker got a demonstrated guest-to-host escape, and Hightouch got frontier judging accuracy at roughly 4x lower cost by letting a small decision model ask Opus for help. Governance moved too: Congress is opening investigations into SpaceXAI, Meta and Apple, Huawei put a ‘100% domestic’ desktop into production, and a Dutch ethical bank locked GrapheneOS users out of payments. Plus Facebook’s contested $16B scam-ad figure, AI bots crowding university libraries into closing public access, an App Store census finding 9.8% of apps name-drop AI, and Blender’s UI extracted as a standalone library that benchmarks twice as fast as GPUI. The US shift adds 14 more: Britain moving to mandate salary disclosure in job ads (447 comments), the FAA replacing air-traffic floppy disks by 2028, an 820-domain fake-delivery skimming campaign, 15,579 infected websites with no way to notify their owners, pwn.college, sodium-ion batteries approaching LFP cost parity, Ubuntu 26.04’s sudo-rs breaking classic sudoers rules, the agent-drivable Rune UI framework, devvm’s per-project VMs, the Firenook emulator, Google taking over Crusoe’s Wyoming data center site, NVIDIA’s NVKE cluster control plane, Memorilla’s 16 memory tokens, and protesters entering Microsoft’s Amsterdam data center construction site. 38 items.
1. 329 comments on Anthropic IPO marks cut to as low as $1.1T
Who was quoting $2 trillion in October?
Bloomberg reported on Nov 6 that prospective buyers of Anthropic’s IPO are now discussing valuations as low as $1.1T; the story drew 137 upvotes against 329 comments, the day’s most-contested. Timeline: expectations ran as high as $2T in early October, investors had told underwriters Goldman Sachs and Morgan Stanley that $1.4 to $1.5T was fair, and after the election-day AI selloff some backers moved to $1.2 to $1.3T; the WSJ reported the same day that major investors are broadly skeptical of the $1.5T figure. Anthropic filed its S-1 on Oct 26 (ticker ANT), with H1 2026 revenue of $16.3B against an $80.7B net loss (about $72B of it non-cash); at a $75 to $100B raise it would be the largest IPO ever, and Reuters separately reported the AI backlash could delay the offering. All of these are press-reported negotiation marks, the S-1’s price range is still blank. Anyone reading primary-market signals, this is the first time buyers have openly pushed back on the sticker price. Bloomberg · WSJ · HN discussion
2. One day, $2.5T gone. Nasdaq fell 5.5% as the AI trade unwound
Ballots not certified, positions already cut.
Reuters reported the Nasdaq Composite fell 5.5% on Nov 5, wiping out about $2.5T in value, its biggest one-day drop since the April 2026 tariff shock. The trigger was the Democratic midterm sweep, on fears that AI infrastructure expansion gets slowed by regulation and permitting; Asian markets followed on Nov 6 with AI data-center names leading the decline. Context from yesterday’s edition: the Nasdaq 100 had already shed $2T in late October on doubts about AI capex returns, so this is the second trillion-dollar drawdown in two months. If you value AI companies off data-center revenue multiples, the denominator held while the numerator got cut. Reuters · HN discussion
3. AI R&D now runs 2x faster, and METR says AI itself is why
Extrapolate that slope and where does it end?
Ajeya Cotra published her read of METR’s public assessment of Anthropic on Nov 6 (35 upvotes, 34 comments). Her summary: capability progress from October 2025 to October 2026 ran about 2x the prior year’s pace, with the report attributing roughly 80% of the speedup to AI automation itself rather than more compute or headcount; she converts that into an implied labor uplift of roughly 4x to 9x, close to the ~12x figure Anthropic staff cite internally. Two caveats she flags: a majority of the supporting evidence is redacted from the public version, and if the acceleration compounds, doubling times could shrink from ~6 months toward ~3, with a naive extrapolation pointing at systems far beyond human level within a year. She notes she did not work on the assessment. For anyone building long-horizon capability forecasts, this is the first institutional evaluation that turns ‘the speedup comes from AI itself’ into a checkable number. Source · HN discussion
4. 15-second video at 30% off, Gemini Omni 1.2 Flash goes GA
Cheaper per second means the whole pipeline quote changes.
Google DeepMind announced Gemini Omni 1.2 Flash on Nov 5 (263 upvotes, 204 comments), now generally available. Clip length goes from 10 to 15 seconds, settable in 1-second steps from 3 to 15; 720p generation costs $0.07 per second, a 30% price drop from Omni 1.1. New: multi-turn conversational editing (swap backgrounds, lighting, aspect ratio without regenerating), reference inputs mixing character/product/style images, 3-second motion clips and audio cues, continuous camera moves (FPV flythroughs, zoom-outs, orbits) with fewer jump cuts. The API and AI Studio are open now, drafts can be upscaled to 1080p or 4K, and outputs carry an imperceptible SynthID watermark. If you build video tooling, the per-second cost line in your spreadsheet just moved. Source · HN discussion
5. Thinking Machines ships Inkling 2 Small, 594B total, 17B active
Weights go open next week. Is the wait worth it?
Thinking Machines Lab released Inkling 2 Small on Nov 5 (279 upvotes, 90 comments): a mixture-of-experts model with 594B total parameters and 17B active, taking image inputs with variable effort. Training specifics: the H1 pretraining phase ran 8.4T tokens over 68 hours on NVIDIA GB300 NVL72 racks in native NVFP4, later phases pushed context to 1M, and the joint RL run consumed 2,048 GPUs for 9 days. Scores: HLE-Diamond 28.7% (highest among the open-weight models listed), SWE-Bench Pro v2 Hard 86.3%, DeepSWE v1.1 72.1%, Artificial Analysis Intelligence Index 30.6; on safety, FORTRESS average risk fell to 16.2 and the internal misbehavior rate dropped from 66.8% to 14.9%. Weights are available on Tinker only for now, with an Apache 2.0 release promised for next week after external safety testing and red-teaming. If you self-host frontier weights, that is the one window worth watching this week. Source · HN discussion
6. Open-weight models win at 3% of price points, down from 41% in June
Cheap and strong now come from the same labs?
village.dev ran the numbers on its LLM Atlas (a capability score plus cost-per-task model), finding that open-weight models were the best buy at 41% of budget points in June 2026 but only 3% on Oct 31 (76 upvotes, 72 comments). Of 21 frontier price points, open models hold 5, all in the cheaper half (DeepSeek V4.1 Flash, Gemma 4 E4B, MiMo-V2.6-Flash and similar). Three explanations offered: closed labs distill their own flagships into cheap tiers in-house (Claude Haiku 5.5 shipped two weeks after Opus 5.5), they release whole model families that cover most price points, and their inference scale (Google on its own chips) serves tasks more cheaply. Robustness checks (10+ benchmark families required, budgets narrowed) show the same pattern. If you run a self-host vs API cost model, the 3% figure belongs in it. Source · HN discussion
7. 207,000 sessions mapped: how scientists actually use Claude
So what do scientists do with Claude all day?
The Anthropic Institute published a report on Nov 5 (82 upvotes, 34 comments) analyzing 207,000 research sessions on consumer plans from August to September 2026 (128,914 on Claude, 78,135 on Claude Code), plus a survey of 528 scientists. Data analysis dominates: 33.7% of Claude sessions and 48.1% of Claude Code sessions; writing and communication take 32.0% of Claude sessions, with manuscript drafting alone at 18.6%. The striking gap: ideation and hypothesis generation get only 4.2% of Claude sessions even though scientists rank it as the most important activity for discovery. On Claude Code, computer science is 32.7% of sessions versus a 4.3% global share of researchers. Scenario modeling: current capabilities plus wide adoption could free up 27% to 52% of research time and roughly double output in computer science; 40% of surveyed researchers expect AI to run full projects on its own by 2030. If you build research tooling, that 4.2% cell is the unmet demand. Source · HN discussion
8. Perplexity’s always-on Teammates agents hide behind $200/month Max
10,000 agents on the payroll, where do they sit?
Perplexity launched Teammates on Nov 5 (56 upvotes, 48 comments): persistent agents with assigned roles and memory that run 24/7 in the background, can hand work to each other, manage other agents, ping humans when a decision is needed, and integrate with Slack, each with its own browser, cookies and desktop. There is no separate price; access is bundled into the $200-per-month Max plan ($2,000 billed annually) with 10,000 Computer credits a month. CEO Aravind Srinivas says Perplexity deployed ~10,000 Teammates internally in 30 days, about 20x headcount, with roughly 1,000 active on a peak day and a median of 7 agent-hours per employee per week; the default model is Claude Opus 5.5, switchable per task. If you build agent products, ‘20x our headcount in internal agents’ is a ready-made pitch, and a compute-bill warning. Source · HN discussion
9. Replit rebuilt its sandbox on microVMs, 96.1% of CVEs avoided
Ten years of containers, back to VMs overnight?
Replit published the rebuild of its sandbox engine (18 upvotes, 6 comments): shared-kernel Linux containers were replaced wholesale with microVMs (Cloud Hypervisor plus KVM, the same lineage as Firecracker), each sandbox running a stripped custom kernel (tinyconfig, module loading disabled), migrated across hundreds of millions of projects without disruption. The drivers: container escapes via Linux kernel CVEs became a growing risk as AI accelerated vulnerability discovery, and per-VM kernels finally let tools like Docker run safely. The numbers: sandbox acquisition takes ~500ms; of 6,211 CVEs audited from June to November 2026, 96.1% were avoided (70.4% fixed before disclosure, 10.9% excluded by configuration); security updates hot-reload to all production sandboxes in under 15 minutes. The block-storage backend was their first fully AI-written component, later reworked into a sync/async hybrid. If you run an agent runtime, this is the most complete cost accounting yet in the containers-vs-microVM debate. Source · HN discussion
10. 948 kernel root exploits from one $1.2M OpenAI pipeline run
From 21 to 948 in four months, what’s the next print?
Phoronix reported that OpenAI’s cyber pipeline scaled its results to 948 validated Linux kernel local-privilege-escalation exploits, each with a candidate patch (32 upvotes, 16 comments on the day). The funnel: 2,308 candidate findings, 1,724 reproduced as real vulnerabilities, 948 usable root exploits, at roughly $1.2M in API spend per run, or about $1,265 per exploit-and-patch pair; OpenAI claims at least 10 previously unknown exploitation techniques in the set. Timeline: o3 found CVE-2025-37899 in May 2025, the Patch the Planet program launched with Trail of Bits and HackerOne in June 2026 with 8 pointer-leak PoCs and 24 LPEs, and four months later the count hit 948. Mind the framing: these are validated exploits with candidate patches, not fixes shipped into the kernel; human review remains mandatory and no exploit code was released. Yesterday’s edition logged the 21-LPE bill; today’s number is 45x that. Phoronix · HN discussion
11. Guest to host root: Firecracker’s full escape, demonstrated
If the sandbox base leaks, what holds the agents?
PWN.AI published PwnEscape Part 2 (93 upvotes, 52 comments): building on the Part 1 host-memory read/write primitive (CVE-2026-99999), the team hijacked a Firecracker VMM process’s vCPU thread from inside an untrusted guest, rewriting the saved register frame the host kernel stores during the KVM_RUN ioctl, so Firecracker executed the attacker’s write(2) on return to userspace, printed the 8-byte marker N8V23M2 to stderr, then resumed its normal path. The setup: Firecracker v1.16.1 on x86-64, host kernel 7.0.0-1011-aws; no host executables were modified and the VM stayed alive and reachable over SSH throughout. The key caveat: the demo ran Firecracker as root without the jailer; production containment (privilege drops, namespaces, seccomp) needs a separate assessment, and the root cause remains the Part 1 memory primitive; cut the address leak and the chain dies. If you run agent sandboxes on Firecracker, the jailer is not optional. Source · HN discussion
12. Ask Opus when the small model hesitates, 4x cheaper, more accurate
A judge that admits doubt beats a confident one?
Hightouch engineers (Albert Jin, Neil Desai, Delaney Mackenzie) wrote up an identity-resolution grading setup (130 upvotes, 36 comments) using Jev, TypeSafe AI’s decision model that outputs a 0 to 1 confidence score instead of generating text. Because Jev’s mistakes cluster near 50% confidence, they route only the 30% to 90% band (22% of cases) to Claude Opus 5.5 for a second opinion. Results on the hard test set: the hybrid hit 95.3%, edging out Opus alone at 95.1%, at $2.94 per 1,000 records versus $5.39 (about 45% cheaper); on the easy set it scored 98.8% versus 99.0% at $0.88 versus $3.86, roughly 4x cheaper, and runs about 4x faster (0.63s vs 2.2s per identity). The control, Luna (GPT-6’s decision model), was over-confident on hard cases and rarely escalated, so the hybrid barely helped. If you run LLM-as-judge pipelines, confidence-routed escalation to an expensive model is a pattern you can ship this week. Source · HN discussion
13. 164 upvotes for ‘I struggle to believe AI R&D is ingenious’
Would you put your name on its papers?
A LessWrong comment by Arjun Panickssery drew 164 upvotes and 172 comments, more comments than upvotes. After watching OpenAI’s Astra agent across several side projects, he concludes its judgment cannot carry real R&D: it fails to recognize the most important papers when reviewing prior work, runs experiments with zero information value (training volumes too small to distinguish architectures), and repeatedly drifts into trivial directions, in an Ultimate Tic-Tac-Toe project it surfaced a blog post containing a key idea without recognizing its importance, despite the author’s strong arena rank and the post’s recency as obvious signals. His exception: fast-feedback hill-climbing, where repeated prompting took a Scrabble move-listing SOTA 10x higher, though the two biggest breakthroughs came from his own suggestions. His closing jab: Astra’s reports are extremely hard to understand, requiring interrogation to establish what was actually done, and the temptation to just say ‘whatever’ and let it rip is real. If you treat agents as researchers, this is an operational antidote. Source · HN discussion
14. Hugging Face’s CEO puts his P(doom) ‘close to zero’
The real risk lives behind closed doors, not open weights?
Wired reported (34 upvotes, 18 comments) that Hugging Face CEO Clement Delangue, speaking at the Wired World Fair in Miami on Nov 5, put his P(doom) ‘close to zero’ and argued most AI risk is created by a handful of players building behind closed doors: agent misbehavior so far has come from big labs’ secret frontier testing, and future risks like bio-capabilities or recursive self-improvement are more likely to emerge from closed labs than open-source systems. He also disclosed a July 2026 rogue AI agent attack on Hugging Face linked to OpenAI agents; the company declined to pursue legal action, but he argued agentic cyberattacks should be illegal. Platform numbers: over 20 million AI builders and 50,000 sold Microduck robots; Nvidia announced a $12.93B acquisition of Hugging Face in September. Whatever you make of his stance, ‘the risk comes from closed-door testing’ is an argument that belongs in your security assessment checklist. Source · HN discussion
15. Cuban says Facebook knowingly runs his AI deepfake scam ads
Victims email him weekly, who takes the cut?
Mark Cuban told the Pivot podcast (Kara Swisher, Scott Galloway) that AI deepfake ads cloning his face and voice sell a $294 ‘diabetes cure’ he never endorsed, that elderly victims email him every week, and that Facebook, in his words the worst of the worst, knows users are being defrauded (487 upvotes, 222 comments). The figure traces to Reuters’ November 2025 report on leaked Meta documents: roughly $16B, about 10% of 2024 revenue, came from ads for scams plus prohibited goods, with a separate document putting clearly scam-like ‘higher risk’ ads at about $7B annualized; the automated systems allegedly required 95% confidence of fraud before banning an advertiser, and below that threshold suspected scammers paid higher rates and stayed live. Meta spokesman Andy Stone called the 10% figure ‘rough and overly-inclusive,’ citing a 58% drop in scam-ad complaints and 134M+ removals in 2025; the SEC is reportedly investigating Meta over fraud-linked ads. Framing matters: $16B is an allegation from leaked documents combining scams with banned goods, not an audited number, and Cuban’s specific ads were not independently verified. If you buy ads, that 95% threshold explains why scammers’ CPMs beat yours. Source · HN discussion
16. AI bots are swarming university libraries, public access at risk
A one-click bulk download exists, why don’t bots use it?
CBC reported on Nov 5 (119 upvotes, 53 comments) that libraries at MIT, the University of Michigan and other research universities say AI scraper bots now make up the overwhelming majority of their web traffic, forcing defensive measures that end up blocking legitimate readers and putting public access at risk. Project Gutenberg executive director Eric Hellman called the bots poorly programmed ‘monstropities’ destroying open access last year, this week he told CBC ‘everything’s gotten worse.’ The telling detail: Gutenberg is happy for its public-domain books to be used as training data and even offers a one-click download of the entire library, ‘but the bots are too stupid to use it,’ preferring to hammer the site at thousands of requests per second in every format. The dividing line is simple: good bots honor robots.txt (Google, the Internet Archive); bad ones ignore it and sweep every format. If your training or eval pipeline depends on the open web, the open web is actively closing on you. Source · HN discussion
17. Zalando’s broker lets agents act without your credentials
Your account, its hands, who’s liable?
Zalando open-sourced the Agentic Identity Broker on GitHub (18 upvotes, 3 comments), MIT-licensed, written in Go, 264 commits: agents get scoped, revocable access to services while users keep their own credentials. The broker sits in the call path via gateways like Agentgateway, handling OAuth ceremonies, token vaulting and translation centrally; a consent UI grants agents granular permissions across services like Google, GitHub, Databricks and Linear; privileged clients such as Slack or Google Chat bridges can start delegation chains without the user present; and sensitive actions like sending email get centralized, auditable tool approvals. The design emphasizes fail-closed behavior, log redaction and cryptographic validation; the team says it already runs in production for multiple internal MCP servers, with agent-to-agent token exchange on the roadmap. If you design agent permissions, this is a full generation above stuffing an API key into the agent. Source · HN discussion
18. CloneGuard sniffs out credential stealers before you run the repo
You’d execute a fake recruiter’s demo as-is?
Show HN: CloneGuard (tharun208, MIT, pure Go, zero dependencies), 24 upvotes and 6 comments, scans a project for credential-theft tricks before you run it: reads of SSH keys, AWS/GCP/Azure/K8s credentials, shell history, browser passwords and wallets; connections to raw IPs, Discord webhooks, Telegram bots and other drop sites; download-and-execute patterns like eval, new Function and curl-piped-to-sh; npm install hooks and VS Code tasks that fire on folder open; obfuscated identifiers, whitespace-hidden code, and compiled binaries checked into the repo. It only reads, never executes: files get classified, checked line-by-line with regex rules; combined signals in one file (reads keys plus contacts a suspicious server) escalate straight to critical, with exit codes 0 to 3 for git hooks and CI. The origin story is a real incident: a fake recruiter’s ‘demo app’ tried to steal SSH and AWS keys. The author’s own caveat: a clean result is not proof of safety, and pattern matching loses to a determined attacker. If your coding agent clones and runs unfamiliar repos, this is a cheap first gate. Source · HN discussion
19. Microsoft is handing back Oracle’s AI compute
With $190B a year of your own capex, why rent racks?
Reuters reported on Nov 6, citing people familiar with the matter (69 upvotes, 12 comments), that Microsoft is winding down its arrangement to rent AI cloud capacity from Oracle as part of a reassessment of billions of dollars in cloud partnerships. Background: Microsoft’s FY2026 capex guidance sits around $190B, concentrated on self-built data centers, and in June it reportedly walked away from $3B of Oracle compute talks because Oracle’s infrastructure missed the FedRAMP bar for federal workloads; Oracle’s response at the time called the partnership collaborative and fruitful. The signal is clear enough: hyperscale buyers are starting to pick suppliers using their own capacity plus compliance requirements. If your multi-cloud strategy leans on Oracle OCI, the renewal negotiation table just got reset. Source · HN discussion
20. Congress targets SpaceXAI, Meta and Apple, ‘lawyer up’
New majority, first subpoenas, who bleeds first?
The Washington Post reported on Nov 6 (105 upvotes, 97 comments) that after the Democratic midterm sweep, Congress is preparing an investigative blitz with SpaceXAI, Meta and Apple among the first targets. On the record: the incoming House Oversight chairman’s advice to ‘lawyer up,’ and Senator Ed Markey’s line about planting new forests for all the subpoenas they’ll be printing. Named threads: the Trump White House allegedly using Grok to plan the kidnapping of Venezuelan president Maduro; Meta requiring NDAs from local governments before data-center talks; and Apple’s relationship to tracking apps used by ICE. The background numbers: Democrats took 240+ House seats and 54+ Senate seats, the new Congress convenes January 3, 2027, and government funding expires December 11, with warnings from Senators Sanders and Markey already knocking the stocks. If you sell AI into the big labs or handle their government relations, subpoena risk belongs in next budget season’s plan. Source · HN discussion
21. Huawei ships a ‘100% domestic’ desktop with no Intel or Qualcomm
Sanctioned out of everything, it closed the loop anyway?
SCMP reported (41 upvotes, 45 comments) that Huawei has put its first ‘100% domestic’ desktop into production: no Intel, no Qualcomm. This is the Qingyun HM650 series announced earlier: a self-developed Kirin X90 chip, HarmonyOS for PC, dual-layer cooling, aimed at office work, engineering design, simulation modeling and programming, presented by Zhu Dongdong, president of Huawei’s tablet and PC product line; the HarmonyOS computer ecosystem now counts 19,000+ apps. It is not Huawei’s first domestic-chip PC, Qingyun W515/W525 government models have been on sale since roughly 2,250 yuan, but this one comes with a whole-chain-domestic claim and a consumer posture. For calibration: Huawei announced in September that cumulative Kirin shipments passed 100 million. If you track China’s domestic-substitution market or supply-chain evolution under sanctions, a full PC with no US parts went from slogan to SKU. Source · HN discussion
22. 602 upvotes as a Dutch ethical bank locks out GrapheneOS users
Your own money or your privacy-respecting phone, pick one?
Issue #105 in the PrivSec-dev banking-compat report was reopened on Nov 6 and hit the front page at No. 2 with 602 upvotes and 394 comments: Dutch bank ASN’s payment flow now blocks non-Google-certified Android, leaving GrapheneOS users stuck at the Wero step (Wero is the European payment system run by EPI), on Pixel 6 Pro and other devices. EPI’s stated reason: hardware key attestation ‘does not verify the integrity of the calling application and the runtime operating system.’ Commenters push back line by line: GrapheneOS’s hardware attestation certificate carries verifiedBootState, boot key hashes and app identity, while Play Integrity amounts to one SDK call to Google returning an opaque verdict. The irony writes itself: ASN positions itself as the ethical bank while a European payment system outsources its integrity judgment to an American company, and users are already organizing complaints via Bits of Freedom. If you run banking or payment apps on a de-Googled phone, this issue is your coming attraction. Source · HN discussion
23. 9.8% of App Store apps mention AI, per a 178K-app census
Same ratings, so is AI a feature or a sticker?
Obelisk Club published an App Store census (59 upvotes, 52 comments) covering 178,637 apps across 173 regions on the charts since October 9, 2026, plus 165,735 US reviews of 3,912 apps. Headline: 9.8% (17,456 apps) mention AI in their name or description, but the distribution is lopsided: 43% in Graphics & Design versus 2.6% in Navigation, and the ratings gap is negligible (4.49 vs 4.4 stars). Monetization: 26.7% of free-to-download apps advertise in-app purchases or subscriptions, from 60.2% in Graphics & Design down to 11.5% in Shopping; 67.3% of charting apps come from indie developers, rating-for-rating equal with big publishers (4.41 each). Abandonment: 20.3% haven’t been updated in over a year and 9.8% in over three, worst in Books (36.3%) and Reference (33.6%). The top complaint in 1- and 2-star reviews: ‘subscription & price’ at 12.7%. Framing caveat: a keyword proves AI is in the marketing, not that the features depend on it. If you ship indie apps or judge AI wrappers, 9.8% is the baseline, not the ceiling. Source · HN discussion
24. Twice GPUI? Blender’s UI, extracted and benchmarked
Everyone mocks Blender’s UX, but the code bench-tests?
BUI, the day’s No. 1 story with 641 upvotes and 261 comments, is ppmpreetham’s extraction of Blender’s interface system into a standalone UI library: widget drawing, layout, input rules, themes, search and expression behavior, damage-based redraw, all in a small C runtime with bindings for Python, Rust, JS/React/Svelte, Dear ImGui and more, billed as ‘twice as good as GPUI.’ Measured on a Ryzen 7 5800U: startup at ~40ms CPU versus GPUI’s ~231ms, RAM at ~15MB versus ~29MB, and in a 3,000-cell animation test 59.9 FPS versus GPUI-fast’s 31.5. The default renderer is a CPU software raster with damage tracking, idle CPU drops to zero and hovers repaint only affected rectangles, with an opt-in GPU path (D3D 11, OpenGL 3.3) that produces identical pixels. The cost list is honest too: no accessibility, no IME, no RTL shaping yet, and Cocoa remains experimental. If you are choosing a UI framework for a desktop tool, the thirty-year-old Blender codebase just posted a score worth reading. Source · HN discussion
25. Britain will make employers print salary numbers in job ads
No more “competitive salary, DOE” dodge?
Reuters reported on Nov 5 (589 upvotes, 447 comments) that Prime Minister Andy Burnham announced mandatory salary disclosure in job ads, arguing that just over two in five employers always publish a range and fewer than one in five always give a specific figure (CIPD estimates cited by his office), while about one million British 16- to 24-year-olds, 13%, are out of work, education or training, one of Europe’s highest rates. The consultation ran July 14 to October 27, 2026, with no company-size threshold in the draft; roles that are not advertised would require salary information in writing before interview, and no implementation date is set yet. For context, the EU Pay Transparency Directive took effect on June 7, so ranges in continental job ads are already the norm. If you hire in Europe, the “salary negotiable” era is ending. Source · HN discussion
26. Floppy disks finally leave US air traffic control, by 2028
My flight weather rode a diskette until now?
The FAA said on Nov 5 (347 upvotes, 234 comments) it will finish installing the Enterprise Information Display System at 435 facilities by the end of 2028, replacing four legacy systems from the 1980s that handle weather, surface observations, airport and airspace status, and en-route flight information; the four operate as isolated silos whose parts are so scarce the agency has bought replacements on eBay. The first five sites are Atlantic City, Nashville and Rochester airports plus data hubs in Atlanta and Salt Lake City. The money behind it: Congress approved $12.5B for ATC modernization in 2025 and Transportation Secretary Sean Duffy is asking for $30B more, while RTX and Indra are separately contracted to replace 612 radar systems by summer 2028. If you maintain legacy infrastructure, this is a textbook budget case for replacing it. Source · HN discussion
27. 820 fake delivery sites pair real menus with a fake checkout
That sponsored “official” link deserves a second look?
InfoHawk published its Scamwich findings on Nov 5 (138 upvotes, 71 comments): 820 domains impersonate 35 restaurant brands, 808 of them on the .delivery TLD; Jimmy John’s alone accounts for 404, Firehouse Subs 121, Jersey Mike’s 80. Each clone is a reverse proxy pulling live pages from the real restaurant, so menus, prices and carts are genuine; only the checkout is swapped for a form harvesting card number, expiry, CVV and the bank’s SMS verification code, which an operator appears to use in real time to push charges through 3-D Secure. Distribution ran through Google Search ads, at least 47 of them between July 7 and October 20 across 46+ advertiser accounts, and 367 domains were still live on Nov 4. If you order from search results, check the domain before you pay. Source · HN discussion
28. 15,579 websites are infected and nobody can notify the owners
Knowing is the easy part?
Scott Helme of Report URI wrote (159 upvotes, 136 comments) that his firm’s threat-intelligence crawls have confirmed 15,579 websites currently serving malicious JavaScript (the true count is higher since cleaned sites leave the list), with a combined 135M+ visits per month. The injections range from card skimmers and ClickFix lures to iOS spyware (DarkSword, Coruna), entering through WordPress flaws, third-party script supply chains and “DeadTag” domains re-registered by attackers but still loaded by sites. The problem is notification: WHOIS contact data went mostly useless after GDPR, CDNs and proxies hide the real host, and publishing the list would hand attackers a target inventory; in his words they find infections “considerably faster than we could ever hope to notify” the owners, and he is asking national CERTs for help. If you run websites, auditing your third-party script supply chain beats waiting for a notification that may never come. Source · HN discussion
29. pwn.college is ASU’s free dojo for learning to break software
White belt to blue belt, beats grinding LeetCode?
pwn.college hit the front page (319 upvotes, 54 comments): a free, open-source cybersecurity curriculum maintained by Arizona State University, organized as dojos where learners progress from white belt through orange, yellow, green and blue belts via hands-on challenges in exploitation, web and crypto, backed by YouTube lectures and livestreams. Educators can import challenges or run private dojos for their own classes, and the platform already powers much of ASU’s security curriculum; the one hard rule is no public writeups. Graduates get pointed to CTFs, wargames and ASU research internships. If you are breaking into security or training a team, this is a mature course infrastructure you can use today. Source · HN discussion
30. Sodium-ion batteries close in on LFP at $229 per kWh
Time to put sodium in the data-center backup quote?
pv magazine reported on Nov 6 (200 upvotes, 93 comments): Wood Mackenzie puts sodium-ion’s current global average at about $229 per kWh versus roughly $200 for LFP, with a 21% lower total cost of ownership over 20 years thanks to no capacity-expansion cycles and lower operating costs, stable across 2-to-4-hour storage durations. Supply agreements already top 70 GWh across China, the US and Europe; BYD is targeting 30 GWh of sodium-ion manufacturing capacity and CATL aims to ship over 1 GWh by end-2026. Timeline: cell-level cost parity with LFP arrives by late 2026 for leading makers, system-level parity only after 2030. The pitch is raw-material security and a wider operating temperature range, so cold climates need less thermal management than lithium. If you spec energy storage or data-center backup power, start asking for sodium quotes for 2027. Source · HN discussion
31. Ubuntu 26.04’s Rust sudo deliberately breaks classic sudoers rules
Sold as a drop-in, minus the safety features?
Chris Siebenmann of the University of Toronto wrote (188 upvotes, 165 comments) that Ubuntu 26.04’s sudo-rs deliberately does not support classic sudo’s bracket wildcards: a rule like /usr/sbin/exim -bp[acru], which allows exactly four argument combinations without a password, fails under sudo-rs, which keeps only the broad * wildcard, and the limitation sits at the end of a doc section describing functionality that does not exist. Real casualties already: Dokku Postgres and Xymon both hit “wildcards are not allowed in command arguments,” and Ansible become can hang. The escape hatch is sudo update-alternatives —set sudo /usr/bin/sudo.ws to return to the C implementation. Siebenmann’s verdict, “we aren’t Canonical’s free debugging resource”, came with action: he has switched all his own 26.04 machines back. If you administer servers, run your sudoers files through regression tests before upgrading. Source · HN discussion
32. Rune, a GPU-rendered Rust UI framework built for agents to drive
Even the issue filing gets delegated to an agent?
Rune, on GitHub (172 upvotes, 125 comments), is anomalyco’s native GPU-rendered Rust UI framework with its own renderer, text engine, layout and spring animation, targeting macOS, Windows, Linux and the web. The distinctive choice is headless-first: every app can run headless so an agent can render it to PNG, dump the element tree and drive it with scripted input; the contribution policy matches, with outside PRs closed and users asked to let their agents open issues, which the maintainers say they answer quickly. It is explicitly pre-release, with dependencies pinned to a rev on the dev branch, at roughly 1,640 commits. If you build agent-native apps or desktop tools, “UI as an agent-drivable surface” is an idea worth borrowing. Source · HN discussion
33. devvm gives every project its own VM, coding agents included
The era of hoping your agent behaves is over?
Show HN: devvm (169 upvotes, 64 comments), MIT-licensed: every project folder gets an isolated VM with an auto-detected toolchain (package.json gets Node, pyproject.toml gets Python), built on Lima (Apple’s Virtualization.framework on macOS, QEMU on Linux). Sensitive paths like ~/.ssh, ~/.aws and ~/Documents stay on the host unmounted; networking defaults to an allowlist routed through a host proxy, and agent tokens are redacted inside the VM while the proxy injects real credentials. Claude Code, Codex and opencode can run inside the VM, ports auto-forward, and once the base image is built, new VMs start in seconds. If you let agents run unfamiliar code, this is hard isolation instead of hope. Source · HN discussion
34. Firebase emulators reimplemented with 40x faster writes
How many years of 30-second Java emulator starts?
Show HN: Firenook (170 upvotes, 30 comments), an Apache 2.0 drop-in replacement for the Firebase emulators: change one word in the start script and it runs the same firebase.json, flags and ports. Measured on a real app (211,205 Firestore docs, 6.7GB Storage, 2 hours of replayed production traffic): median writes 208ms versus 5.1ms (40x), peak memory 22.6GiB versus 2.11GiB (~10x), restart with data 30.8s versus 0.59s (52x); it answered 25 of 26 query shapes faster. It emulates eight services including Firestore, Auth, Storage and Functions, requires no Java, and firenook mcp exposes 54 tools to coding agents. If you develop against Firebase locally, this is a direct swap. Source · HN discussion
35. Google takes over Crusoe’s Wyoming data center site
6GW signed, and pre-development itself is now a business?
Crusoe announced on Nov 6 (85 upvotes, 33 comments) that Google has acquired its southeast Wyoming data center campus site; Crusoe handled site selection, government approvals and power strategy through pre-development before handing off. The campus was initiated jointly with Tallgrass Energy around a hybrid power plan of fuel cells, grid power and gas turbines, with a path to carbon capture hooked into Tallgrass’s regional CO2 sequestration hub. Crusoe’s books: more than 6GW contracted and over 40GW in the pipeline, and it is separately developing Google’s campus outside Amarillo, Texas. If you track the AI power chain, “energy developers de-risk the site, the cloud buys it ready” is becoming a business of its own. Source · HN discussion
36. NVIDIA ships NVKE, a control plane for bare-metal GPU clusters
One kubectl per tenant cluster?
NVKE (NVIDIA Kubernetes Engine) on GitHub (123 upvotes, 27 comments) is NVIDIA’s management plane for GPU Kubernetes clusters on bare metal: a throwaway seed cluster stands up a management cluster, which then provisions isolated workload clusters, each with its own control plane, etcd and node pools. Security design includes TPM-based SPIFFE/SPIRE attestation before nodes receive join credentials and optional per-cluster etcd encryption; it follows the Cluster API pattern under Apache 2.0, and the local quickstart needs 6 CPUs, 12GiB of memory and 100GiB of disk. Caveats: it is a preview, the API is unauthenticated plain HTTP by default after install, and the README says not to run it in production. If you run GPU infrastructure, the quickstart is worth an afternoon. Source · HN discussion
37. Memorilla fits a whole document library into 16 memory tokens
Short on context? Stop carrying the raw text.
Memorilla, a Stanford-led paper (arXiv 2610.21375, 108 upvotes, 13 comments), gives frozen LLMs latent semantic memory: documents are encoded once into reusable embeddings, and the model learns to retrieve query-specific information and place it in the input embedding space, using only 16 memory tokens. It beats Mem0 and MemoRAG on 5 of 8 benchmarks spanning personalization, long-document QA and knowledge-intensive QA; on FactKG it scores 84.33% versus 80.43% for full-context inference at 1/108th the tokens, and in TextWorld agentic continual learning average success rose from 18.6% to 24.2%. If you work on long-context or memory systems, this is hard evidence for compressing retrieval into embeddings instead of text. Source · HN discussion
38. 100 protesters storm Microsoft’s Amsterdam data center site
AI’s power bill just reached the doorstep?
The Verge reported on Nov 6 (39 upvotes, 37 comments) that roughly 100 protesters entered the construction site of Microsoft’s new Amsterdam data center; some scaled a crane to hang a “fuck AI” banner while others unfurled “Microsoft powers genocide,” organized by the local action group Geef Tegengas against three planned 85-meter towers. Context: Amsterdam halted new data center permits at the end of 2023 and the ban now runs to 2030, but this project was planned back in 2016 and pushed through loopholes, with Microsoft as sole tenant and Britain’s Pure Data Centres building it. The same week activists staged a mock Israeli military checkpoint outside Microsoft’s Dublin data center, and last month Microsoft abandoned its largest UK data center project after grid officials said power would not arrive until the 2030s. If you site data centers or sell AI infrastructure, community consent is now a schedule risk, not a PR issue. Source · HN discussion