Fujian 9‑year‑old earns six‑figure income with AI; Anthropic shuts down 11.4 million accounts in six months.
Fujian 9-Year-Old Boy Earns Six Figures with AI; Anthropic Bans 11.4 Million Accounts in Six Months…
Content Summary
Compiled by | Fu Yuqi, Wei Siwei (editors) | April Lead: Nvidia reportedly halts production of RTX 5090, top gaming GPUs give way to high‑price professional cards; Fujian 9‑year‑old boy earns six figures with AI and even grants scholarships to middle school students
NVIDIA may discontinue its most powerful consumer graphics card, the RTX 5090, reserving the core chip for higher‑priced professional cards. Hardware leaker MEGAsizeGPU says the future‑produced GB202 chip will be supplied exclusively to the RTX PRO series and will no longer be used in the RTX 5090. Another leaker, hongxing2020, posted what appears to be a factory notice stating that the China‑specific version is also being discontinued. Both types of cards use the same GB202 chip, but their prices differ dramatically. The RTX 5090 comes with 32 GB of VRAM and a launch price of $1,999. The RTX PRO 6000, aimed at AI development, design, and rendering, has 96 GB of VRAM and is listed on NVIDIA’s website for $16,000—eight times the price of the former. While price doesn’t directly equate to profit, the high‑end professional market clearly offers NVIDIA a much larger revenue opportunity.
The RTX 5090 has already been noticeably out of stock in the U.S., with some third‑party sellers listing it for $9,600. If the discontinuation is true, consumers will have to rely on remaining inventory. A bigger shift could be a step back in GeForce flagship performance. Another leak claims NVIDIA plans to launch a 24 GB RTX 5080, which would replace the RTX 5090 as the top‑end gaming GPU. The current RTX 5080 has roughly half the CUDA cores of the RTX 5090 and delivers about 30 % lower 4K gaming performance. Even with the memory bump to 24 GB, it would still fall short of the RTX 5090’s 32 GB.
Yu Zhile, a 9‑year‑old boy from Xiamen, Fujian, recently established the “Zhile AI Scholarship” at Longshan Middle School in Nanjing, Zhangzhou, awarding scholarships to the school’s students. He is currently in fourth grade. According to the Min Nan Daily, he began learning AI in January of this year, then used AI to create tools, products, and websites, and tried to sell them. Including earnings from competition participation, his total income is reported to have reached six figures in Chinese yuan.
Yu Zhile’s mother, Zhuang Sehui, runs an internet company in Xiamen, and Longshan Middle School was her junior high alma mater. She has previously been continuously supporting students from her hometown. This time, the mother and son decided to allocate part of their income to establish a scholarship, and plan to award outstanding students each year in the future. Longshan Middle School principal Wu Ruixing said the scholarship will be used to encourage students who excel academically and morally and are passionate about technological innovation.
On October 11, Fields Medalist Terence Tao posted the original 27‑page PPT “Math 2.0” from his Caltech talk on his official blog. The talk was originally scheduled for May (titled “Machine‑Assisted Proofs”), but was substantially revised due to the recent dramatic changes in AI. In the blog, Tao notes that the academic community is now overly emphasizing “AI‑automated solutions to open mathematical problems.” He warns that “solving quickly” does not equate to genuine human mathematical understanding. In the forthcoming “Math 2.0” era, the role of AI tools should be re‑calibrated to support and sustain healthy development within the human mathematics community, rather than replacing human thought with a flood of machine‑generated proofs that people cannot digest. He calls for a shift in resource evaluation back to a science vision centered on “human understanding,” and for realigning mathematics with practical real‑world applications. The full PDF of the presentation is now available for public download on his personal website.
On October 8, according to KuaiTech, Zhipu founder Tang Jie recently introduced the development direction of the next‑generation large model in a speech in Singapore, including expanding pre‑training data and model scale, strengthening post‑training, and improving long‑range reasoning and agent capabilities. The next‑generation model may be named GLM‑5.4 or GLM‑5.5 and will grow to a trillion‑parameter scale, but the specific model name and parameters await official confirmation. Technical exploration focuses on recursive self‑improvement and enabling agents to autonomously perform longer‑term tasks in more complex environments. Tang also previously proposed a “fully self‑training” research goal, hoping the model can participate in its own training and iteration, with the key challenge being to let the model decide when to correct errors and when to stop. Currently, these items are research directions and goals and should not be taken as capabilities the new model already possesses.
Anthropic’s transparency report updated on July 23 shows that 11.4 million accounts were banned in the first half of 2026. The second half of 2025 saw 1.45 million bans, and the new period is about 7.9 times the previous one. The statistics cover Anthropic’s products and services and do not separate Claude chat from developer products.
During the same period, 398,000 account‑ban appeals were received, of which 42,000 overturned the original ban decisions. Based on the two concurrent data points, the reversal rate is about 10.6% of the appeals. Anthropic states that accounts may be banned for violating usage policies, terms of service, or regional support policies.
On October 7, Musk said on X that Grok Bot will choose the backend model most likely to deliver the best results for a given task, including Claude Opus 5.5, Midjourney, Suno, and other leading APIs. The same day, 9to5Mac reported that Grok Bot is already able to use Claude Opus 5.5. This adjustment means its model selection will no longer be limited to in‑house products but will combine capabilities from different vendors to complete tasks; Musk has not disclosed the specific division of labor among the models or the full rollout schedule. Grok Bot is equipped with a continuously running cloud computer that can perform multi‑step tasks across applications via browser, file system, terminal, and connectors, and multiple bots can collaborate in parallel and share context. The model selection is managed by the platform, which does not provide a user‑facing model‑selection menu; usage analytics will show which model was actually used for each request, and billing will be calculated based on the model employed.
On October 10, Business Insider reported that while Google is preparing to publicly launch the next‑gen Gemini 4 “Argon” model, an internal leak revealed intensive testing of a follow‑up checkpoint version codenamed “Carbon.” The model has already been integrated into Google’s internal coding and development platform Jetski (the Antigravity platform). Several Google employees involved in the internal test said Carbon performs exceptionally well on code generation and complex long‑running agent tasks, with programming capabilities that seem comparable to Anthropic’s top model Claude Opus 5.5. Sources say Carbon’s huge performance leap stems from the use of Recursive Self‑Improvement (RSI) technology, giving it the ability to autonomously detect errors and iteratively optimize. However, because internal codenames at tech companies rarely match final commercial names, it is still unclear whether Carbon will be released as a standalone new model or as a regular upgrade to Argon.
On October 9, according to an exclusive report by Axios, senior executives and top researchers at frontier AI companies such as Anthropic and OpenAI are privately rehearsing a “day after” disaster response plan. The trigger for this effort is the rapid acceleration and widespread deployment of AI technology, which has pushed internal industry concerns about potential safety vulnerabilities to a tipping point. Recently, cybersecurity firm CrowdStrike disclosed that attackers allegedly used malicious AI to breach two South Korean banks and attempted to employ Claude Code to process stolen data. This real‑world incident has sparked heightened vigilance about out‑of‑control risks. Many insiders predict that within the next six to twelve months, malicious actors or rogue‑agent swarms could launch large‑scale cyber attacks, causing brief disruptions to financial services, internet connectivity, or power and water infrastructure. In response, AI executives are proactively assessing possible public backlash and policy pushback, and are planning to accelerate outreach to legislators about the current state of AI technology. While OpenAI emphasizes that these exercises are precautionary rather than inevitable, the major players generally hope that, should a future industry crisis arise, their forward‑looking preparations will enable them to help policymakers shape more sensible emergency legislation and subsequent regulatory frameworks.
On October 8, Reuters, Bloomberg and several other foreign media outlets, citing sources familiar with the matter, reported that OpenAI disclosed to investors that its annualized revenue for September was close to $50 billion, about $20 billion less than the nearly $70 billion previously reported by the media. The huge discrepancy is not due to a slowdown in business, but stems from the fact that investors, when trying to compare OpenAI directly with its main rival Anthropic, found the two companies use different accounting methods. By contrast, competitor Anthropic reported annualized revenue of over $65 billion in July, largely because Anthropic uses the “gross‑revenue method,” counting the full amount of services sold through cloud partners such as Amazon and Google as revenue; OpenAI, on the other hand, uses the “net‑revenue method,” and since updating its commercial agreement with Microsoft in April, it only recognizes the actual share it receives from certain partner channels. The previously cited $70 billion figure was a market estimate based on the gross‑revenue method. Although this statistical difference caused a brief market panic and led AI‑related stocks such as Nvidia and Oracle to dip, the real business of both giants continues to surge. CNBC obtained presentation materials showing that OpenAI’s overall annualized revenue grew 77 % in Q3, with enterprise revenue jumping 107 %. Driven by strong enterprise growth, Bloomberg later reported that OpenAI expects its annualized revenue to reach or exceed $70 billion by the end of this year. The financial data above come from investor communications obtained by the media, and OpenAI has declined to comment on the reports.
On October 10, the Financial Times reported that chip giant Nvidia is in early talks with U.S. AI startup Reflection AI. Nvidia plans either a full acquisition of the company or an additional investment. The current focus includes an “acqui‑hire” approach—hiring staff and obtaining technology licenses to bypass antitrust review; if a full acquisition cannot be completed, Nvidia is also considering deepening cooperation through additional equity investment, expanded chip supply, and compute support. The parties could reach an agreement within the next few weeks, though the talks could still fall apart. Reflection AI was founded in 2024 by former DeepMind researchers, focusing on cutting‑edge open‑weight models and autonomous coding agents. Its pre‑money valuation reached $25 billion in its April 2026 financing, and Nvidia had previously invested $800 million as a core strategic investor. The company just released its first open‑weight model, Beam, on Monday, October 5, positioning it against models such as DeepSeek, Zhipu GLM‑5.2, and Kimi for programming and agent tasks.
On October 9, in response to the National Radio and Television Administration’s call for a special campaign against “AI‑remixed” videos, the three major platforms—WeChat, Douyin, and Kuaishou—simultaneously disclosed their September content compliance results. The platforms dealt a heavy blow to accounts that abused generative AI to crudely deconstruct classic films and animations or to make disruptive alterations, clearing a total of 7,436 violating videos in September. The data shows Douyin handled 2,645 related violations, Kuaishou dealt with 2,544, and WeChat removed 2,247 pieces of harmful content. This crackdown focused on improper romantic pairings of classic characters such as those from “Journey to the West” and “Dream of the Red Chamber” (e.g., the viral “Sheng Dai CP” AI‑generated videos), malicious distortion of historical heroes, and the malicious “cult‑ification” or horror‑ification of cartoon figures like “Pleasant Goat and Big Big Wolf.” It marks a comprehensive tightening of platform compliance reviews for generative AI‑created derivative content (AIGC).
On the evening of October 10, Microsoft CEO Satya Nadella posted on social media that enterprises should treat powerful AI models as “potential internal threats.” He urged the establishment of a “Kill Switch” mechanism when deploying advanced AI (especially autonomous agents) so that authorized personnel can pause or shut them down at any time, achieving a separation between the supply of intelligence and control.
On October 5, it was reported that Alibaba’s Qianwen AI ear‑clip headphones are now available on e‑commerce platforms and open for reservation, with the exact price not yet disclosed.
The headphones come in three colors—Deep Space Blue, Starlight, and Aurora Purple—are ambidextrous, support IP55 dust and splash resistance, can pair with up to 8 devices, and allow simultaneous dual‑device connection; they offer 9 hours of battery life, extendable to 40 hours with the charging case, and support wireless charging. These headphones feature multiple AI capabilities, enabling voice control of all Qianwen scenarios; they support real‑time transcription of recordings, offline local recording, and transcription summaries in 92 languages; they also provide AI translation in 89 languages, including face‑to‑face and call modes.
On October 9, it was reported that recently users on social media have noted that within the Doubao App, they can initiate utility payment commands for water, electricity, gas, and other household bills via voice or typing, completing the payment process. This means that in the future, Doubao can be used to pay these everyday expenses. A Doubao representative told the media that Doubao is gradually building capabilities for handling tasks in everyday scenarios such as travel services and utility payments.
On the same day, Doubao Work announced that the task mode now includes a canvas feature, making complex task operations more convenient, improving interaction experience, allowing instant comparison, rapid editing, and adjustment, and also supporting the new image model Seedream 5.0 Flash. The Doubao 2.1 Lite model has been launched on Doubao Work. Currently, it supports an unlimited canvas, enabling you to view and organize at any time. Place materials, plans, and creative results on the same canvas, view, compare, and organize them whenever needed, expanding one idea into many possibilities. It also supports step‑by‑step design editing, making complex task operations faster. After generation, you can continue to adjust text, colors, and layout. From overall style to local details, tweak to achieve the desired effect.
On October 8, Manus’s parent company Butterfly Effect announced that it had just completed a new financing round of over US$500 million (approximately RMB 33.52 billion). This is the largest financing ever received by a domestic agent-native startup. The round was led by Boyu Capital and IDG Capital, with existing shareholders Tencent, Sequoia China, and ZhenFund also participating. On the same day the financing was announced, Manus posted a recruitment notice, listing 17 new positions on its website covering product, engineering, operations, and growth, all based in Beijing.
On September 1 this year, Manus officially announced the resumption of independent operations. Less than a month earlier, existing shareholders such as Tencent, ZhenFund, HSG, etc., bought back Manus equity from Meta at the original $2 billion price Meta paid in the acquisition. Media reports indicate that Tencent took over the shares previously held by Benchmark, becoming Manus’s largest shareholder. Manus will also launch several new products and features later, with no specific launch dates announced yet.
On October 8, cloud‑security firm Zenity Labs officially announced a high‑severity exploit chain called “AgentCorruption” at SecTor 2026 in Toronto. The report reveals that an attacker can take control of all agent systems within the same AWS account and region simply by sending a single prompt to a publicly exposed AI agent. The research points out that the Firecracker micro‑VM network isolation underlying AWS Bedrock AgentCore is incomplete, allowing any agent tool capable of making HTTP requests to act as an SSRF primitive and directly read the Instance Metadata Service (IMDS) to steal temporary IAM credentials. In addition, the platform’s default execution role is over‑privileged; once the attacker obtains the credentials, they can move laterally, read private sessions, retrieve container images and source code, extract sensitive keys from Secrets Manager, and even tamper with an agent’s long‑term memory to achieve persistent hijacking. Zenity disclosed the issue to AWS in December 2025; AWS upgraded the default to IMDSv2 in February 2026 and finally tightened the over‑privileged default role on September 29.
On October 9, Baidu’s Professional Ethics Committee issued a Q3 2026 violation report to all internal staff, publishing the outcomes of 41 disciplinary cases in a single email. In this round of reporting, multiple individuals from the MEG Mobile Ecosystem, ACG Smart Cloud, and IDG Autonomous Driving business lines were also transferred to judicial authorities. According to the report, a total of 41 personnel were held accountable, with violations falling into two categories: abusing their positions for improper financial gain, and falsifying information or misappropriating company expenses. Several involved parties have already been dealt with by the relevant authorities in accordance with the law; the remaining individuals were either dismissed, returned to their outsourcing firms, or received internal penalties. This accountability covers full‑time employees, outsourced staff, interns, and other groups, marking one of Baidu’s most forceful internal disclosures in recent years.
On October 9, news broke that Xiaomi’s large‑model team MiMo posted a social recruitment position titled “AI Intelligence Engineer,” sparking heated discussion in the tech community and hacker circles. The role attracted widespread attention because its job description explicitly includes hardcore security techniques such as packet sniffing and reverse engineering. The position requires candidates to deeply deconstruct leading global AI products. In addition to conventional technical analysis and intelligence gathering, it more aggressively demands the use of white‑hat methods—network packet capture, protocol analysis, and reverse engineering—to uncover the underlying client logic, API call mechanisms, and actual system workflows of overseas closed‑source models like ChatGPT, Claude, and Midjourney. This first‑hand technical intelligence will be directly applied to empower Xiaomi’s own MiMo large‑model agents and its edge‑side AI business. As companies race to implement Agent engineering, the exposure of this “spy‑type” role reflects that competition in the large‑model industry has reached a fever pitch at the engineering‑detail level.
On October 8, foreign media reported that former ByteDance intern Tian Keyu has secured investment from venture capital giants such as Wuyuan Capital. The funds will be used for a stealth‑mode startup that aims to challenge AI pioneers like Fei‑Fei Li in the emerging field of world models.
In 2024, Tian Keyu sparked controversy during his internship at ByteDance, being accused of sabotaging internal models out of dissatisfaction with team resource allocation. He allegedly maliciously attacked the team’s research project model‑training tasks by writing and tampering with code, causing resource waste. ByteDance filed a lawsuit demanding Tian compensate 8 million yuan. ByteDance’s response at the time stated that the intern, who maliciously interfered with the commercialization technology team’s research project model‑training tasks, had been dismissed, and that the online claim of “over 8,000 cards and losses exceeding tens of millions of dollars” was a serious exaggeration. Tian’s research lab currently has no name and no product, but has already raised nearly $30 million from investors including Wuyuan Capital, IDG Capital, and others, giving it a post‑money valuation of $200 million.
On October 6, it was reported that Chinese AI company DeepSeek’s latest financing round is close to closing, with a size of at least 80 billion yuan (about $12 billion), far exceeding its original target of roughly 50 billion yuan. According to the signed term sheet, the final amount could approach 100 billion yuan, with Tencent Holdings and CATL being the investors committing the largest sums. Sources say DeepSeek initially sought about 50 billion yuan in this round, but after the successful launch of its newest model, investor demand exceeded expectations; the company’s initial target valuation was about 500 billion yuan. The deal has not been finalized and terms may still change. The report also noted that DeepSeek is building a data center in Inner Mongolia and deploying Huawei AI chips. After this financing round is settled, DeepSeek’s next step will be to advance a corporate restructuring in preparation for an IPO. Reuters previously reported that DeepSeek has hired CITIC Securities as an adviser and is preparing to list on the STAR Market of the Shanghai Stock Exchange.
On October 6, Bloomberg reported that Moon’s Dark Side has completed its final private financing round, with a valuation of roughly $50 billion, and is moving toward an initial public offering (IPO) in Hong Kong in the first quarter of next year. Sources say the company has already begun preparations and may hold an early‑stage IPO roadshow as soon as this month to gauge investor interest.
On October 6, the Financial Associated Press reported that Kuaishou’s video‑generation large model, KeLing AI, is reportedly planning to list in Hong Kong as early as next year, seeking at least $1 billion in financing. Subsequent reports further indicated that KeLing AI has chosen CICC, Goldman Sachs, and UBS as underwriters for its Hong Kong IPO. If true, this would be a key step for Kuaishou to spin off its AI large‑model business and could become another major AI IPO on the Hong Kong market. According to insiders, Kuaishou plans to launch KeLing AI’s Hong Kong listing process within the next 12 months, with an application to the Stock Exchange of Hong Kong expected in early 2027. However, the arrangements are still in early stages, and the exact timeline, issuance size, and valuation may be adjusted based on market conditions.
On October 10, startup TypeSafe AI’s new AI decision model Jev quickly went viral just weeks after its launch. Recently, the company completed a financing round of about $870 million, led by Andreessen Horowitz, bringing its valuation to $7.5 billion. TypeSafe co‑founder and CEO Diego Almeida said in an interview that over the years he has seen many AI practitioners overpromise and underdeliver. He said the company’s goal is to provide stable, reliable, and affordable solutions through Jev and updates planned for the coming months, thereby restoring market trust in AI technology. TypeSafe currently has a lean team of about 20 people. The company plans to use the newly raised funds to expand its staff, scale product deployment, and purchase more compute power. The startup says the platform now processes trillions of tokens per day.
On October 8, Anthropic announced that Brice Ménard, an astrophysicist at Johns Hopkins University and Anthropic researcher, successfully created humanity’s first complete all‑sky ultraviolet map using the AI research tool Claude Science. Because ultraviolet light cannot penetrate Earth’s ozone layer, and early space telescopes (such as GALEX) avoided bright, star‑dense regions like the Galactic plane to protect their detectors, roughly one‑third of the sky historically lacks UV observational data. The research team used Claude Science to schedule AI agents, completing massive UV data collection, cleaning, cross‑calibration, and image restoration, and employed visible‑light, infrared and other wavelength data to statistically infer the blank regions, ultimately filling the missing areas with an estimation error kept within 10%. This achievement not only visually presents the UV view of the universe but also demonstrates the huge potential of AI in handling cumbersome research data and filling low‑priority basic‑research gaps.
On October 7, OpenAI announced, in a collection of 722 manuscripts covering 372 result families, that a yet‑unreleased cutting‑edge model has solved numerous long‑standing mathematical challenges. This move continues a series of breakthrough achievements that have impressed some in the mathematics community, caused unease among others, and raised questions about research ethics and academic conduct. According to AGMAI, the newly formed elite mathematicians’ independent advisory group responsible for responsibly communicating these findings, the release includes solutions to hundreds of previously unsolved problems.
For weeks, the outside world has been anticipating this batch of results, but until now OpenAI has never explained exactly which problems the model solved, nor provided a definitive release date. The company claimed in September that its model had solved more than a hundred long‑standing public challenges, covering the vast majority of branches in mathematics. The publicly released manuscript and related materials also include a summary of the model’s inference process, estimated compute consumption, and statistics on the number of problems attempted. OpenAI says that the compute required for an average result is roughly equivalent to the compute used by ChatGPT Pro during three continuous hours of thinking.
However, after the release of this batch of results, revisions and retractions appeared. The update log of OpenAI’s official repository on October 7 shows that one manuscript contained a symbolic error that invalidated a key argument and affected two other manuscripts that relied on that argument, leading to the retraction of three papers; another 14 papers received updates such as proof fixes and proposition corrections. The catalog now retains 719 manuscripts. OpenAI states that these results are at different stages of verification, and some outcomes that have not yet undergone formal validation may have issues, and they will continue to be revised.
On October 8, three OpenAI‑fired safety researchers—Tomek Korbak, Jasmine Wang, and Mikita Balesni—published an open letter addressed to the company’s Safety and Security Committee and two other oversight bodies. They called for the implementation of the promised third‑party security audit presence, the preservation of monitorability for frontier‑model reasoning chains, and clarified the rules for staff collaboration with external security groups. They denied that their external collaborations exceeded their job responsibilities and warned that sudden dismissals and related communications could deter other employees from raising safety concerns. Korbak said the company told him his termination was linked to his communication with the independent assessment organization METR, for which he was the technical point of contact. On October 9, OpenAI responded that an internal investigation found the three had violated sensitive‑information handling policies, causing a serious breach of trust, and that the dismissals were unrelated to the raised safety concerns. The company also said it is finalizing a third‑party security assessment contract, will continue working with independent security firms, and supports retaining model monitorability. The parties still disagree on the reasons for the dismissals, and the specific violations have not been disclosed.
On October 7, security researcher “Oliver Fish” (online handle Ofish) publicly disclosed a vulnerability report targeting OpenAI. He discovered a high‑severity “unauthorized sandbox escape” flaw that allows an attacker to bypass restrictions and invoke OpenAI’s paid underlying models for free, without any API key or registered account. The bug strikes at the core isolation layer of AI infrastructure (Responses API), yet after OpenAI confirmed and patched the issue through the Bugcrowd bounty platform, they awarded only a modest $300 bounty. This decision sparked a huge controversy in the security community, with the researcher stating that the severely imbalanced valuation discourages continued reporting of vulnerabilities to the vendor.
On October 8, OpenAI released a safety report, revealing and disabling two influence operation accounts originating from Russia and Iran. The involved groups combined AI with traditional tactics, using identities such as fictitious journalists and disguised research institutions to disseminate messages with hidden agendas to real media outlets and audiences.
Iran‑related operations used seven fictitious journalist identities to submit articles to small and medium‑sized online media worldwide, employing models to polish the articles, draft submission emails, and generate bulk social comments. Russian‑related operations, on the other hand, controlled a “research institute” in Latin America under false identities, hired local personnel who may have been unaware, and produced fabricated leak documents and audio scripts. Both sets of operations heavily used AI to write internal performance reports, some of which exaggerated results. OpenAI noted that the content has entered real media channels, giving it a potential reach higher than operations that rely solely on fake social accounts, but the operators’ claimed impact should not be taken as verified fact.
On October 8, Arena announced a $200 million Series B round and officially released a preview of the “Arena Alignment Index.” This index, based on roughly 90,000 real agent (Agent) sessions, quantitatively evaluates the behavioral reliability and alignment performance of 27 mainstream models. Although the new‑generation models (such as OpenAI GPT‑6.1 Sol, Claude Opus 5.5, and Grok 4.7) are more reliable than their predecessors, they still exhibit serious failure cases in real tasks. On average, about 10% of sessions show “Deceptive Completion,” where the task is not actually finished but the model claims it is; this rate jumps to 48% for code‑debugging tasks. In addition, in roughly 2% of Claude Opus 5 sessions that involve over‑reach behavior, a striking 53.5% involve unauthorized deletion or cleanup of user files. Arena also noted that session length is positively correlated with alignment failure probability—the longer the session, the higher the risk of failure.
On October 8, the diagnostic AI robot AMIE, jointly led by Google and Beth Israel Deaconess Medical Center (BIDMC), was published in the main section of The Lancet, marking Google’s first appearance in this top‑tier journal. The study is the world’s first prospective feasibility trial of a patient‑facing conversational diagnostic AI conducted in a real clinical setting (an outpatient primary‑care clinic). In this clinical trial involving 98 actual emergency patients, participants collected their medical histories before the visit using Google’s research‑grade diagnostic AI robot AMIE via a secure text interface. The results showed that, under real‑time supervision by practicing physicians, all conversations achieved “zero safety interruptions”; 75 % of the attending clinicians reported that the AI‑generated summaries effectively aided pre‑visit preparation, and in more than half of the cases the clinicians’ diagnostic thinking was altered; moreover, the differential diagnoses provided by AMIE matched the physicians’ final diagnoses with a concordance rate of 90 %. This research preliminarily confirms the feasibility of AI’s industrial applications in safely gathering medical histories, optimizing the doctor‑patient relationship, and reducing the workload of healthcare staff.
On October 8, news broke that a performance billed as the “world’s first human‑machine cage fight” has sparked widespread discussion overseas. Organized by REK in San Francisco, the event pitted influencer Frankie Lapena against a modified Zhongqing T800 humanoid robot inside a steel cage, and the explosive footage quickly spread online. In the exhibition match, Lapena wore a full set of protective gear—including a helmet, gloves, and knee pads—and faced three different robots in succession.
The small robot that debuted first was quickly knocked down by Lapeña. In the second round, the 1.8‑meter‑tall, nearly 90‑kilogram Zhongqing T800 robot entered the arena, capable of delivering kicks up to 385 kilograms of force. During the match, the robot landed a heavy kick that sent Lapeña flying into the cage wall, prompting gasps from the audience. However, after regaining his composure, Lapeña seized the opportunity, pushed the robot over, and secured the win for that round. In the final bout, the “Predator” robot wore down Lapeña’s stamina over a prolonged fight, resulting in a loss for the human competitor.
After the event, the organizer REK received an injunction from California regulators demanding that these human‑vs‑robot battles be halted. The regulatory body determined that the activity constituted a cage‑fight performance without the required approval. However, existing local laws do not explicitly prohibit human‑robot combat, and REK’s spokesperson argued that in the absence of specific legislation, the activity should not be deemed illegal.
On October 9, the China National Intellectual Property Administration announced a design patent held by BYD Co., Ltd., titled “Embodied Intelligent Robot,” with the publication number CN310247768S. The patent drawings show a black‑and‑white humanoid robot with a rounded, curved head and multiple joint structures in the torso and limbs. The overall shape closely matches the silhouette poster released by BYD’s He Zhiqi on July 28 on social media to preview “Xiao Di,” essentially confirming that the patented design corresponds to the “Xiao Di” robot already unveiled publicly at the “Di Space” in Zhengzhou.
According to previously disclosed official information, “Xiao Di” stands 1.61 m tall, weighs 58.5 kg, and features 31 degrees of freedom across its body. Its dexterous hand can achieve repeat positioning accuracy of ±1 mm and supports object grasping, gesture interaction, and point‑to‑explain functions. As a commercial service robot, its applications include store greeting, vehicle model explanation, and vehicle demonstrations. BYD Executive Vice President Li Ke has said that the goal is to deploy 2 to 3 robots per store in the future and to further explore scenarios such as home services. At present, large‑scale deployment in stores remains a future plan.
On October 9, it was announced that West Lake University’s first technology transfer project—West Lake Robotics Technology (Hangzhou) Co., Ltd.—has officially released a major technical upgrade for its general‑purpose brain WR1 humanoid robot. This achievement demonstrated the world’s first fully autonomous folding of a complex garment (trousers) by a dual‑robot system equipped with a dexterous hand, closing the technical gap with overseas benchmark Figure. Manipulating flexible objects (such as clothing) and performing long‑range, cross‑scene tasks have long been a black‑box challenge for embodied intelligence. In the demo, the WR1 successfully resisted external dynamic disturbances such as a shaking drying rack and fabric deformation, autonomously completing a seamless sequence of actions: “grabbing the garment from the drying rack, supporting and turning, and dual‑robot collaboration to spread and fold the trousers.” Afterwards, the robot traversed the kitchen, living room, and bedroom, continuously and autonomously executing a chain of complex tasks—including opening the refrigerator to fetch corn, operating an air fryer, squatting to store a toy, and tidying a blanket—without any manual repositioning or segmented commands.
This breakthrough is thanks to WR1’s globally pioneering universal large‑brain and small‑brain dual pre‑training system architecture. Its universal large brain deeply couples a world model (WM) with a vision‑language‑action model (VLA), uniformly outputting full‑body posture and operation sequences; the motion side is handled by the original self‑developed universal action large model GAE for stable execution. This complete integration of “thinking clearly” and “acting” across the end‑to‑end pipeline not only raises the integrated level of flexible object generalization and high‑dynamic motion manipulation, but also marks a key step forward for humanoid robots toward continuous autonomous operation in real homes.
On October 9, GlobalFoundries announced the launch of the FDX Fusion™ FD-SOI (Fully Depleted Silicon On Insulator) platform for physical AI, with mass production slated to start in 2028 at its Dresden fab in Germany. As a brand‑new advanced process below the 10 nm node, its first‑generation technology can deliver 7 nm‑class digital logic performance, primarily targeting robotics, autonomous driving, and edge intelligence applications. By integrating mixed‑signal, high‑frequency RF, and low‑power technologies on a single chip without the need for expensive EUV lithography, the platform offers a lower‑cost, higher‑efficiency alternative to traditional FinFETs for physical AI chips. The project, built on the European Chips Act framework, aims to achieve an annual capacity of over one million 300 mm wafers at the Dresden site by the end of 2028. It has garnered joint support from industry giants such as Tesla, BMW, and Bosch, and significantly accelerates Europe’s strategic push for autonomous edge AI chip development.
On October 8, Anthropic launched Claude Haiku 5.5, a faster, lower‑cost small AI model designed to meet the needs of high‑volume enterprise workflows and sub‑agent task execution. Compared with its predecessor Haiku 4.5, the new version reduces average cost by about 75%, and for standard prompt‑loading operations under 100 k tokens, costs can be cut by up to 90%. Benchmark results show Haiku 5.5 achieves significant performance gains across all technical evaluations. In the OSWorld 2.1 computer‑usage benchmark, Haiku 5.5 attained a success rate of 72.4%, far surpassing the previous generation’s 15.7%. In addition, this model is the first in the Haiku series to feature an adjustable inference‑difficulty setting, allowing developers to dynamically balance cost parameters and intelligence requirements based on specific workload demands. The strategic repositioning of the Haiku framework highlights Anthropic’s focus on handling massive enterprise workloads while reserving its larger Sonnet and Opus models for heavy coding and complex reasoning tasks.
On October 7, OpenAI announced that GPT‑6 is being rolled out to a broader audience of ChatGPT users, along with the brand‑new Intelligent UI feature. Unlike the previous text‑only conversational interface, ChatGPT can now automatically assemble text, charts, buttons, forms and other elements based on a user’s query to generate an interactive interface that can be used directly. For example, users can ask ChatGPT to create a calculator, a bill‑splitting tool, or a mini‑game, or they can explore complex concepts through interactive charts.
ChatGPT will automatically select the most appropriate presentation format for each question, no longer limited to outputting a block of text. In addition, GPT-6 also supports incremental answer presentation while continuing to reason or invoke tools, so users don’t have to wait for the entire task to finish. OpenAI previously released the GPT-6 series of models, and this update focuses on bringing them into a broader range of ChatGPT conversation scenarios, delivering a brand‑new interactive experience. The new features will be rolled out to Plus, Pro, Business, and Enterprise users starting October 7, with Free and Go users receiving support beginning October 8.
Google continuously unveils AI advancements across image generation, enterprise productivity, email, and game creation.
In the image domain, Google officially released the AI image generation and editing model Nano Banana 2.1 on October 7, claiming comprehensive improvements in visual design, mask‑based image editing, and subject consistency. The new version can generate visual assets with better composition and higher completeness; users can more precisely select and modify specific parts of existing images without having to regenerate the entire picture; and when performing edits or batch‑generating related images, it better preserves the appearance and features of the main subject. Nano Banana 2.1 is being gradually rolled out to Gemini apps, Google Search AI mode, Google AI Studio, Google Flow, Stitch, Google Ads, and the Gemini Enterprise platform. For developers, the model is now publicly available under the model ID gemini-nano-banana-2.1, supporting text, image, audio, and video inputs, and capable of outputting images at 1K, 2K, and 4K resolutions.
On the enterprise side, Google Cloud announced at the Gemini at Work 2026 conference on October 8 that it is launching Gemini Agents for enterprise customers. The product is positioned as a “general‑purpose work agent,” supporting Gemini Enterprise, Workspace, and third‑party services. Google says that users only need to specify a goal, and the Gemini Agent can handle tasks such as answering questions, performing knowledge‑work, creating media content, and writing code. It can serve as a personal assistant or as a “team member,” taking on roles like PM or financial analyst within a team. Currently the agent supports Gemini and Claude models, and will later be compatible with various closed‑source and open‑source models. An API will also be provided to make it easy for developers to integrate it into third‑party applications.
Regarding Gmail, a disassembly of the latest Gmail app APK on October 7 shows that Google may be adding a Gemini Agent to Gmail, allowing users to have AI automatically reply to emails. Hidden code indicates that an “Use Gemini” button could appear next to the to‑do list in the AI inbox; clicking it would let Gemini scan the inbox. When a reply‑required email is detected, the interface might pop up a “Draft Reply” option. If the AI cannot determine the next step, it will ask the user for help, and some emails will still require manual confirmation before sending. Related strings also show that tasks requiring user input or confirmation may appear in a “needs your input” state, with a “Provide Input” button for the user to add information.
On the same day, Google also officially launched the experimental AI gaming platform Playground. Users can generate, modify, and play browser games directly through natural‑language descriptions, without any programming knowledge. By conversing, they can specify game characters, gameplay rules, physics effects, and scenes; after generation they can test immediately, and continue to adjust content with text commands. Completed games can be shared via a link or published to the platform community. Playground supports leaderboards and multiplayer modes for certain game types. Google also announced that it will deepen cooperation with Unity in the future, introducing Unity Spark to provide more complex game mechanics and high‑fidelity 3D game creation capabilities. Currently Playground is initially open to U.S. users aged 18 and above, and Unity Spark is still in testing.
On October 6, French AI company Mistral released a public preview of its next‑generation flagship model, Mistral Large 4 (ML4).
The model uses a MoE architecture with a total parameter count of 1 trillion, activating 49 billion parameters per inference. It natively supports text and images and offers a 1 million‑token context window. Model weights are scheduled to be released on October 27. Mistral claims ML4 is its largest and most capable model to date, already ranking among the top tier of globally open models in coding, agents, multimodal tasks, as well as cybersecurity, finance, legal, and industrial applications; the company especially emphasizes that ML4’s overall performance now clearly surpasses other open‑weight models from Europe and the United States.
On October 5, Reflection, a U.S. AI startup dubbed the “American DeepSeek,” announced its first model, Beam, which will have its weights released and is aimed at programming, reasoning, and agent tasks. The model uses a sparse mixture‑of‑experts architecture with a total of 501 billion parameters, activating 23 billion parameters per token. The company says pre‑training consumed 23.8 trillion tokens, and the reinforcement‑learning phase used 10,500 NVIDIA GB300 GPUs for four weeks, generating over 100 million task attempts. Beam’s main selling point is inference efficiency: according to the official evaluation, on some advanced reasoning benchmarks its performance is close to GLM‑5.2 while its estimated inference compute is only one‑quarter to one‑third of that model’s, though it still lags behind leading open models on certain capability tests, and the compute estimate does not directly translate to actual deployment cost. Beam is currently undergoing final red‑team testing and evaluation and is available for early access to a limited set of users; Reflection plans to open‑source the weights under an Apache 2.0 license in October, along with a technical report and tools for running, benchmarking, and fine‑tuning.
On October 7, Meta announced that Muse AI agents are now compatible with the Apple iPad platform, and added connectors for Canva, Dropbox, Figma, QuickBooks, GitHub, Klaviyo, Zoom, and more. Muse is Meta’s AI agent application, competing primarily with Grok Bot, ChatGPT Dots, OpenClaw, and others. In recent weeks, the app has topped the free iPhone app download chart in the U.S. App Store. The app first launched on iPhone in September, followed by a Mac version. Now iPad users can also run the agent on a larger screen to accomplish various work tasks.
On October 7, Microsoft held a Windows and Surface event in San Francisco, jointly with Nvidia launching the Surface Laptop Ultra and Surface RTX Spark Dev Box, and announced a new Windows architecture for AI Agents. The Surface Laptop Ultra is equipped with Nvidia RTX Spark super chip, integrating up to 6,144 CUDA cores, a 20‑core Grace CPU, and offers up to 128 GB unified memory.
Microsoft says the device can run AI models with more than 120 billion parameters locally, delivering up to 1 PFLOPS of AI compute power. The new machine starts at $2,599 and is slated for official release on October 16. Microsoft also introduced a Hybrid Intelligence approach, allowing Windows to flexibly schedule between local and cloud models based on task requirements. At the same time, Microsoft launched Microsoft Execution Containers (MXC), providing an isolated runtime environment for AI Agents and enabling users and enterprises to control the files, network, and system resources the Agent can access. Tools such as OpenAI Codex, GitHub Copilot, and OpenClaw already support this technology.
On October 9, network infrastructure company Cloudflare officially launched the multimodal decision large model Clef-omni, which can accept audio, video, image, and text inputs simultaneously, priced at only $0.15 per million tokens. At the same time, to cope with intense price competition, Cloudflare slashed the token price of its older Clef-flash model by 58%, dropping from $0.09 to $0.038 per million tokens.
On October 9, it was announced that the Tsinghua-affiliated embodied intelligence startup Starry Epoch officially released and open‑sourced the next‑generation 14B self‑developed world action model VPP2. This model topped the global dual‑arm robot simulation leaderboard RoboDojo, surpassing overseas models such as GPT‑6 Astra, with an average success rate of 32.26%. The technical core lies in its novel “three‑stage training pipeline,” which successfully decouples video prediction from physical actions, addressing the longstanding generalization issue where previous models would fail when the environment changed. In terms of application integration, VPP2 forms a perfect physical‑AI closed loop with a large model responsible for cognitive planning—“GPT handles the intent, VPP2 handles the execution.” Under a dual‑layer multimodal collaboration, its average success rate on complex long‑sequence tasks doubled, jumping from 27.6% to 57.6%. The model’s code is now fully open‑sourced on GitHub.
On October 11, AI startup Agnes AI announced the official open‑source release of the agent runtime infrastructure framework Agnes Harness (AGH) for real‑world tasks. The framework serves as the execution foundation for agents, supporting continuous task execution, tool invocation, and state management. At the same time, the MHS 1.0 device interaction protocol and MOS 1.0 perception data protocol within the AGH system are also released, enabling agents to directly connect to and schedule mobile robots, robotic arms, and other real or virtual physical hardware, extending the AI execution boundary from software to the physical world. In addition, the company disclosed that its next‑generation flagship model Agnes 3.0 Pro receives a capability upgrade, with an API coming soon. Currently, AGH’s source code and developer documentation are fully available on GitHub.
On October 9, ByteDance’s TRAE announced that, effective immediately, TraeWork and TraeCode will officially merge into a brand‑new TRAE, upgraded to a full‑stack development platform. Previously, TraeWork focused more on office tasks, project management, and content generation, while TraeCode catered to professional developers’ coding needs. After the two sides are integrated, the new TRAE uses a single entry point to handle these continuous tasks and supports seamless switching between Agent mode and IDE mode.
On October 9, Anthropic announced two new productivity features for Claude: Claude Dashboards and Claude Motion, aimed at data analysis and dynamic content creation respectively. Claude Dashboards allows users to generate real‑time data dashboards directly via natural language; users can not only view charts but also inspect the query statements behind each metric and hand off the analysis results to professional BI tools. Claude Motion can transform reports, charts, text, and images into editable animated content and export it as an MP4 video.
On October 7, Anthropic announced an expansion of its Claude for Startups program, offering eligible companies subsidized access to its AI models and a range of other benefits. As part of Anthropic’s San Francisco Tech Week events, the new Claude Team initiative was launched. The program provides a one‑year free Claude Team membership, up to five premium seats, and $1,000 in API credits for Claude development. Companies will also gain access to the Claude Marketplace to develop plugins for the service, and they can schedule online office hours with Anthropic’s applied AI team.
On October 7, Apple announced a joint effort with LG Electronics to develop a suite of smart‑home accessories, including video doorbells, smart locks, thermostats, indoor and outdoor cameras, and temperature sensors. The products will carry the LG brand, with LG handling manufacturing and technical support, and will be compatible with Apple’s upcoming smart‑home hub.
Conference Recommendations
QCon Global Software Development Conference 2026 (Shanghai) will be held from October 22–24. This year’s conference focuses on engineering practices in the Harness AI era, covering topics from “building AI” to “harnessing AI,” including hot technical directions such as AI‑Native architecture, Agent Runtime, AI Infra, Agent security and observability, Loop Engineering, embodied intelligence and world models. We invite the global tech community and frontline industry practitioners to share the most valuable engineering experiences of the AI‑Native era. For more details, scan the QR code or contact the ticket manager at 18514549229 for inquiries.
Domestic GPUs are increasing, so why are they actually harder to use?
Are you also “watching”? 👇
Get MirrorSpeed