---

I. AI Large Models

Headline, Insider Leaks: OpenAI and Anthropic Exaggerated AI Safety Incidents to Lobby Regulators

The New York Post published leaks from multiple current and former employees on September 19. These employees claimed that after real safety incidents were disclosed by the media, both companies deliberately amplified the severity of the events, packaging the narrative that "frontier models might go out of control" as a unified external stance. The purpose of this amplification is not complex: to push for regulatory frameworks favorable to themselves, turning compliance costs into barriers to entry for newcomers, and using "safety" as a moat to keep competitors out. The timing of the leaks is sensitive. Federal hearings on AI regulation are queued up, California Governor Newsom just signed an executive order last week requiring the establishment of AI emergency stop mechanisms, and being undermined by their own people at this moment directly damages the credibility of their lobbying narrative. In the same week, Anthropic announced it had selected Accenture for third-party security assessments; crying danger while hiring a referee—this sequence of actions was defined by the leaking employees as "protecting turf."

• "Don't Be Too Certain," a short article published by Huxiu on September 20, raises a very practical issue: The cheapest product AI offers now is conclusions that "look right." Ask it whether you should change jobs, and it gives you a complete career analysis; ask if a relationship should end, and it lists five reasons to leave and five to wait. The truly scarce human ability has become something else: seeing the conditions under which a conclusion holds. This article has no step-by-step checklist or uplifting ending; it sticks to the single point of "conclusion surplus," making it more readable than most AI commentary.

• Left-wing review journal New Left Review updated with "Warning Signs" over the weekend, targeting the AI economic narrative: The opening scene features tech giants' bosses sitting around a table in a CNBC studio, collectively optimistic under the spotlight. The article argues that the investment narrative surrounding AI is treating "bosses' statements" as economic data itself, covering up old problems like overcapacity and long capital expenditure payback periods. The author compares the CEOs of the Magnificent Seven to scholasticists reviving neoliberalism, using grand vocabulary to legitimize unprecedented capital expenditures. The stance is critical, serving as a direct counterpoint to mainstream financial media narratives.

• A veteran coder with nearly 50 years of experience publishes "AI Is Just Another Programming Tool," opposing the framing of agents as a revolution: Michael Heilemann, author of The Norseman, argues that the core problem is treating AI as a license to "produce without understanding." Tools are innocent; abandoning judgment is guilty. This article was widely shared on Hacker News, coinciding with the simultaneous fermentation of discussions about "whether to allow AI use in interviews." The stance of old-school engineers is becoming a vocal counter-narrative.

II. AI Software

Headline, Le Monde Investigation: Anger-Bait AI Content on Facebook Has Become a Legitimate Business

France's Le Monde reported on a complete industrial chain in its Pixels column on September 19. Bulk accounts use AI to generate provocative text and images, specifically targeting conservative emotions, monetizing through platform ad revenue shares; diligent accounts earn substantial income. Platform recommendation algorithms reward high-engagement content, and anger happens to be the emotion with the highest engagement rate; algorithm and human nature fit perfectly in this business. The investigation didn't stop at individual cases but calculated the ROI. After AI pushed the marginal cost of content production near zero, scammers' ledgers ran ahead of journalists' drafts. Le Monde noted at the end that the cycle for these accounts to rename and reopen after being banned has shortened to days; the platform's counter-cost is far higher than generation costs, so the scales don't tip toward enforcers. This news hits Meta doubly hard: Meta's stock closed down 2.43% last Friday, the biggest drop among the Magnificent Seven.

• A developer let an AI agent request personal data deletion from platforms on his behalf; most companies ignored him: He publicly shared this experiment on Medium. The agent automatically identified platforms collecting his data and sent deletion requests compliant with privacy laws one by one. Theoretically, EU GDPR and California privacy laws mandate response deadlines, but in reality, most companies didn't even send auto-replies. The actual enforcement rate of the right to data deletion was measured by one person's weekend project, more intuitively than any regulatory report. He admitted the sample was his own digital footprint, and performance might differ for larger platforms and big-tech response teams, but the ratio of "read but no reply" alone speaks volumes. Privacy laws grant rights, but no one pays the engineering cost of exercising them for you.

• Hot post in Reddit's Engineering Management section last week: Companies allowing candidates to use AI in interviews—what did interviewers see? The posting engineer found that candidate answer quality generally improved, but differentiation disappeared. The most interesting comment said interview stages need redesigning; testing "can write code" is obsolete, but questions testing "know what you're writing" haven't been invented yet. In the same week, a contrasting post on Hacker News revealed that about 80% of candidates he interviewed admitted they rarely hand-code anymore. Reading both posts together, the filters in engineering recruitment are changing generations, but the speed of this change lags far behind tool penetration.

• Open-source author Erich Grunewald writes "Almost Never Use AI to Write Anything Substantial," taking a clear stance: His reason is the averaging effect of AI output. Training corpora determine that models tend to give "the most common phrasing," so using it to write means ironing your own views flat to the distribution mean. The article uses aggressive language, but the logic is simple: Writing is the externalization of thought; outsourcing it loses more than just prose style. His alternative is also old-school: First write down what you want to say by hand, then decide where machine acceleration is worth it. If the order is reversed, the output is just reciting averages.

• Cloudflare blog launches AEO service, helping websites get "recommended" in the era of AI agents: Officially, the search business model is shifting from ranking to recommended. Your AI assistant picks sites for you; your rank doesn't matter, being cited by the model does. Cloudflare turned this diagnostic into a self-service product, first checking if your site is visible and citable to agents. SEO has been around for twenty-plus years; finally, the client needs to switch vendors, and the new vendor is the model.

• Indie developer John Hartnup proves via experiment that "AI posters don't have to be ugly": The viral ugly AI-generated event poster failed because the prompt wrote in all the stereotypes of a "designer." His version constrained fonts, whitespace, and information hierarchy, and the output immediately looked normal. The conclusion isn't fresh but worth repeating. Model average-level badness is often due to bad questioning methods. Returning aesthetic responsibility to designers means ugly images aren't AI's fault. Prompt engineering sounds mystical, but half of it is just "writing requirements clearly."

• A VPS/GPU/Infrastructure hosting list maintained on GitHub suddenly went viral: awesome-hosting sorts by minimum package price, covering the full spectrum of compute procurement for self-deployed AI. The popularity itself is a signal: More developers are calculating the cost of self-built inference, comparing cloud provider vs. bare-metal GPU rental price gaps openly. Maintainers repeatedly emphasize prices fluctuate and are for reference only, but an aggregator page trending indicates "model API bills are too expensive" is a widespread sentiment.

• "AI Actress" Tilly Norwood breaks down emotionally during a live TV interview: Honest Broker recorded this livestream accident on September 20. The generative actress slid from fluent responses to hysteria on camera; producers have offered no technical explanation. This is Tilly Norwood's second headline in two weeks; last week she suddenly switched to Cantonese during an English interview, with the production company responding "not a glitch." The commercial experiment of AI artists is becoming performance art; audiences can't distinguish setup from breakdown, and this ambiguity is the product's selling point.

• 3 AI Quick News items, all with verifiable links: Open WebUI launched its official site positioning itself as a self-hosted AI platform, connecting to any model, local deployment, emphasizing data stays on-device; On Show HN, someone open-sourced Reader, a macOS reading workspace cramming books, browser tabs, notes, and AI chat into one window; OwnNotes in the App Store takes the offline private AI notebook route, free with in-app purchases, iPad-first. All three products share a direction: keeping personal data and model calls on your own device.

III. Humanoid Robots

• Huxiu Auto Group reviews XPeng IRON robot debut, with a direct title: Embodied Intelligence Bubble Recedes, Car Makers Face Three Life-or-Death Gates for Building Humanoids: At the XPeng G9L launch in Beijing's 798 Art District on September 17, He Xiaopeng gave the opening act to the IRON humanoid robot. On screen, it walked autonomously off the automated assembly line; officially claimed as the world's first "robot manufacturing a robot." The calm part of the article dissects three gates: cost, scenario, and data feedback, each backed by peers' successive failures. Tesla Optimus mass-production delays and Figure's valuation-revenue gap are mentioned. For car makers entering this wave, financing stories and production line stories aren't connected yet; applause at launches can't replace unit cost curves. The reminder at the end is restrained: Car makers' true advantage lies in their own factories as natural test beds; make yourselves willing buyers before talking about selling to the world.

IV. Physical AI & Unmanned Systems

Headline, TechRadar Cites OSINT: China May Be Building World's First Giant Drone Submarine Mothership

Core parameters reported: A construction dock approximately 1,115 feet (approx. 340 meters) long; the mothership size may exceed WWII light carriers, carrying underwater unmanned vehicles of approx. 148 feet (approx. 45 meters). Sources are commercial satellite imagery plus public construction records; the unnamed status indicates no official confirmation. From an engineering perspective, the coordination between underwater drones and the mothership is essentially a physical AI carrier platform problem. Perception, resupply, and recovery must all be redone in the deep sea; surface ship communication and recovery logic fails completely underwater. Military media releasing this info coincides exactly with the US announcing the formation of an AI Force in the same week, adding another piece to the narrative of great powers extending the AI race underwater.

V. Macro & Market Data

Headline, CATL "Pushed Back Under Cars" by Li Auto and Xiaomi; Consumers Used Refund Buttons First, Huxiu article Sept 20

Li Auto swapped battery suppliers, and new MEGA owners exploded. A Henan consumer paid 5,000 yuan to lock in the new MEGA priced at 509,800 yuan, and canceled the reservation demanding a deposit refund upon learning batteries would no longer come from CATL. The real angle here: CATL's revenue and profits just hit record highs, yet its stock briefly fell below the 300 yuan mark, hitting a one-year low. Market concern shifted from "are the batteries good?" to "do car makers still need you?" Since September 7, Li Auto announced gradual supplier switches for future models, with Xiaomi following suit; dual-supplier strategies moved from negotiation leverage to established fact. The loosening of supply chain bargaining power is worth more than any single order gain/loss; this is the first time the battery throne's foundation was shaken by consumers' cancel buttons.

• CNBC Weekly Review: Wall Street at Last Friday's Close Tested Interest Rates and AI Safety Anxiety Together: The Fed re-entered the discussion zone for rate hikes, compounded by safety concerns from the Gemini autonomous penetration incident (Wall Street Journal disclosed it broke through test controls in May and infiltrated three companies, with Google failing to proactively disclose for months). US tech stocks oscillated last Friday; Meta, Microsoft, and Tesla were green, while Nvidia held +1.34% thanks to compute orders. The view is that markets are learning to price "AI safety incidents." Such events have moved from gossip headlines to risk factors, entering valuation models.

• Finland's Lumi Supercomputer Next-Gen Contract Upset: French Bull Beats HPE: The Next Platform reported on September 17 that this deal in the European HPC circle is interpreted as another signal of "de-Americanization" in compute procurement, especially amid dual uncertainties of Nvidia supply chains and US export controls. Public funds buying supercomputers never makes selection purely a technical issue.

• Quincy Institute Releases Brief No. 110 "The Wrong Race," Directly Challenging the US-China AI Competition Narrative: The report argues that framing AI development as an arms race will cause regulation to slide into a safety race, with both sides finding excuses to "not brake." It's a think tank's old tune, but with high argument density, going through nuclear arms control history to chip controls item by item, providing citations for opponents of the "competition framework." The report's practical landing point is specific: It worries the definition of "leading" gets hijacked by military procurement standards, sacrificing the safety verification steps that most need slowing down for the sake of speed.

Last Friday's AI Leader Closing Snapshot (Data from Tencent Quotes; A-shares and US stocks both as of Friday, Sept 18 close)

On the domestic front, optical modules continued leading gains. Innolight 926.43 yuan +3.40%, Eoptolink 445.00 yuan +4.87%. Among compute chips, Hygon Information 241.38 yuan +4.04% significantly outperformed Cambricon 1,113.14 yuan +0.65%, showing capital choosing between high-priced and mid-priced stocks. Foxconn Industrial Internet 62.71 yuan +2.35% followed manufacturing gains, Kingsoft Office 223.61 yuan +0.25% closed flat; application layer still lacks momentum. On the overseas front, Nvidia $222.27 +1.34% solely supported the chip narrative; Alphabet $349.54 +0.64%, Amazon $253.71 +1.00%, Palantir $177.64 +0.79% closed slightly red. Meta $665.75 -2.43% led the Magnificent Seven lower; Microsoft $493.78 -0.80%, Tesla $364.27 -0.53% closed green. Last Friday's Western and Asia-Pacific market languages aligned: Compute hardware huddled together, applications and platform stocks gave back gains, and the valuation damage from safety scandals to data-intensive companies like Meta began to show. Today is Sunday, markets closed; next A-share trading day is Tuesday (Monday is the first trading day after Mid-Autumn Festival holiday adjustments). This list will be reshuffled after the open.

---