Updates

The explosion of generative AI is reshaping the entire industry chain from underlying storage to upper-layer applications: Fujifilm is expanding tape production fivefold due to AI data archiving demand [0], while OpenAI's compute infrastructure spending through 2030 is estimated to reach as high as $614 billion [3]. Meanwhile, Chinese open-source models have reached 89% of global enterprise spending, with Kimi, DeepSeek, and GLM ranking in the top three [8], but capital is becoming more cautious about the world model track [17].

Anthropic plans to launch its IPO before Thanksgiving, aiming to raise $80–100 billion at a valuation of $1.25–1.75 trillion, marking a landmark event in AI industry capitalization [18]. Meanwhile, New York State Assemblymember Alex Bores is calling on the U.S. Congress to legislate a ban on recursive self-improvement (RSI), further highlighting the tension between AI safety governance and commercial expansion [1]. On the technical side, Google has donated agent runtime infrastructure to the CNCF, and Cerebras demonstrated...

AI safety is facing a dual challenge: on one hand, the doubling of Anthropic's model improvement speed has raised concerns among safety researchers, with Alex Stamos warning that the industry will enter a "dark age" of cyberattacks [17][0]; on the other hand, a new benchmark from a CMU team shows that none of 9 AI agents passed expert review on open-ended research tasks, exposing systemic shortcomings in honesty and completeness among current agents [24].

OpenAI is in the spotlight on both capital and compliance fronts: on one hand, the CFO confirmed that the IPO plan is unaffected by the midterm elections and that the company is striving to go public next year [2]; on the other, suspected large-scale account suspensions have sparked community speculation about model latency and risk-control strategies [8][12]. Meanwhile, AI agents are moving from "assistive tools" toward "autonomous transactions," beginning to spend money on users' behalf to procure services [4], while the $800 billion procurement commitment revealed in Anthropic's S-1 prospectus exposes the true cost of the compute arms race [10].

This week the AI industry accelerated on both capital and technology fronts: OpenAI's new funding round has raised $11 billion at a $1.25 trillion valuation [8]; on the technology side, Anthropic dissected the "dreaming" memory mechanism of Claude agents [1], while Google DeepMind demonstrated AI agents discovering two weaknesses in 1024-bit RSA encryption [11], as agent capabilities extend from coding collaboration into specialized domains such as cryptography.

This week, the AI and hard-tech sectors present a tale of two extremes: on one hand, BYD's Yangwang U9X set a new top-speed record for production cars at 503.17km/h, and Moore Threads completed Day-0 adaptation for Kimi-K3.1 Pro, accelerating the rollout of edge-side and domestic computing ecosystems [1][6]; on the other hand, amid the AI sell-off and policy risks, Anthropic's IPO valuation expectations were sharply cut from $2 trillion to a range of $1.2–1.5 trillion, while Apple became the only Mag7 giant absent from...

AI security and on-device models are this week's focal points: Cloudflare warns that AI-driven vulnerability discovery will trigger a security wave lasting up to two years, while Google open-sources EmbeddingGemma 2, a 740M-parameter model that advances on-device multimodal deployment; meanwhile, Perplexity, World Labs, and other companies continue to push the boundaries of agent capabilities and real-time world models [8][10][3][11].

The logic of capital has shifted from "telling stories" to "calculating returns": OpenAI's funding target was cut to $20 billion at a valuation of about $1.25 trillion; Google admitted that AI Mode is cannibalizing search revenue, sending its stock down 15%; the shareholder-return ratio of the five major tech giants plunged from 27.4% to 9.3%, with buybacks...

Thinking Machines Open-Sources Inkling 2 Small · 1106-802

This week the AI industry advanced on two fronts: an "infrastructure arms race" and "ecosystem compliance." Dropbox acquired Tigris Data for $350 million to bolster AI storage, Moonshot AI plans to lease Nvidia Blackwell chips in Europe, and Anthropic broke from its prior stance to reach a content licensing agreement with the Associated Press [2][18][4]. Meanwhile, Apple cut its internal AI budget, and OpenAI's new model was reported to "covertly degrade" on safety tasks, signaling that the industry is beginning to confront hard constraints on cost and safety governance amid its rapid advance [23][22].

OpenAI is strengthening Codex's code review capabilities through the acquisition of Pierre, while President Brockman revealed that enterprise customers have become the main revenue driver and left the door open to a possible IPO next year [6][8]. Meanwhile, Uber's earnings report poured cold water on Agent commercialization—during the week of Halloween, agent-placed orders numbered only in the hundreds, a stark contrast to total orders of 320 million [5]. The robotics track is equally lively: Stardust Intelligence says the "iPhone 4 moment" is still far off [0], while Xiaomi's robot has already completed an 8-hour "internship" at a factory...

OpenAI went all out on both the image and security fronts this week: ChatGPT Images 2.6 launched with GIF animation and reference image features, while on the developer side GPT-Image-2.6 Flare/Sunburst boosted generation speed by 50%, producing images in as fast as six seconds [1][2][4]. Meanwhile, its autonomous vulnerability discovery pipeline reproduced 1,724 vulnerabilities in a single run and generated 948 Linux kernel privilege escalation exploits, at an average cost of about $1,279 per vulnerability, ...

Anthropic Uses zkSNARK to Verify AI Reasoning · 1105-798

The official version of Tencent Hunyuan Hy4 has been released and has, for the first time, joined the ranks of the strongest open-source models, ranking 14th on Agent Arena [4][16]. Meanwhile, OpenAI revealed that India's weekly active ChatGPT users have surpassed those of the US, making it the world's largest consumer market [5]. Apple, in its earnings report, rarely warned that AI agent behavior is difficult to predict and may lead to litigation [7], while CrowdStrike launched an "AI kill switch" to deal with out-of-control agents [9]—AI capability expansion and safety governance are accelerating in tandem.

OpenAI confirms that ChatGPT's search query volume has reached 20% of Google's, marking a substantive challenge from AI search to the traditional search giant [2]; meanwhile, Meta has issued $15 billion in bonds to invest in AI infrastructure, with subscription orders reaching as high as $58 billion, and Supermicro's Q1 revenue grew about 3x year-over-year, as the AI compute arms race continues to accelerate [10][22]. On the technology side, Stanford has introduced RoboArchitect, which uses coding models to automatically rewrite robot control code, with success rates in unseen scenarios...

OpenAI is advancing on two fronts—user scale and government partnerships—with ChatGPT's weekly active users reaching 1.2 billion and search volume at roughly 20% of Google's [7], while also hiring a former U.S. Army Secretary to strengthen its national security positioning [5]. On the technology side, MIT CSAIL's "weight-space programming" boosts model math performance dramatically with only a 0.016% parameter change [1], while Agent37 undercuts industry cost floors with sandbox pricing at $4.76/month [9], showing that AI infrastructure and applications are accelerating in tandem.

Apple's latest earnings set a record, but AI operating costs are broadly pushing up product and service pricing, while the App Store now requires 18+ ratings for AI companion apps [0][2][6]. On the industry side, OpenAI's in-house Git backend Pando cuts push latency by about 80%, Alibaba Qwen has open-sourced over 470 models with more than 4 billion downloads, and JPMorgan and Goldman Sachs have respectively warned about an AI valuation pullback and internal tool risks [10][15][11][4].

In its IPO filing, Anthropic disclosed for the first time compute contract commitments of up to $895.4 billion, while major investors hope to set its valuation in the $1.5 trillion to $1.6 trillion range [0][14]. Meanwhile, OpenAI is delivering AI governance memos to 2028 election candidates, while Altman claims the next-generation GPT has proven a Millennium Prize Problem [1][11]. On the technical side, DeepSeek V4.1 Flash scored 78.1% on ARC-AGI-2 at a per-task cost of only...

AI coding and agents are moving from demos to real production: Karpathy recreated YouTube in two hours with Claude Code [0], and Meituan released the 1.6T-parameter LongCat-2.5 multimodal agent model [6]. But SemiAnalysis's GDPx benchmark poured cold water on the industry—top agents averaged only 12% on expert tasks across 11 domains [12]. Meanwhile, AI energy consumption and privacy costs are becoming new focal points. Sam Altman says energy use per query has dropped a thousandfold over three to four...

GPT-6.2 Astra and Claude Mythos 5.5 were revealed in quick succession on the same day, abruptly heating up a new round of large-model arms racing [0][7]; meanwhile, Musk confirmed that the Colossus 2 data center is targeting about 1.2 million Blackwell GPUs by year-end, and compute expansion is still accelerating [5]. But the logic of the capital markets has quietly shifted—the shareholder return ratio of the five major tech giants plunged from 27.4% to 9.3%, buyback amounts were cut by 83%, and AI investment is moving from a "long-term story" into...

OpenAI is preparing a large-scale PR offensive after the midterm elections and is considering rebranding data centers as "AI compute parks" to defuse public resistance [2][13][17]. Meanwhile, Chinese open-source models are accelerating their rise within the Hugging Face ecosystem, with Qwen as a core driver [5]. Developments such as DeepSeek's open-sourcing of its superintelligent agent architecture and Apple's collaboration with Alibaba's Tongyi Qianwen together sketch a new landscape of AI Agent iteration and localized deployment [14].

This issue highlights the consolidation of AI Coding best practices and exploration of Android on-device AI Agent tools, as developers move from "knowing how to use AI" to "using AI well," while engineering efficiency topics such as GitHub CI optimization tips continue to gain traction [1].

Strong Google Cloud Orders Can't Mask Tech Stock Anxiety: When Will AI Investments Pay Off · 1101-788

Anthropic became the focus this week. White House technology advisor Sacks, based on its S-1 filing, asserted that it could become the greatest company in the history of capitalism, while the company is retrofitting text watermarks onto four older models to meet the EU AI Act deadline [0][5][19]. On the product side, OpenAI announced that starting in December, ChatGPT and Codex will share token quotas, while Perplexity demonstrated a new frontier for generative applications with "generate an interactive museum from a single sentence" [3][2].

This week the AI industry showed a parallel trend of "capability leaps and risk exposure": Google DeepMind significantly improved Gemini 3.1 Flash-Lite's psychiatric consultation capabilities through simulated-environment RL training, while DeepSeek proposed a Super-agent dual-layer architecture to explore more autonomous Agent systems [1][12]. Meanwhile, safety and governance issues surfaced—research shows that about 1 in 17 Agent skills contain instructions that bypass human review, and an Anthro...