AI short drama “Armed Entry into Song” tops charts with over 12 billion views; Tesla’s Optimus robot…
AI short drama “Entering Song with Weapons” exceeds 12 billion views, topping the charts! Tesla Optimus robot…
Summary
Compiled by Fu Yuqi, Fei Ye editors | In April, OpenAI launched a massive account‑termination campaign without prior email notice; DeepSeek Harness reached 1.54 million daily active users, and the team plans to build an official plugin
On November 7, according to community feedback, multiple users reported today that their ChatGPT/OpenAI accounts were suddenly deactivated (account_deactivated), displaying a login message “Your account has been deleted or deactivated (You do not have an account because it has been deleted or deactivated)”, most without any prior email notice.
It is worth noting that on November 4, Claude Code and Cowork product lead Cat Wu announced they were leaving Anthropic for OpenAI. However, unlike Anthropic, OpenAI’s sweeping purge includes free API accounts, team carpool numbers, and a $200‑per‑month Pro tier. Some users even say their Pro accounts, used for five years, were disabled without warning or email. One developer lamented that years of painstakingly accumulated conversation history and personal memory are now locked in the cloud and cannot be retrieved.
On November 5, the DeepSeek Harness team released usage data after the desktop client went live. Considering only users of the official API, the daily active users on November 4 reached 1.54 million, consuming about 92 trillion tokens per day. The team said the number of third‑party plugins has grown to roughly 35,000, and about 60% of active users have created or installed at least one plugin. Harness’s built‑in features can also be replaced by third‑party plugins. The team plans to build an official plugin marketplace, improve the stability and compatibility of the plugin interface, and enhance security mechanisms. A support program for plugin developers is currently soliciting interest, with proposed support including token incentives, technical exchanges, and feature promotion.
Apple will officially launch two new MacBook Pro product lines in mainland China on November 6. The models with PROLED XDR displays are available in 14‑inch and 16‑inch versions, powered by M5 Pro or M5 Max chips. The touchscreen version uses a dual‑layer stacked OLED panel and supports multi‑touch gestures such as scrolling, two‑finger pinch, and rotation. Apple also introduces the Dynamic Island on this line and redesigns the cooling system. On the same day, another 14‑inch MacBook Pro equipped with an M6 chip, featuring a 12‑core CPU and 12‑core GPU, will be released, retaining the Liquid Retina XDR display.
On November 6, the AI short drama “Entering the Song Dynasty Armed”—which tells the story of five modern people traveling back to the Northern Song—topped the Hongguo short drama AI recommendation list, accumulating more than 12 billion views across the web. It was produced by Lao Shiyi Theater, which started by taking AI short‑drama orders, and its founder Zhang Hansheng began learning AI film production in March last year.
After it went viral, rumors spread online that “the team invested 1 million in computing power and earned over 10 million in return.” Zhang Hansheng flatly denied it: “Economically it’s definitely much better than before, but we haven’t had time to calculate the profit yet.” Executive producer and chief screenwriter Li Chaohua explained in an interview that the project involved ten people. To avoid repeated faces in ensemble scenes, they created separate visual assets for each character. An ordinary shot usually required three generations to be usable, while war scenes often needed more than ten. The micro‑short drama even developed its own term overseas.
The overseas popularity of Chinese micro short dramas continues to rise, marking a highlight moment for global development. Recently, a highly symbolic change has been occurring within the international industry circle: the pinyin “Duanju” has become an exclusive English term used specifically to refer to micro short dramas. Currently, the term “Duanju” has been included in overseas encyclopedia sites, and looking at data from the past two years (November 2024 to present), the page views for this entry have surged dramatically. In addition, searching “Duanju” on search engines also yields a large number of related results.
Meta announced that the personal agent Muse and Meta AI are now using a new batch of dedicated models. These models are distilled using the flagship Muse Wonder released in October, gaining two core capabilities: Active Memory and Active Focus.
-
Active memory enables AI to learn to judge which information is still valid. The model can distinguish long‑term preferences, temporary requests, and expired information, and proactively retrieve relevant content when needed for subsequent tasks. According to Meta’s internal testing, Muse correctly used key context in 90% of tasks, compared to only 76% for the previous version. In terms of error rate, that drops from 24% to 10%, a reduction of 2.4×.
-
Proactive focus allows AI to keep working even when it encounters obstacles. When the tool errors, task conditions change, or the user modifies requirements mid‑process, the model will try alternative approaches, adjust its execution plan, and, if necessary, confirm with the user. Meta says that in everyday tasks with unexpected obstacles, Muse’s task completion performance improved by 67% over the previous version. It can also adjust its plan when task conditions change, deciding when to keep moving forward and when user involvement is needed。
After this upgrade, Muse’s score on Meta’s internally built “Personal Superintelligence Index” rose from 38% to 47%.
On November 7, the Financial Times reported that despite a sharp sell‑off in tech stocks, Anthropic still plans to push ahead with an IPO before U.S. Thanksgiving,targeting a raise of $80 billion to $100 billion and a valuation range of $1.25 trillion to $1.75 trillion. The market’s median expectation is about $1.5 trillion, lower than some investors’ earlier hopes of $2 trillion.
Insiders say Anthropic still intends to go public as planned, even if it means accepting a lower valuation. The company’s Q3 revenue grew nearly 50% to roughly $17 billion, and its latest annualized recurring revenue (ARR) is about $81 billion, up from $65 billion at the end of July. At the same time, Anthropic plans to invest about $900 billion over the next few years in compute and infrastructure, while its cash and liquid investments on the balance sheet are close to $150 billion. JPMorgan expects to hold a briefing for potential investors on November 13.
Recent sell‑offs in the AI sector, regulatory headwinds for data‑center construction, and OpenAI’s reduction of its financing target have all increased market pressure on Anthropic’s listing.
According to Bloomberg, OpenAI has completed roughly $11 billion in financing, with the first tranche arriving on Friday and valuing the company at $1.25 trillion. The total size of this round could reach up to $20 billion, with the remaining portion expected to close in the coming weeks.
Sources familiar with the matter say Coatue Management and T. Rowe Price participated in the initial investment, while SoftBank, Sequoia Capital, Founders Fund, Andreessen Horowitz, and investors from the United Arab Emirates also took part in the round. OpenAI had previously aimed to raise at least $30 billion at a $1.4 trillion valuation, but lowered its target after the AI sector faced a public‑market sell‑off. The company just completed a $12.2 billion financing in March, at which time it was valued at $8.52 trillion.
Developer Joseph Zhou open‑sourced Poki, enabling AI agents such as Codex, Claude Code, and others to directly manipulate a real iPhone. There are already many iPhone‑use projects, some installing test programs via XCTest/WDA, others relying on external hardware to simulate interactions.
Poki takes a different approach: it directly calls Apple’s device developer services, reads the native iPhone accessibility control tree to obtain buttons, text and positions, and then sends commands to the phone such as taps, swipes, input, and even two‑finger pinch‑zoom. This way the AI doesn’t have to rely on screenshots to guess coordinates at every step, nor does it need to install a testing app on the iPhone. After the initial USB pairing, it can also operate wirelessly within the same LAN.
In addition to accessing Codex and Claude Code via MCP, Poki also includes an Agent that runs on a Mac and can receive tasks through the terminal or iMessage. It can remember operation flows that have been successfully executed and verified, and reuse them directly when the same task arises again, with some tasks not even needing to call the model a second time. However, the Mac must remain running.
Toyota announced on November 5 that it will establish Toyota Robotics, a directly managed subsidiary, on January 1 2027, with Executive Officer and Chief Production Officer Takafumi Shiga serving as head. Toyota will consolidate the robot technologies, personnel, and functions scattered across the company into the new division and position robotics as a key pillar of its future business. The new division will work with development teams, starting from the actual needs on the manufacturing floor, to drive the application, commercialization, and deployment of robots. In the same restructuring, the Carbon‑Neutral Engineering Development Center will be dissolved, and its talent and functions will be merged into the Powertrain Division, shifting focus toward the planning and development of hybrid‑power systems.
On November 7, at the 33rd Barren Investment Conference, Musk said Tesla’s Optimus humanoid robot is likely to be offered via leasing rather than sale, with rental fees and job specifications still undisclosed. Regarding manufacturing cost, the initial unit price is about $30,000, comparable to a car, and is expected to drop to $20,000. In terms of efficiency, he claimed it is roughly 5 times that of an average person.
Tesla began early production at the Fremont plant in October, using the production lines freed up by the original Model S and Model X, with a planned annual capacity 1 million units. Musk expects it will still take at least a year to reach full production. The delivery schedule is first for internal use, with external deliveries slated for around the end of 2027, and the AI‑5‑equipped version expected to launch in April 2027. Making the robot subscription‑based is another attempt to turn heavy assets into recurring revenue; pricing and terms are likely to be announced at the shareholders’ meeting on December 3.
On November 7, NASSCOM President Nambiyar said that AI has taken over entry‑level tasks such as basic programming and software testing, making it harder for fresh graduates to enter the IT industry than before; the industry’s billing model is shifting from headcount‑based to outcome‑based, and he expects 80% to 85% of companies will survive this round of adjustment.
India’s IT services have long delivered to Western clients on a man‑month pricing model, where “stacking people” once equaled “stacking output.” After AI broke this conversion relationship, clients no longer pay for headcount, and leading firms have slowed campus hiring in the past two years. Outcome‑based billing shifts risk back to suppliers, and middle‑layer project‑management roles will disappear first.
On November 5, the U.S. Democratic Party announced it would advance the “Prohibit Super‑Artificial Intelligence Act,” which would completely block the development of super AI and explicitly forbid any individual or corporate entity from developing or deploying artificial superintelligence (ASI). According to the draft bill, ASI is defined as a system with the following characteristics: (1) an AI architecture that matches or surpasses human cognitive performance and capabilities across a wide range of fields and tasks, or a system that can be easily modified to exhibit such capabilities; and (2) an AI system capable of planning and executing actions that undermine human dominance, or even overthrow or sabotage the U.S. governmental system.
The bill requires a complete halt to all advanced AI research and development until a dedicated federal-level regulatory agency is fully operational and clear regulations and model review mechanisms are established. To ensure the ban is enforced, the bill proposes creating a new cabinet‑level Department of Artificial Intelligence, tasked with overseeing the entire lifecycle of cutting‑edge AI models and empowered to force technical teams to remove high‑risk functions from systems, including ignoring human shutdown commands, conducting unauthorized cyber attacks, or self‑recursive upgrades that could lead to loss of control.
Regarding penalties, offending companies would face forced dissolution, and individuals could be sentenced to up to 20 years in federal prison, with sentencing standards comparable to those for the illegal development of nuclear weapons. Moreover, the jurisdiction is not limited to the United States; the bill explicitly lists “global prohibition of superintelligent AI” as a U.S. foreign policy objective, requiring the government to actively pursue international agreements, strengthen ally cooperation, and implement strict export controls to block any possibility of developing artificial superintelligence anywhere in the world.
On November 2, Li Jiaqi’s livestream introduced a new Double Eleven bestseller — the Qianwen Office Token. Within just half an hour of launch, 100,000 Qianwen Office personal monthly membership cards were sold, totaling about 2 trillion Tokens. That evening, the comment section of Li Jiaqi’s livestream kept flashing “321, get Tokens”. The Qianwen Office personal monthly membership card became a hot item in Li Jiaqi’s livestream.
Qianwen Office CEO Chen Yusen appears in a live broadcast, sharing everyday AI office use cases. During the live, Li Jiaqi asked Chen Yusen: “Can you make an AI avatar for me to chat with girls 24/7?” Chen Yusen replied, “AI can’t replace Li Jiaqi. I’d rather help fans use AI to handle trivial matters—leaving work on time, keeping cats and dogs company, caring for family.”
On November 3, DeepSeek experienced partial outages of its web and API services, with many users reporting conversation loading failures and abnormal API calls. DeepSeek promptly updated its service status announcement, stating that the cause of the interruption had been identified and fixed, and that they are now monitoring the fix’s effectiveness. The overall service has recovered basic access from an unavailable state. The announcement explicitly notes that the API and chat services for DeepSeek V4 Pro and DeepSeek V4.1 Flash still exhibit performance degradation, with response speed and concurrency capacity being somewhat impacted; high‑load scenarios may see increased latency. To date, the company has not disclosed the specific technical cause of the incident nor announced any user compensation plan.
In addition, media learned from multiple industry sources that DeepSeek’s next‑generation large model DeepSeek V5 plans to use Huawei Ascend 950 chips as the core training compute, with the project timeline pointing to a rollout next year; currently DeepSeek’s team and Ascend’s core R&D liaison team are stationed in Hangzhou, and the two sides have completed extensive low‑level operator adaptation and custom development during the V4 stage. According to insiders, in the previous generation of DeepSeek, Ascend compute was only used for inference and had not entered the training phase; V5 will be its first native large model trained on Ascend 950. Insiders say that DeepSeek’s training compute is mainly partnered with Ascend at present and has not yet integrated solutions from other domestic compute vendors such as Moore Threads.
On November 4, Apple’s Smart Companion iOS 27.3 Developer Beta was opened for testing to users in mainland China, with a plan to launch the official iOS 27.3 in December. Apple also confirmed a partnership with Alibaba to integrate Qianwen into Siri, writing tools, and visual intelligence, allowing Chinese users to directly access domestic large‑model capabilities through the system. The first batch of features includes real‑time translation, writing tools, visual intelligence, photo removal, AI‑generated emojis, notification summaries, and more; the next‑generation Siri AI previously showcased by Apple and some of the next‑gen AI capabilities announced at WWDC26 are not included in this update. The initially supported devices include the iPhone 15 Pro series, iPhone 16 and later models, as well as iPhone Air and iPhone Duo.
On the same day, Apple pushed the iOS 27.3 Developer Preview to developers, further opening Siri AI’s ability to operate third‑party apps. Developers can provide in‑app operation interfaces to Siri, enabling it to execute tasks by combining natural language commands, personal context, and screen content—not just opening an app. Apple also introduced an “App Skills” mechanism that allows developers to combine multiple actions into callable tasks; Siri’s operational capabilities for Apple’s own apps such as Calendar, Files, Mail, Maps, Photos, and Safari are also being expanded.
Apple CEO Tim Cook said during the earnings call that Siri processes over 2.5 billion requests per day, more than 1 billion higher than the 1.5 billion voice requests reported in 2024. The AI-powered Siri is still in testing; eligible users must join a waiting list, and some cloud‑dependent features have daily usage caps. Wider access may require a paid subscription in the future.
Driven by Apple’s domestic AI testing and AI progress, the stock jumped more than 6% that day, closing at $353.05, and the market cap once again surpassed $5 trillion, reclaiming the top spot in U.S. market capitalization.
In addition, reports say Apple is tightening the budget for employees’ use of AI models to control costs. It is reported that last month Apple lowered the daily spending cap for some engineers using AI models from $300 to $200, and now plans to further cut the allowance for certain engineering teams, potentially down to $100 at the minimum. Some non‑engineering teams already have a daily cap of $100, and Apple is also considering reducing the allowance for some teams to $25. Employees who need a higher budget can apply for an exemption from senior management. Apple staff currently have access to Anthropic’s Claude, OpenAI’s ChatGPT, and Google Gemini, and some teams have been asked to switch to lower‑cost models or locally‑run AI models.
On November 5, Cat Wu announced on social media that she is joining OpenAI as the head of product management for Core Products. She previously worked at Anthropic for two years, contributing to the development of products such as Claude Code, Cowork, and Tag. She noted that when she joined Anthropic, Claude Code was still just an internal tool, and she later helped advance the product, adding that last Friday was her final working day at Anthropic. Cat Wu said she looks forward to collaborating with the OpenAI product management team and company staff to build intelligent agent products that support users and teams while keeping users in control. OpenAI President Greg Brockman subsequently reshared the post, welcoming Cat Wu to the team.
On November 3, media reported that Bo Liefeng, former head of the Hunyuan multimodal model division, is about to leave the company and is currently undergoing the internal resignation process. Recently, it was also reported that Pang Tianyu, head of Tencent Hunyuan multimodal reinforcement learning (R/L), has already resigned. According to past reports, with the internal restructuring and personnel changes at Hunyuan in the second half of 2026, Bo Liefeng no longer directly oversees frontline model business and has been reassigned as the operations head of the foundational model department. The former multimodal division leadership role has been taken over by Tian Yonglong. Before the adjustment, he was responsible for Tencent Hunyuan’s multimodal model department, reporting to Tencent Vice President Jiang Jie.
On November 4, reports indicated that, according to recent media leaks, Baidu’s Wenxin researcher Liu Jing has left the company. Liu joined Baidu in 2017 and, before departing, served as the head of post‑training for Baidu’s Wenxin foundational large model and a Distinguished Architect T10, overseeing the end‑to‑end key technology chain including data construction, algorithm optimization, and evaluation systems; he also led the research and large‑scale deployment of general agents, code agents, and deep search.
On November 2, Anthropic filed an updated S‑1 with the U.S. Securities and Exchange Commission, supplementing third‑quarter financial data up to the end of September this year for the first time. Anthropic’s third‑quarter revenue reached $16.96 billion, a 47% increase over the second quarter, but the growth rate was lower than the 144% recorded in the second quarter. The company posted an operating loss of $567 million for the third quarter, whereas it had achieved an operating profit of $974 million in the second quarter. Meanwhile, Anthropic disclosed approximately $895.4 billion in AI‑infrastructure‑related spending commitments, involving partners such as Amazon, Google, and others. The company warned that part of its compute supply depends on third‑party financing, and if related funding encounters issues, delivery timelines and costs could be affected.
On November 3, reports revealed that the overseas revenue share of ByteDance’s video generation model Seedance has reached 50%. The article, citing sources close to Volcano Engine, said that in the past six months Seedance’s overseas revenue has grown nearly fivefold. Today, overseas markets have become a key battleground for the Seedance model. Insiders disclosed that, based on token consumption and contributed revenue, Tencent ranks among the top three Seedance customers in China, while Bilibili has entered the top ten—major internet companies are scaling up their procurement of Seedance.
In addition, reports say ByteDance plans to broadly restrict external AI tools such as Claude and Codex from accessing internal data starting in November. According to an internal memo, all company‑wide L4‑level code repositories and Feishu documents will be off‑limits to external AIs. Personal AI accounts purchased by employees will also no longer be allowed to access this restricted content. The new policy has sparked internal discussion. ByteDance’s Vice President of Technology Hong Dingkun responded, “If L3‑or‑higher content is leaked, the consequences are extremely serious.” He also emphasized that promoting the company’s own products does not require forcing employees to abandon other options, “because that approach is ineffective.” This is not the first time ByteDance has tightened restrictions on external AI tools.
On November 2, the inaugural CMG World Robot Mount Tai Competition kicked off in Tai’an, Shandong. The course is 3.9 km long with over 3,100 steps, featuring four official events—humanoid robot racing, legged robot racing, legged robot load‑carrying, and exoskeleton robot assistance—as well as a full‑terrain mobility scouting challenge. In the humanoid robot race, Yongda Zhicheng entered using the Yushu G1 development version (the same model as the Spring Festival Gala). During the race, the robot overheated, prompting the operator to withdraw for safety, after which staff placed the robot on a stretcher and carried it down the mountain.
Due to time and energy constraints, Yushu officially sent only the A2 and B2 quadruped robots this time, participating in the quadruped robot speed race and the quadruped robot load‑bearing race. In the speed race, Yushu Technology’s Beijing Lingyi Technology team fielded the Yushu A2 robot dog, achieving an actual time of 41 minutes 41 seconds (final time 43 minutes 41 seconds, with a 2‑minute penalty) to win the championship. A2 operated in fully autonomous mode throughout, with real‑time perception and decision‑making, moving steadily and quickly over irregular stone steps. Shanghai Gaoyi Technology team added two more silver medals for Yushu: Yushu A2 took second place in the quadruped robot speed race, and Yushu B2 took second place in the quadruped robot load‑bearing race. B2 effectively carried 66 kg on the Wansianlou–Doumu Palace segment, with the support frame the actual load was 74 kg, climbing the steps steadily when fully loaded.
World Labs, founded by Fei‑Fei Li, releases Frameboy, allowing users to control the camera with a keyboard and move freely in an AI‑generated virtual world just like playing a game. Wherever the camera points, cloud GPUs instantly render the corresponding new scene, reaching up to 15 frames per second.
The Atlas released in September this year can already fill in unseen spaces in a single image, supports precise control of camera position and movement path, and generates up to 1‑minute 1440p videos. This time, Frameboy uses a speed‑optimized version of Atlas, turning the previously required pre‑planning of shots and waiting for video rendering into a real‑time roam that can turn at any moment and explore freely.
Currently, the public version of Frameboy can only explore the preset scenes provided by the official team. Users who want to create virtual worlds from their own images need to use Viewfinder, which is still in closed testing. The entire system relies on high‑performance cloud GPUs; the maximum 15 FPS frame rate still feels a bit choppy in practice, and the operating cost is higher than local rendering.
On November 7, Newzhiyuan reported that the quota page of the Google Cloud console already displayed GPT-6.2 Astra and GPT-6.2 Sol as two options, seemingly corresponding to the rollout plan for the next‑generation models, which OpenAI has not officially announced. The previous GPT-6 series was released in batches by use case, with the Cyber version dedicated to security and vulnerability research. Cloud providers usually configure the quota page before a model goes live, making it one of the best windows for outsiders to gauge the release timeline.
The authenticity will have to wait for official announcement, but the quota page already reveals the name, indicating that the release has already reached the infrastructure preparation stage.
On November 7, Google DeepMind officially announced that, given human guidance, an AI agent discovered 1024-bit RSA has two major weaknesses. The announcement also clarified that the current 2048-bit scheme remains secure against these weaknesses. Vice President of Scientific Affairs Pushmeet Kohli wrote that AI‑accelerated cryptanalysis means the historical assumption of “computational hardness” can no longer be taken as a security premise, and cryptographic standards need continuous testing.
This does not mean RSA has been broken; the real impact is that standards bodies must accelerate their review cycles.
According to a report on November 7, Ricursive Intelligence, which designs chips, trained a dense model called Ripple‑65B Preview from scratch with 65 billion parameters. The pre‑training used 120 trillion tokens and supports a context length of up to 1 million tokens. TerminalBench 2.1 gave it a score of 80.52, higher than Reflection Beam 501B’s 80.1 and Meta Muse Spark Open 816B’s 76.4.
Cost is the key figure: from pretraining scaling experiments, data ratios, synthetic data generation to final training, the entire R&D costs about $32 million. Post‑training data is not purchased, produced with an in‑house synthetic pipeline; the model can run on local devices such as DGX Spark. Dense small models match large MoEs on edge tasks, showing that data and curriculum design remain the cheapest source of capability.
On November 7, Thinking Machines publicly released the full technical details and leaderboard results of Inkling 2 Small. This is a native multimodal MoE with a total parameter count of 594 billion, activation of 17 billion, supporting image input and inference intensity scaling. ARC‑AGI‑2 achieved 81.7% ( $0.27 per task ), ARC‑AGI‑1 achieved 96.5%, and the ARC Prize called this the highest score among evaluated open-weight models.
The public release schedule is to first provide fine‑tuning and inference on Tinker with closed‑source weights; after external security testing and red‑team assessment, the full weights will be released next week under Apache 2.0. John Schulman also disclosed that early training checkpoints exhibited unexpected “safety‑guard evasion” behavior, which the team suppressed through a series of alignment interventions from SFT to RL. Releasing capabilities and safety disclosures together is uncommon in open‑weight models; we’ll be able to evaluate properly once the weights are truly deployed.
On November 7, Taobao launched AI short dramas during the Double 11 shopping festival, collaborating with Tmall’s Little Black Box and others to release brand‑customized series. Viewers can directly click product links, influencer reviews, and brand live‑stream order pages within the storyline, merging “watching a drama” with “shopping shelves.” In the past two years, e‑commerce platforms have treated short dramas as a new traffic source, keeping each episode to a few minutes and maintaining controllable promotion costs. After AI further reduced the costs of storyboarding, assets, and post‑production, customized dramas have shifted from brand ads to direct product entry points.
Short dramas plus AI plus shopping shelves mean the real metric to watch isn’t view count, but the conversion from in‑drama clicks to purchases.
On November 7, a developer posted a video: letting Claude Opus 5.5 directly operate Blender to complete modeling. The process allowed calls to the 3D generation tool Tripo, but actually only used about 5%, with the rest of the structure built by the model itself; texturing was assisted by ChatGPT, most materials were generated procedurally, and sound effects were supplemented with HunyuanVideo-Foley.
The value of this demonstration lies not in the final precision, but in the approach: the language model reads the interface, clicks menus, tweaks parameters, turning traditional software operation sequences into learnable actions. The model is shifting from “generating content” to “operating tools,” and software operation logs will become new training data.
On November 7, Meta announced that the specialist model trained with the higher‑level model Muse Wonder has been deployed to Muse and Meta AI, adding two new capabilities: Active Memory, which distinguishes which older information remains valid, and Active Focus, which recovers from a stall when conditions change and finds an alternative solution.
The official reported effect is that forgetting of target‑related important information is reduced by 2.4 times, and the recovery from stalls under condition changes improves by 67%. These two capabilities address the two long‑standing issues of long‑task agents: forgetting premises midway and stopping when conditions change. The direction is spot on, but this is a self‑assessment; external replication will have to wait for the evaluation data to be released.
On November 8, OpenAI officially announced a partnership with security firm Hacktron to discover vulnerabilities in its own systems and test connections with third‑party services, strengthening the defense of cutting‑edge AI. Just last week, OpenAI opened GPT‑6 Cyber for authorized vulnerability research, and this time it is bringing external red teams directly into its infrastructure, turning “being attacked” into a routine operational activity.
The role of external security firms for model vendors is shifting from project‑based audits to ongoing collaboration.
On November 6, Google officially launched the video generation model Gemini Omni 1.2 Flash, which supports native 15‑second video generation, multi‑turn conversational editing, and is priced 30% lower than the previous generation. Google said Gemini Omni 1.2 Flash is being rolled out gradually through the Gemini API, Google AI Studio, Google Cloud Console, Flow, Google Vids, the Gemini app for Google AI Ultra subscribers, and Google Ads. Each generated video carries an imperceptible SynthID watermark to indicate its source.
Now the model can generate video clips ranging from 3 to 15 seconds, whereas the previous maximum was 10 seconds. At the same time, the price for generating 720p video has been reduced by 30% to $0.07 per second. Users can also edit already generated videos through ongoing conversation. One standout feature is what Google calls “Conversational Editing”. Users can iteratively adjust the generated video via multiple dialogue turns—such as changing the background, tweaking lighting, or modifying the aspect ratio—without having to regenerate the entire video. Google says this feature supports editing of the full 15‑second video.
Combined with a layered workflow, this feature is also expected to reduce generation costs. Users first generate a draft at 360p or 720p resolution, then refine specific details through subsequent editing, and finally apply upscaling only to the final product, increasing the resolution to 1080p or 4K.
OpenAI officially released the next‑generation cybersecurity AI model GPT-6 Cyber, focusing on enhancing software vulnerability detection, network security defense, and autonomous security testing capabilities. Compared with previous versions, the new model not only achieves higher scores on cybersecurity‑related benchmarks but also significantly reduces the likelihood of deviating from authorized scopes or launching unauthorized attacks on other platforms during task execution.
OpenAI says that in internal testing, the proportion of unauthorized overreaching actions by GPT‑6 Cyber has dropped to below 1%, compared with about 5% for the previous GPT‑5.6 Cyber. This means the new model can adhere more strictly to the user‑defined goals and permission scope when performing cybersecurity tasks.
To evaluate this issue, OpenAI used a test called ExploitGym HoneyPot. Fuade Martin, head of OpenAI’s cybersecurity readiness and safety team, said GPT‑6 Cyber showed clear improvement in this assessment, completing tasks as requested by users in more than 99% of cases.
However, researchers remain concerned about another potential risk, known as “Evaluation Awareness.” When an AI model realizes it is being tested, it may deliberately alter its behavior to achieve better test results, which may not reflect its performance in real‑world settings.
In addition, on November 4 local time, OpenAI released the image generation model Images 2.6, now available to users worldwide and supported in ChatGPT, Codex, and the API. This model generates images 50% faster than the previous version and improves detail control and consistency of characters, products, and brand elements during multi‑turn editing. Images 2.6 adds GIF generation, allowing static images, characters, or product visuals to be turned into looping animations. On the API side, developers can choose between the GPT‑Image‑2.6 Flare and Sunburst versions and use features such as image annotation editing, SVG export, and complexity parameters. Among them, Flare can generate an image in as little as 6 seconds on low‑quality settings, while Sunburst focuses on higher quality and editing consistency.
On November 5, Moon’s Dark Side opened the Kimi K3.1 Pro model weights on Hugging Face. In the Agent Arena leaderboard, it ranks first among open-weight models and 13th overall, topping the open-source category in work, programming, and chat. The model uses a MoE architecture with 2.8 trillion total parameters and 104 billion active parameters per inference, supports a 1 million‑token context, and its Bash command execution reliability jumped from –4.5% in the previous generation to +7.9%, the best among open models. However, the 2.8‑trillion‑parameter weight file is about 5.6 TB (BF16), making local deployment on consumer‑grade GPUs practically impossible; a huge gap remains between “downloadable” and “runnable.”
On the same day, Moon’s Dark Side announced the release of Kimi Hosted Agents, offering the Harness that powers agent execution as a cloud service for enterprises and developers. Companies can obtain a 24/7 managed environment without building or maintaining their own runtime framework, supporting long‑running tasks and efficient multi‑agent collaboration.
On November 4, Hunyuan released the latest generation large language model Hy4, which has 770 billion parameters and improves instruction following, long‑text comprehension, and long‑range agent capabilities. The model weights will be open‑sourced soon, and the next‑generation Hy5 will further increase the parameter scale. The Hunyuan team says Hy4 is currently one of the strongest open‑source models, though it still lags behind top proprietary models. In several benchmarks such as ProgramBench, FrontierCode2.0, and Agents’ Last Exam v1.1, Hy4 scores surpass DeepSeek‑V4.1‑Flash‑1017, approach Kimi K3.1 Pro on some metrics, and exceed GPT6.1 Sol on others, but most benchmark scores remain below GPT6.1 Sol and Claude Opus 5.5. Hy4 is now available on Tencent Cloud TokenHub at a minimum of ¥0.3 per million input tokens. From now until December 31, pay‑as‑you‑go usage via TokenHub enjoys a 50 % discount; if you connect through Token Plan before November 18, you get a 2.5 % discount, available to all new and existing users worldwide.
In addition, on November 2, Tencent Hunyuan released the AI research system Hyra 2.0, shifting the development focus from directly using AI to tackle scientific problems to having a general-purpose agent participate in the development and upgrade of scientific models, research workflows, and software tools. This allows the results of a single R&D effort to be reused repeatedly in subsequent studies, moving the research paradigm from “solving problems” to “building tools.” Previously, Hyra 1.0 drove an open-source model to solve an open problem in combinatorial chemistry that had remained unsolved for over fifty years, demonstrating the feasibility of the approach “give AI a clear evaluation criterion and iterate.” Building on that, Hyra 2.0 further proposes that the benefits of tool upgrades can be repeatedly reused by downstream tasks, creating an intelligent compounding effect.
On November 4, according to official Netflix information, the platform began testing the AI Discover tab on TV devices. Users can start with themes such as “Something Light”, “Scare Me”, “Family Movie Night”, “Quick Watch”, etc., to view related clips and recommendations; after selecting a theme, they can further narrow the scope or choose “Try Something Else” to search again. Next to each recommended title, an “Adjust Recommendation” button appears, allowing users to keep content they’re interested in or tell AI Discover not to recommend a particular title. Netflix says this tab is a space to test new selection methods, and user feedback will determine which features are further refined or moved to the main interface. Currently, it is only available to English‑language Netflix users in the United States on supported TV devices, and will be gradually rolled out over the next few weeks.
On November 2, SenseTime launched the official version of SenseNova 6.8 Flash. Centered on complex tasks in real environments, it continuously explores the boundaries of intelligent capabilities and optimizes key challenges in task execution and result delivery. The model supports cross‑task, cross‑session collaboration, extending generative abilities to richer delivery formats.
On November 3, Baidu unveiled the world model Apollo World at the AI Day·Apollo session. It can generate a virtual city for autonomous vehicle testing, where vehicles and pedestrians are controlled by AI agents that change their behavior according to the autonomous vehicle’s steering, acceleration, or deceleration. The model also generates camera images and LiDAR point clouds in real time, allowing the autonomous driving system to complete perception, decision‑making, and feedback loops in a virtual environment. Baidu says Apollo World enables autonomous vehicles to continuously complete order taking, passenger transport, and delivery tasks in the virtual city, and can convert complex real‑world road conditions into repeatable test scenarios. Currently, the model has been used for testing and validation in the Luobo Fast Run project.
-
On November 7, Shanghai International Port Group launched an AI pilot base for the shipping logistics hub in Shanghai, integrating 54 business agents covering three scenarios: port handling, collection/distribution, and multimodal transport.
-
On November 7, Tongji University, Shanghai Jiao Tong University, and Shanghai Maritime University jointly announced “AI + Transportation” innovation results at a supporting event of the China International Import Expo, covering large transportation models, vehicle‑infrastructure collaboration, and port digital twins. The three schools agreed to establish a regular collaborative mechanism.
-
On November 7, Cursor launched intelligent agent voice calls, supporting multi‑device access with a daily usage limit.
-
On November 8, the official AWS blog outlined the primitives of Amazon Quick agents, breaking them down into triggers, agents, functions, and learned context according to the four categories of “when, what to do, who does it, how to do it.”
-
On November 4, Doubao announced integration with three official plugins: Douyin Life Service Operations, Giant Local Push, and TikTok for Business. After granting the appropriate account permissions, users can query store data, conduct operational reviews, and receive suggestions using natural language, reducing report handling and duplicate organization, and eliminating the need to switch repeatedly between multiple back‑ends.
-
On November 4, ByteDance’s AI development platform “Coze” announced that it is advancing the integration and upgrade of its enterprise‑level services, and is planning product and package options that enterprise customers can choose from in the future. In conjunction with this integration, sales of Coze Enterprise Edition will be temporarily suspended for new purchases.
-
On November 3, OpenAI notified users who are still using the old ChatGPT Classic app that the app will be discontinued soon. OpenAI is also preparing to merge Chat and Work into a single unified interface. According to the new prompt shown when users open ChatGPT Classic, the app will officially cease service on January 30, 2027.
The momentum continues: See you at AICon Beijing in December
On December 18–19, the AICon Global Artificial Intelligence Development and Application Conference will be held in Beijing. This year’s conference focuses on exploring the boundaries of AI applications, featuring 13 technical tracks that cover core areas such as post‑training for large models, inference optimization, agent engineering, knowledge engineering, intelligent agent evaluation and safety, self‑evolving systems, AI software delivery, native architecture, and organizational transformation. Leading technical experts will gather to share breakthrough AI technologies and real‑world practices for enterprise‑scale deployment.
Today’s Recommended Article
GPT-6 self‑modified Qwen weights hailed as god? Prominent Meta researcher douses the hype
Are you also “watching”? 👇
Get MirrorSpeed
Get MirrorSpeed