2026-11-11

Daily AI Frontier Briefing: 2026-11-11

Hacker News front page for 2026-11-10 (UTC), 86 stories, 16 with more comments than upvotes. Models: day two of the Fable 5.2 thread, 984 comments fighting over whether you still need to understand your own code (860 pts); Google launched Gemini 4 Carbon at $2/$10 per million tokens, doubling after Nov 30 (259 pts/141 comments); Amazon detailed replacing Sonnet 4.6 with the 4B Strands Decider, cutting cost per ticket from $0.0222 to $0.0005, 44x cheaper (672 pts/233 comments); Fable 5.2 built a 12-level beat 'em up overnight on ~340M tokens (472 pts/424 comments); DeepMind's AlphaProof Code formally verified Signal's protocol implementation with 6M lines of Lean in four weeks (532 pts); Claude shipped Whiteboards, Sheets, and Data (107 pts); Chrome teams open-sourced modern-web-guidance, 250 modern web features packaged for coding agents (167 pts). Industry: Anthropic opened its IPO roadshow at $70 to $78 per share for 825M shares under ticker ANT (58/79 and 29/49), with WSJ reporting a $1.3T target valuation; CoreWeave's 2032 bonds yielded above 16% and CDS topped 1,200 bps (191 pts/182 comments); an OpenAI–Upwork study says writing freelance postings fell ~31% in a year; ClickHouse acquired Xata (96 pts); Val Town announced it is winding down, open-sourcing Feb 2027 (216 pts). Engineering: a former Meta engineer argues AI usage generates fragile codebases (276 pts/343 comments); Expo open-sourced the Exact cross-platform framework (131 pts/106 comments); Blender 5.3 adds native Gaussian Splat support (361 pts); HolyArrs replaces six *arr apps in a 14.9 MB binary (161 pts); Zen 6 deep-dive shows ~21% single-thread gains over Zen 5 (130 pts); a Lean line-by-line formalization of CPython covers 14.3% with 613K lines proved (123 pts); Starlink Mini 2 arrives at $999 with a year of service (175 pts/184 comments); Synaptics stopped an open-source DisplayLink driver with 125 patents and a $25K weekly claim (190 pts); an 11-rule smart home stack that survives without internet (501 pts); deleting all on-disk SSH keys for hardware-backed credentials (190 pts); TV broadcast signals sensing motion through walls with an SDR (256 pts); the Fireworks AI breach touched Notion AI chats (157 pts); China drafts lifetime ID codes for every robot made domestically (19 pts). 25 items.

The HN front page for 2026-11-10 (UTC) had 86 stories, 16 of them with more comments than upvotes. Fable 5.2, released the day before, still owned the top spot on day two, with 984 comments arguing not about specs but about whether you still need to understand the code you ship. Google launched Gemini 4 Carbon the same day, and Amazon published a full cost ledger for swapping Sonnet 4.6 out of a massive internal pipeline. Anthropic’s IPO roadshow opened and CoreWeave’s bond yields crossed 16%. Tools and security had a heavy day too. 25 items.

1. Fable 5.2 day two: 984 comments on understanding your own code

Ship it without understanding, then who fixes the break?

The discussion thread for Anthropic’s Fable 5.2 stayed pinned all day, 860 points, 984 comments. The fight is not about pricing: 1. Multiple users say they can no longer tell Fable 5.2 apart from Opus 5.5 on routine coding, while the max-output truncations and slowness that plagued 5.1 are gone. 2. The central argument is about understanding, one camp says productivity belongs to whoever lets go of understanding fastest, the other cites Bainbridge’s 1983 Ironies of Automation: the more automated the system, the more the human operator’s judgment matters. 3. The compiler analogy (‘nobody reads generated assembly anymore’) got pushed back on: compilers are deterministic, models are not, and generated code has no ground truth to check against. If your team ships agent-written production code, read this thread in full before setting your review bar. HN discussion

2. Gemini 4 Carbon is out at $2/$10 per million tokens, for now

One more contender on the bench, then?

Google launched Gemini 4 Carbon on November 10, positioned as its strongest model for cybersecurity and long-horizon agentic work, 259 points, 141 comments. 1. Through November 30 it stays at Argon’s price, $2 per million input tokens, $10 per million output, then doubles to $4/$20. 2. Scores: first place on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench; on the safety side it blocks 99% of simulated attack trajectories, scores 0% prompt-injection success on Gray Swan IPI, and Trajectory Labs found zero jailbreaks across 132 hours and over 13,000 requests. 3. Carbon is the model that powered AlphaProof Code’s formal verification of the Signal protocol. If you run agent workflows, benchmark it at the introductory price before deciding whether to lock in volume before the increase. Post · HN discussion

3. Amazon replaced Sonnet 4.6 with a 4B decider model, 44x cheaper

It’s just labeling tickets, did it ever need a big model?

Amazon engineers published on November 9 how one of the world’s largest internal ticketing systems swapped Claude Sonnet 4.6 for Strands Decider E4B, a Gemma 4-based small model, 672 points, 233 comments. Decider generates no free text; it outputs a probability and confidence per given option, at about 640 ms per call. The ledger: 1. Cost, per ticket dropped from about $0.0222 to about $0.0005, roughly 44x cheaper; a million tickets went from ~$22,200 to ~$500. 2. Latency, average call fell from 4,215 ms to 897 ms, output tokens from 170 to 1. 3. Rollback, a three-way AppConfig switch (BEDROCK/DUAL/DECIDER) let them shadow live traffic, and reverting was a 45-second config change. The authors state the boundary clearly: closed-choice tasks with a stable catalog, no explanations needed, context under a few thousand tokens. If you have labeling, routing, or triage workloads, this is a ready-made playbook; the repo is open source at strands-labs/strands-decider. Post · HN discussion

4. A 12-level action game, built by Fable 5.2 overnight

The nephew’s dream, subcontracted just like that?

The author of notesbylex asked Fable 5.2 to build the game his nephew had been dreaming of: he sent the prompt before bed and woke up to a 12-level side-scrolling beat ‘em up, 472 points, 424 comments. The finished game has 18 unlockable characters, 12 bosses modeled on the kid’s drawings, cutscenes, and 4-player co-op; the model wrote a bot that beat all 12 levels in a headless browser and checked screenshots for problems. The bill: about 600 Claude requests and ~340M tokens, ~500 images from OpenAI, 11 music tracks and ~100 sound effects from ElevenLabs, playable at aidens-revenge.pages.dev. The author admits a June attempt with Opus 4.8 and Fable 5 produced something ‘not worth playing for more than a couple of minutes,’ and since he never tested Opus 5.5 he cannot say one-shotting was exclusive to 5.2. If you want to know how far ‘one shot’ is from your own project, this invoice is more useful than a benchmark score. Post · HN discussion

5. AlphaProof Code proves Signal’s crypto in 4 weeks, 6M Lean lines

Great, when does my codebase get its turn?

Google DeepMind introduced AlphaProof Code on November 10, 532 points, 169 comments. The process in three steps: 1. Researchers and models translate security goals into formal Lean specifications, each reviewed by a human. 2. Agents powered by Gemini Carbon generate proofs, over six million lines of Lean in total. 3. Agents audit the proofs and flag only about 1% of the codebase for human review; Lean machine-checks the rest. The target was not a protocol design document but Signal’s actual compiled implementation; as a validation, the team reintroduced two previously fixed real bugs and the system produced machine-checked proofs that both versions were insecure. DeepMind calls this the first formal verification of a widely deployed cryptographic protocol implementation at this depth, and the artifacts are open-sourced at google-deepmind.github.io/code-verification. For high-assurance software, this is the first large-scale demonstration of the ‘AI writes proofs, humans set goals’ division of labor. Post · HN discussion

6. Claude adds Whiteboards, Sheets, and Data in one push

Whose job description is this eating?

Anthropic shipped three products on November 10, 107 points, 81 comments. 1. Claude Whiteboards (beta): a shared board where teams brainstorm with Claude in real time, with box-select questioning and export to Lucid and Miro. 2. Claude Sheets (beta): spreadsheets with tabs, formulas, and dropdowns from a plain prompt, with each edit explained, exportable to .xlsx/CSV. 3. Claude Data (generally available, formerly Dashboards): connects to Redshift, BigQuery, and ClickHouse, adding certified-query badges, data stories, and live sharing for viewers without connector access. All three are on paid plans; Team plans get everything by default, on Enterprise Whiteboards and Sheets default off while Claude Data turns on by default November 19. If Claude is already a team tool for you, it is worth testing whether some of your spreadsheet and BI work moves over. Post · HN discussion

7. Chrome’s teams packaged 250 modern web features for coding agents

Someone’s finally grading the agents’ legacy JavaScript?

The GoogleChrome repo released modern-web-guidance, 167 points, 46 comments. The premise: coding agents produce legacy code because training data skews old, and they generate bloated JavaScript for tasks the platform now handles natively. Contents: 250 modern web features (140 CSS and layout, 36 HTML and DOM, 74 JavaScript and APIs) and 203 real-world use-case guides, written for token efficiency and pruned by evals, covering View Transitions, container queries, the Popover API, Trusted Types, and passkeys. It is stewarded by the Chrome, Edge, Firefox, and MDN teams, installs with npx modern-web-guidance@latest install, and supports Claude, Cursor, Codex, and other agent formats. If you write frontend with agents, install it and run its legacy-scan once. Repo · HN discussion

8. Anthropic’s IPO roadshow opens at $70 to $78, 825M shares, ticker ANT

$1.3 trillion, would you take that price?

Two stories landed the same day: the WSJ reports Anthropic is seeking a $1.3 trillion valuation, 58 points, 79 comments, and the company itself posted its IPO roadshow notice, 29 points, 49 comments. Prospectus highlights: 825 million Class A shares, about 801.4 million from Anthropic and 23.6 million from existing stockholders; an expected range of $70.00 to $78.00 per share; underwriters get a 30-day option for up to 123.75 million more shares; planned Nasdaq listing under the symbol ANT, led by Goldman Sachs and Morgan Stanley. The SEC registration has not taken effect, so no orders are being taken yet. The $1.3T figure drew heavy skepticism in the thread, with commenters comparing it against the largest IPOs in history. If you track this story, the registration going effective and the final pricing range are the only two events that matter. Post · HN discussion

9. CoreWeave’s bond yield tops 16% as AI credit stress rises

The spread says what the equity won’t?

CNBC reported on November 6 that CoreWeave’s 2032 bonds, coupon 9.625%, now yield above 16%, 191 points, 182 comments. Its five-year CDS spiked to 1,289 basis points (Deutsche Bank data) before easing to about 1,155 on Friday; for comparison Nvidia sits near 139 and Oracle near 298. The backdrop: 1. Five AI bond deals totaling $67B priced since October 29, bringing 2026 hyperscaler issuance to $296B; Goldman forecasts $440B next year, and BofA projects AI-related credit issuance up 31% to $685B in 2027. 2. Goldman notes hyperscalers collectively reported negative free cash flow in Q3, capex exceeded operating cash flow. 3. Jefferies’ Christopher Wood: the most plausible end to the arms race is ‘the credit card being taken away.’ If you allocate along the AI infrastructure chain, CDS spreads speak earlier than stock prices. Post · HN discussion

10. Writing freelance postings fell ~31% in a year, says an OpenAI study

Is the affected list longer or shorter than you thought?

OpenAI published a study with Upwork, Work at the Frontier, on AI’s effect on the freelance market, 24 points, 17 comments. Figures as read from its charts by commenters: content-writing postings dropped 0.57 percentage points to 1.27% share, roughly a 31% year-over-year decline; writing, illustration, and tech support top the list of categories people now ‘handle themselves with AI’; web development moved the most in dollar terms but fell only about 2%. Additional observations from the thread: freelancers increasingly get hired to clean up AI output rather than build from scratch, and one commenter claims AI cleanup work starts struggling past roughly 50,000 lines of code (a contested claim). The report has no executive summary and does not answer whether the overall freelance pie is growing or shrinking. If you take platform gigs, find your category on that chart. Post · HN discussion

11. ClickHouse acquires Xata, and managed Postgres goes agentic

Database M&A, now coming for branching and scale-to-zero?

Xata announced it is joining ClickHouse, 96 points, 9 comments, the same day ClickHouse Managed Postgres hit general availability. The two are complementary: ClickHouse serves high-throughput production workloads while Xata handles large fleets of small, short-lived databases with instant branching and scale-to-zero compute, both running unmodified vanilla Postgres. The technology moving over: the Xatastor storage engine (ZFS zvols over NVMe/TCP, copy-on-write) enabling ~3-second branches and ~1-second wake-ups, plus the open-source projects pgstream (logical replication with anonymization) and pgroll (reversible schema changes). Nothing changes for existing Xata customers in the short term; the plan is feature parity inside ClickHouse Managed Postgres, then migration. Terms were not disclosed. If you use Postgres branches for previews or multi-tenancy, watch the planned agentic tier. Post · HN discussion

12. Val Town is winding down after raising $7M

Is there really no market for end-user programming?

Val Town, a social code-hosting platform, posted its wind-down notice, 216 points, 55 comments. Key dates: Val Town’s code goes open source on February 1, 2027; the thread cites total prior funding of $7M. Commenters place it in the end-user programming graveyard: Eve raised $2M and Darklang $3.5M, and neither made it either. One commenter blames the Deno runtime, its team is joining Cloudflare and will stop maintaining it next year (a commenter’s claim, not a company statement); others argue this category never needed VC-scale money in the first place. If you host vals on Val Town, plan your migration now rather than waiting for the open-source drop. Post · HN discussion

13. ‘AI usage generates a fragile codebase’ hits a nerve: 343 comments

Output is up, but who pays the debt?

Former Meta engineer Evan X. Merz argues that heavy AI coding trades resilience for short-term output, 276 points, 343 comments, the day’s most contested thread. Four claims: 1. Test inflation, agents produce piles of trivial tests that duplicate existing e2e paths, so coverage metrics look fine while important cases stay untested. 2. Over-engineering, extra helpers and rigid DRY application make debugging and comprehension harder. 3. Eroding human understanding, reviewers can no longer recognize bad abstractions. 4. Speed breaks review, faster PRs get skimmed, and small defects accumulate. He says over 90% of his code at Meta was AI-generated near the end, yet he saw no 2x or 10x productivity because he held the quality line, and he watched P0/SEV incidents drift from monthly rarities toward routine weekend outages. He acknowledges he was laid off from Meta in May, partly, he suspects, for being less enthusiastic about AI coding than his peers. To test his claims against your own reality, count your incident tickets from the last three months. Post · HN discussion

14. Expo’s Exact framework, write once, run on web, macOS, iOS, Linux

‘Agent-native’, noted, and we’ll see?

Expo open-sourced Exact, 131 points, 106 comments, a cross-platform application runtime. The structure: 1. UI is described in a declarative language called Contract, data logic in TypeScript or Rust; the compiler emits a compact plan that becomes a ~20 KB ES-module runtime on the web driving real DOM and CSS, while native execution runs on a Rust runner with a patched Taffy layout kernel, AppKit/CoreText on macOS, vello GPU rendering on Linux. 2. No webview and no JavaScript bridge; TypeScript compiles to Hermes bytecode and loads after first paint. 3. Agent-native is the headline: ten driver operations (tree, screenshot, tap, type, state, and friends) behave identically on every host, with a seekable clock to eliminate test-timing flakes. Speed budgets are enforced per commit: 100 ms p50 cold start, 100 ms edit-to-present. It is pre-1.0 and changes daily, Android is early, and Windows remains unfinished. If you want agents writing cross-platform UI, this is a substrate worth cloning. Repo · HN discussion

15. Blender 5.3 ships with native Gaussian Splat support

Does the 3D-reconstruction pipeline go mainstream now?

Blender 5.3 was released on November 10, 361 points, 51 comments. The headline changes: 1. Native 3D Gaussian Splats, import from PLY, SPZ, and glTF, stored as point clouds, renderable in EEVEE, Cycles, and Workbench, editable with Geometry Nodes. 2. Cycles gains Principled BSDF dispersion (OpenPBR v1.1.1) and NVIDIA DLSS 4.5 ray reconstruction denoising (driver 620+). 3. EEVEE switches to Vulkan by default, with new Material Lighting nodes and panoramic/fisheye cameras. 4. Stacked scene effects in the compositor, a rewritten Blade tool and OpenTimelineIO support in the video sequencer. If you run a capture-to-delivery pipeline for splats, native import removes one conversion step. Post · HN discussion

16. HolyArrs replaces six *arr apps with one 14.9 MB binary

Six ports, six logins, somebody finally got annoyed?

HolyArrs, a self-hosted media manager, hit the front page at 161 points, 129 comments. The pitch: one app replacing Sonarr, Radarr, Lidarr, Bazarr, Prowlarr, and Seerr/Overseerr. The numbers: 1. Size, a 14.9 MB download versus 718.2 MB for the six-app stack, a ~48x gap by the author’s measurements (November 8, 2026). 2. Footprint, about 140 MiB idle RAM on a 10,000-title library, web UI on port 6262, a single Go binary in a shell-less Docker image. 3. Migration, a built-in tool reads configs from the six apps plus qBittorrent/SABnzbd without touching them; AGPL-3.0 licensed. No built-in downloader yet (planned), no phone or TV apps. If you run a media server, this is a direct way to shrink your maintenance surface. Repo · HN discussion

17. Zen 6 deep-dive shows EPYC single-thread up ~21% over Zen 5

Beyond clock speed, what got fixed this time?

Chips and Cheese published its AMD Zen 6 family analysis, 130 points, 46 comments, testing the EPYC 9686F (Zen 6) and EPYC 9996 (Zen 6c). Highlights: 1. Single-thread, on SPEC CPU 2026 estimates, the 9686F leads the Zen 5-based 9575F by about 21% in integer and 18% in floating point; at a matched 4.1 GHz, Zen 6c gains about 6.7% per clock in integer. 2. The out-of-order structures split, a 576-entry reorder buffer for Zen 6 versus 640 for Zen 6c, with 12-core 48 MB L3 CCDs and 32-core 128 MB L3 CCDs respectively. 3. Vector units, VNNI units doubled from 2 to 4, lifting 512-bit VNNI throughput from about 1.99 to 3.88 IPC, and a single thread can finally use both decode clusters. The cost: L3 latency rises from 47 cycles on Zen 5 to 55/66. If you pick cloud instances for inference or numerical work, the Zen 6c VNNI doubling is worth waiting for real benchmarks on. Post · HN discussion

18. 613K lines of Lean and 18,898 theorems in, 14% of CPython covered

An interpreter, worth proving?

David Andrews and Chao Zhang’s project copies CPython 3.15.2 line by line into Lean and proves properties about the copy, 123 points, 55 comments. The numbers: 1. Scale, started October 29; by November 10 it has accepted 4,071 functions, producing 613,318 Lean lines and 18,898 theorems and lemmas across 1,539 files, covering 14.3% of 28,393 checkable functions; grammar rules are at 100%, built-in types at 35%. 2. Method, every function gets at least 50,000 adversarial inputs for differential testing, and statements that fail to prove keep their counterexamples; the article’s advice is ‘copy the code, not the documentation.’ 3. By-catch, 418 bug reports filed, roughly 55 of them crash-or-corruption bugs reachable from ordinary Python code. Multithreading and memory bookkeeping are not modeled yet. If you run critical services on Python, that list of 55 crash-level bugs is worth walking through. Post · HN discussion

Work-from-the-woods people, is this the one?

The Starlink Mini 2 discussion thread sits at 175 points, 184 comments. Key specs as assembled from commenters: 1. Price, $999 bundled with one year of service; the Mini 1 recently cleared at $149. 2. Battery, a built-in 71 Wh pack rated up to 4.5 hours, quick-swap, with a USB-C port; the Mini 1 had no internal battery. 3. Power, runs on a 60 W supply versus 100 W for the Mini 1. 4. The trade-off, the Mini 1 accepted 12 to 48 V DC directly, while the new model goes through USB-C PD. Nobody in the thread cited official weight or speed numbers. The debate focuses on plan restrictions on travel use and standby mode not working in motion. If you work in the field and hated dragging a power bank for the Mini 1, wait for battery life testing before deciding. Post · HN discussion

Even clean-room code infringes, so what now?

LWN reports that developer Mike Lothian’s open-source Rust DRM/KMS driver for DisplayLink docks, Vino, was halted by Synaptics, 190 points, 83 comments. Timeline: first patch series on June 17 could drive three generations of DL3 docks; the August 17 cease-and-desist listed 125 patents and eight demands; October 5 brought claim charts for two UK patents plus a $25,000 ‘compensation’ demand escalating $25,000 per week; on October 23 OIN confirmed its protections do not cover codecs. The core dispute: Synaptics argues that under patent law even a clean-room independent implementation can infringe; Lothian says he ‘just can’t afford to defend myself’ and has taken the repo down. DisplayLink’s official Linux software is a proprietary binary limited to x86. If you work on reverse engineering or interoperability, archive every step of this case. Post · HN discussion

21. An 11-rule smart home stack that survives without internet

When the cloud drops, do your lights still listen?

The author distilled a decade of DIY home automation into eleven rules, 501 points, 179 comments. The core rule: the house must run with the connection unplugged, and cloud services like Alexa may only ever be one input method. The stack: ESP8266/ESP32 boards flashed with Tasmota, an MQTT broker, Node-RED flows, Grafana on a Raspberry Pi, and Tailscale for remote access. His scars: Google killed the Nest API he relied on, a Samsung fridge threw a TLS certificate error, and a Camp Chef grill needed Wi-Fi just to show its temperature. The rest of the rules: no cloud device without an API and logs, and no battery-powered device unless it lasts a year. Matter is never mentioned. If you are automating your home and fear vendor rug-pulls, these eleven rules are copyable as-is. Post · HN discussion

22. I deleted all my SSH keys

Nothing to steal if the key never sits on disk?

The author documents deleting every on-disk SSH private key in favor of hardware-backed credentials, 190 points, 80 comments, across three paths: 1. YubiKey, ssh-keygen -t ed25519-sk generates the key on the device, leaving only a handle on disk; the key material never leaves the hardware. 2. TPM, ssh-tpm-agent stores encrypted key material on disk and unseals it with the TPM; note it ignores PCR values, so a firmware or OS change will not lock the key. 3. macOS Secure Enclave, sc_auth create-ctk-identity creates a non-exportable key unlocked by Touch ID, with a handle exported via ssh-keygen. He concedes passphrases plus ssh-agent are probably fine for most threat models; the hardware benefit is that exfiltrated files are useless without the device. The costs: hardware loss is unrecoverable, so enroll multiple keys in authorized_keys, and Touch ID fails in clamshell mode, where a Touch ID keyboard or YubiKey 5C Nano covers you. Even if you do not convert everything, moving your highest-privilege logins to ed25519-sk is a one-afternoon upgrade. Post · HN discussion

23. A TV signal senses people through walls, for the cost of an SDR

The television in your living room is also a sensor?

Using an Ettus USRP B200 software-defined radio receiving a local station’s ATSC digital TV signal, the author demonstrates through-wall motion sensing, 256 points, 122 comments. The mechanism: 1. Multipath reflections cause selective fading, a state extremely sensitive to nearby movement. 2. Walking in the neighboring room produces stripes in the waterfall plot, spaced by Doppler shift. 3. With stripe gaps of about 70 pixels at a 150 Hz waterfall update rate he computed a 2.1 Hz gait frequency, and plugging the 490 MHz carrier in yields 1.3 meters per second, matching a walk. Wi-Fi and LTE signals work the same way, just with messier waterfalls and extra processing. The takeaway: any wideband radio receiver can detect motion through walls. For security teams thinking about indoor-tracking countermeasures, and for passive-sensing researchers, this demo is a reproducible starting point. Post · HN discussion

24. Fireworks breach reaches Notion AI chats in under 1% of workspaces

A vendor’s mistake lands on the user’s desk, how?

Notion posted about the impact of a security incident at third-party AI vendor Fireworks AI, 157 points, 58 comments. Key points: 1. Scope, fewer than 1% of workspaces may be affected, and what leaked was ‘unattributable’ Notion AI chat sessions: no user, workspace, or company identifiers, and no tool-call results. 2. Timeline, Notion proactively disconnected Fireworks and deleted all data in its systems on November 3, was notified of possible access on November 6, and confirmed a new dataset on November 8; the Insights feature is disabled and open-weight models no longer run through Fireworks. 3. Remediation, affected workspaces can review data via the Security incident review panel in the sidebar help menu, with CSV export for Enterprise; Notion recommends rotating any credentials pasted into Notion AI during the window. If you send sensitive context to third-party inference providers, this incident shows the concrete shape of a minimal blast radius. Post · HN discussion

25. China drafts lifetime ID codes for every robot it makes

One machine, one code, what about the ones for export?

China’s MIIT posted a draft rule on November 6: every industrial, service, or special-purpose robot produced in China must carry a lifetime, tamper-proof identification code, assigned under a ‘one machine, one code’ principle, 19 points, 17 comments. Supporting requirements: 1. Makers register each product’s key components and software on a ministry-overseen platform, which may be outsourced to a third party. 2. Each code must be filed within 20 working days, assigned before the robot leaves the factory; robots shipped before the rules take effect keep their existing serial numbers. 3. An earlier humanoid-robot standard from May 22 already uses a 29-character code, with over 28,000 humanoids registered on the platform by that date. Context: China produced 794,781 industrial robots in January–September 2026, up 29.1% year over year. By contrast, the EU Machinery Regulation (effective January 14, 2027) asks only for model, year of construction, and any existing serial number. If you ship robots into or out of China, the comment window closes December 5. Post · HN discussion