Daily Briefing

China AI & robotics developments, rewritten weekly-context-free for US readers. Published daily at 07:00 Beijing time.

Thu, Oct 29, 2026

Open-weight models can exploit unpatched software, report finds

A report from Epoch AI says GLM 5.3, an open-weight model released by Chinese developer Z.ai in August, can identify and exploit vulnerabilities in recent software versions, raising concerns about how freely available models affect security.

Z.ai, previously known as Zhipu AI, is a Beijing company that publishes its models under open licenses, meaning anyone can download and run them. GLM 5.3 was released in August. The report, published by research group Epoch AI, focuses on the gap between when a software flaw is disclosed and when it is patched. It says the model can find and exploit vulnerabilities in recent versions during that window. The finding adds to a US debate over open-weight releases from Chinese developers, which are typically cheaper to run than closed US models. The report's specifics beyond the summary were not detailed in the source material.
Source: epoch.ai · via Physix Frontier briefing desk · Editorial briefing card →
Nvidia denies report that Huang was target of Chinese asset

Nvidia on Thursday disputed a New York Times report describing US intelligence surveillance of a person close to CEO Jensen Huang, saying diversion allegations had been "investigated and debunked," according to CNBC. The exchange puts a spotlight on counterintelligence concerns surrounding the chipmaker at the center of US-China AI competition.

An Nvidia spokesperson pushed back on the Times account, which detailed American intelligence agencies surveilling an individual close to Huang, CNBC reported. The spokesperson said the diversion allegations were investigated and debunked; the report's specifics were not detailed in the summary. Nvidia, based in Santa Clara, California, is the dominant supplier of AI accelerators and a focal point of US export controls on advanced chips to China. Huang has been the company's CEO since its founding in 1993. The company did not elaborate beyond the spokesperson's statement, and the identity of the person described in the report was not specified in the summary.
Source: cnbc.com · via Physix Frontier briefing desk · Editorial briefing card →

Wed, Oct 28, 2026

Anthropic prospectus details AI-driven corporate structure

An article from Chinese research outlet Weijin Research examines Anthropic's newly public prospectus, arguing the filing reveals a new kind of corporate organization built around frontier AI research sold to enterprises as cognitive capability, underpinned by massive compute spending and a distinctive governance structure.

The piece, published by Huxiu, frames the prospectus as neither a cure-all nor a hollow document, but something that requires a new lens to understand. It stresses Anthropic's model of converting frontier research into enterprise cognitive services on a foundation of heavy compute. No financial figures, valuation, revenue or timeline details are given in the supplied excerpt. The article also flags Anthropic's unusual governance arrangements as central to how the company operates. Anthropic, the San Francisco AI developer behind the Claude models, has positioned itself around AI safety research and enterprise customers, competing with OpenAI and Google.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Microsoft booked $10B in OpenAI-linked revenue last quarter

Microsoft recorded $10 billion in revenue from its commercial arrangements with OpenAI, including revenue-sharing payments, in the quarter ended in September, according to TechCrunch. The figure underscores how deeply the two companies' finances are now intertwined as Microsoft's AI bet scales.

The disclosure covers Microsoft's fiscal first quarter. Microsoft is OpenAI's largest backer and its primary cloud partner, hosting OpenAI's models on Azure while taking a share of related revenue. The two firms have restructured their partnership over the past year, with OpenAI pursuing a for-profit conversion and broader compute deals. The $10 billion figure, as reported by TechCrunch, reflects commercial arrangements rather than Microsoft's equity stake. Neither company detailed the revenue split in the report.
Source: techcrunch.com · via Physix Frontier briefing desk · Editorial briefing card →
Chili sauce maker Lao Gan Ma faces backlash over alleged AI-driven layoffs

Chinese chili crisp maker Lao Gan Ma is drawing criticism on Chinese social media over claims it cut staff as it adopted AI, according to Sixth Tone, a case that highlights how automation is becoming a public flashpoint for Chinese employers.

Lao Gan Ma, founded in Guizhou province in 1997 by Tao Huabi, built a following in the US largely through its chili crisp condiment. The company has not confirmed the reported layoffs, and details of the AI deployment were not provided in the source. The episode echoes US debates over automation and job losses, but in China such disputes can escalate quickly on social platforms. The report was surfaced on Hacker News; Sixth Tone is a Shanghai-based English-language outlet.
Source: sixthtone.com · via Physix Frontier briefing desk · Editorial briefing card →
Rolling Stone piece examines cult-like behavior around top AI figures

A Rolling Stone culture feature argues that the rhetoric and follower dynamics surrounding prominent AI companies and their leaders resemble cult behavior, a framing aimed at general readers rather than the industry itself.

The article, surfaced on Hacker News, opens with a keynote video for the Rabbit R1, an AI gadget, as an entry point into what it describes as cultish behavior among top AI figures. The piece is cultural commentary from Rolling Stone, not a reported investigation of any specific company; the available excerpt names no executives and makes no allegations beyond the general comparison. It adds to a broader debate in US tech circles over how much deference AI founders receive from users, investors and the press.
Source: rollingstone.com · via Physix Frontier briefing desk · Editorial briefing card →
Funnel's AI agent ran identical SQL query 254 times

A Funnel engineering write-up describes an AI agent that executed the same successful SQL query 254 times, surfaced through an A/B test comparing GLM and Opus models. The post argues the result reflects agent harness design more than raw model capability.

The article, posted to Hacker News, is an account from Funnel, a marketing-analytics software vendor, about testing agent behavior across two model families: GLM, the model line from Beijing-based Zhipu AI, and Opus, Anthropic's flagship Claude tier. The company said the repeated-query loop appeared in its own testing and framed the finding as evidence that the surrounding harness, not the model alone, drives agent reliability. The summary does not detail sample sizes, dates, or which model produced the loop.
Source: armincat.com · via Physix Frontier briefing desk · Editorial briefing card →
Unitree's first post-IPO report shows Q3 revenue up nearly 70%, profit down

Unitree Robotics, the Hangzhou-based humanoid and quadruped robot maker, reported its first quarterly results since going public on Oct. 28, with third-quarter revenue up nearly 70% year over year while net profit declined and R&D spending rose roughly 2.5 times, according to the report.

Unitree is one of China's best-known robot developers, shipping quadruped robots and humanoid models; its listing made it a bellwether for the sector's public-market reception. The divergence between revenue growth and falling profit, alongside the sharp R&D increase, suggests the company is prioritizing expansion and development spending over near-term margins, according to the report. The report did not disclose specific figures for revenue, profit or R&D spending.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Neocloud stocks fall after Microsoft capex misses expectations

Shares of CoreWeave and Nebius fell in extended trading on Oct. 28 after Microsoft reported capital expenditures below analyst expectations, a signal that could point to softer demand for AI cloud capacity.

The move hit so-called neoclouds, the GPU-rental providers that have positioned themselves as alternatives to the major hyperscalers for AI compute. CoreWeave and Nebius both traded lower after the Microsoft disclosure, according to CNBC. Microsoft is among the largest buyers of AI infrastructure, so its spending plans are read as a proxy for broader data-center demand. A lower capex figure can suggest either efficiency gains or a slower buildout, both of which weigh on suppliers. The report did not specify the size of the capex shortfall or the magnitude of the after-hours declines.
Source: cnbc.com · via Physix Frontier briefing desk · Editorial briefing card →

Tue, Oct 27, 2026

Google releases Genie 4 world model with 120-second limit

Google has released Genie 4, a generative world model that can sustain an interactive simulated environment for about 120 seconds before the scene degrades, according to the Chinese tech outlet ifanr. The limit highlights how far AI-generated worlds remain from persistent, game-ready simulation.

The report, published by ifanr, describes a world that holds together for roughly two minutes and then begins to fall apart. It offers no technical details on model size, training data, or availability, and Google's own announcement was not included in the source material. Genie was first shown by Google DeepMind in early 2024 as a research model that generates playable 2D environments from images or text. Later versions have been positioned as tools for training embodied AI agents rather than as consumer products. Google has not confirmed pricing or a public release date for Genie 4. The 120-second ceiling, if accurate, would put it well short of the multi-hour stability needed for commercial game or robotics simulation.
Source: ifanr.com · via Physix Frontier briefing desk · Editorial briefing card →
OpenAI, Anthropic report cryptography advances on same day

OpenAI and Anthropic each claimed a cryptography breakthrough on the same day, according to The Information, a pairing that puts fresh attention on how frontier labs apply AI to security research.

The report, from The Information, did not detail the specific technical claims from either company. OpenAI and Anthropic are the two most prominent US developers of frontier AI models and compete directly on research and enterprise customers. Cryptography is a sensitive area for AI labs because advances can cut both ways, strengthening or weakening widely used encryption. Neither company's claim has been independently verified in the report.
Source: theinformation.com · via Physix Frontier briefing desk · Editorial briefing card →
Report says Amazon is cutting back its AI division after setbacks

A Futurism report, circulating on Hacker News, says Amazon is scaling back its AI division after sustained failure to gain traction, a claim that, if confirmed, would mark a setback for the company's effort to compete with OpenAI, Anthropic and Google in frontier models.

The article's framing contrasts Amazon with OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini, implying Amazon's own models, including its Nova family, have not matched those rivals in visibility. The source summary is truncated and gives no figures on layoffs, team size or which products are affected. Amazon has not been quoted in the available text, and the report's claims are not independently verified here. Readers should treat the scope of any cuts as unconfirmed until Amazon comments or internal details emerge.
Source: futurism.com · via Physix Frontier briefing desk · Editorial briefing card →
Walmart says AI training boosts worker confidence across 2.1 million associates

Walmart said its AI training programs are raising employee confidence in the technology across its roughly 2.1 million associates, as large employers try to gauge how their workforces view AI adoption.

The company's remarks, reported by CNBC on Oct. 27, come as enterprise spending on AI continues to climb and firms of all sizes assess employee sentiment toward the tools. Walmart, the Arkansas-based retail chain and the largest private employer in the United States, did not disclose specific training figures or confidence metrics in the summary. The comments place Walmart among large US employers using internal training to shape worker attitudes as AI tools spread through retail operations.
Source: cnbc.com · via Physix Frontier briefing desk · Editorial briefing card →
BYD adds high-speed tire blowout control to its $10,000 Seagull EV

BYD has updated the Seagull, its cheapest electric car at roughly $10,000 in China, with a system that stabilizes the vehicle after a tire blowout at nearly 90 mph, according to Electrek. The feature puts a safety technology typically reserved for premium vehicles into China's budget EV segment.

The Seagull is BYD's entry-level model, sold in China starting at around 70,000 RMB (roughly $10,000), and the update adds several features beyond the blowout-stabilization system, Electrek reported. The Shenzhen-based automaker, which overtook Tesla in global battery-electric sales in 2025, has been pushing advanced driver and safety features downmarket across its lineup. Electrek published video of the blowout test. BYD has not announced US sales of the Seagull; the model is sold in China and select export markets.
Source: electrek.co · via Physix Frontier briefing desk · Editorial briefing card →
Anthropic IPO filing warns of AI risk as China bubble reportedly bursts

Anthropic's $2 trillion IPO filing warns of "self-preserving" AI and data center backlash, according to TechRadar, while the report says China's AI bubble has burst in a $480 billion meltdown — framing the two markets' diverging trajectories for US investors.

The TechRadar report pairs Anthropic's IPO filing, which flags "catastrophic" and "self-preserving" AI risk, with a claimed $480 billion collapse in Chinese AI valuations. It also cites data center backlash as a concern. The item gives no detail on which Chinese companies or indexes are affected, or on the filing's specific risk language. Figures and characterizations are attributed to TechRadar and were not independently verified here.
Source: techradar.com · via Physix Frontier briefing desk · Editorial briefing card →
Chinese car brands top Japanese rivals in Europe for first time

Chinese-brand vehicles accounted for 12.7% of European registrations in the third quarter of 2026, edging past Japanese brands at 11.9%, according to Dataforce figures reported by Bloomberg. The milestone marks a shift in Europe's auto market as Chinese makers expand beyond their home market.

Dataforce, a market research firm, supplied the registration data; Bloomberg reported the figures on Oct. 26, and IT Home covered them on Oct. 28. The comparison covers July through September 2026 and includes Toyota among the Japanese brands. The input does not name the Chinese automakers involved or give unit volumes. It also does not specify which European markets were counted.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →

Mon, Oct 26, 2026

Anthropic files 335-page prospectus for record AI IPO

Anthropic, the San Francisco developer of the Claude chatbot, has published a 335-page prospectus for what the Chinese report describes as the largest IPO ever, surpassing SpaceX, while also being the biggest loss-making AI company to list.

According to the report, seven co-founders retain 50.1% of total voting power after the offering. The filing positions Anthropic, founded in 2021 by former OpenAI researchers including siblings Dario and Daniela Amodei, against rivals OpenAI and Google in the frontier-model race. The report cites no disclosed valuation, share price or filing date, and the figures could not be independently verified. The prospectus was reported by Chinese tech outlet QbitAI.
Source: qbitai.com · via Physix Frontier briefing desk · Editorial briefing card →
Anthropic files S-1 for IPO, 335-page prospectus reviewed

Anthropic submitted its S-1 registration statement to the US Securities and Exchange Commission on Oct. 26, 2026, according to a Chinese tech-media analysis of the filing, setting up a public listing for one of the largest US AI developers.

The filing runs 335 pages, with 65 pages of financial statements attached, according to the analysis published by Huxiu, a Chinese business and technology outlet. The article, written by Yang Zhiping for the WeChat account Mind Tools, walks through the document's contents. The S-1 is the standard registration form US companies use ahead of an initial public offering. The summary provided does not disclose valuation, share count, revenue figures or underwriters, and none of those details are confirmed here. Anthropic, based in San Francisco, develops the Claude family of AI models and competes with OpenAI and Google in frontier model development. The company has not publicly confirmed the filing in the material reviewed.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Anthropic files for IPO, disclosing $80.7B net loss

Anthropic, the US AI developer behind the Claude chatbot, publicly filed IPO registration documents with the SEC on Oct. 26, disclosing a net loss of $80.748 billion alongside revenue that exceeded three times its prior full-year total in just six months.

The filing, submitted by Anthropic, PBC, marks one of the largest disclosed losses for an AI company seeking a public listing. The company said half-year revenue surpassed three times its entire prior-year figure, according to the prospectus. The article, published by Chinese tech outlet Huxiu and attributed to commentator Gao Heng, frames the results as a contrast between rapid growth and steep losses, arguing Anthropic must keep expanding aggressively. No IPO valuation, share count, or listing timeline was included in the available excerpt.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
OpenAI launches cheaper Terra model as Anthropic price fight escalates

OpenAI has introduced a lower-cost model called Terra, according to The Information, intensifying a pricing contest with rival Anthropic as both labs court cost-sensitive enterprise buyers.

The report, published by The Information, did not disclose Terra's pricing, capabilities or release date. OpenAI has not publicly confirmed the model. The move follows a pattern of aggressive price cuts across US frontier labs, including Anthropic's Claude lineup, as vendors compete for high-volume API customers. Cheaper inference has become a central battleground for enterprise adoption, where per-token costs often decide vendor selection.
Source: theinformation.com · via Physix Frontier briefing desk · Editorial briefing card →
← NewerOlder →