English text is being prepared.

vLLM Blog Publishes Technical Retrospective on Kimi K3 Inference Optimization

On October 16, 2026, the vLLM blog published a technical retrospective titled "Hyper Optimizing Kimi K3 in vLLM: Moving Fast and Responsibly." According to a TMTPost·AGI report, NVIDIA CentML engineer Benjamin Chislett used a Megakernel written in 12,000 lines of CuTe DSL to optimize inference for Moonshot AI's Kimi K3; the vLLM blog reports over 840 tokens/s at batch 1 and a 1.5x to 2x improvement in decoding throughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog states that agent-optimized inference code is now outpacing human review and maintenance capacity, and the team therefore proposed a fast-track model optimization RFC so that reviewers judge designs from high-level abstractions rather than reviewing code line by line. The report also mentions that NVIDIA is developing optimizations for other models in vLLM, such as DeepSeek V4.1.

Rohan Paul says Anthropic's revenue curve will become legendary

On October 28, 2026, Rohan Paul offered the judgment that Anthropic's revenue curve will become legendary. Citing a Bloomberg chart, he said that Anthropic's revenue in the first half of 2026 was nearly 400 times its full-year 2023 revenue, and that no other company in the chart comes close to that multiple; Nvidia, the second-fastest grower, remained below 50 times throughout its first three years after going public, while Alphabet, Meta, Microsoft and SpaceX essentially stayed near the 1x baseline. The same material also states that Anthropic's full-year 2025 revenue was about 100 times that of 2023, so its revenue in just six months of 2026 was about 4 times its full-year 2025 revenue. At present, this judgment appears only in this single source, Rohan Paul, and there have been no subsequent developments.

Jensen Huang proposes that doubling model speed could double AI lab revenue

On October 29, 2026, Jensen Huang put forward a judgment: under the premise that demand exceeds capacity, doubling model speed (for example, through software or kernel optimization) could double AI lab revenue. The logic is that inference revenue equals the number of tokens served multiplied by the price per token; a faster kernel can increase tokens per second under the same GPUs, power, and data center conditions, and if all the additional tokens can be sold at the current price, a 2x speedup could double revenue with zero additional capital expenditure. Currently, this claim is presented in a report by Rohan Paul, with no other developments or official response yet.

Google releases world model Genie 4

On October 27, 2026, Google DeepMind released the world model Genie 4, focused on real-time interactive simulation and introducing native audio to the series for the first time. According to the official introduction, the model is built on capabilities from the Gemini family: Gemini Omni is responsible for making the simulation match the appearance and behavior of the real world, while other Gemini models provide the complex reasoning needed to maintain the simulation's coherence; users can control scenes in real time using natural language, and each world also instantly generates ambient sound, music, and interactive dialogue. The next day, Google DeepMind further explained that Gemini 3.8 Flash breaks down goals in a loop and checks frames to maintain causal coherence across multi-step events, and that Gemini 4 Argon has also been used to jointly simulate tasks with Gemini Robotics; the company said it remains an early experimental model, that object permanence and complex physical interactions still drift during long runs, that all generated content carries a SynthID watermark, and that it is currently being previewed to trusted testers first. On October 28, Ethan Mollick posted that Genie 4 can run DOOM, and tested it generating a game-like environment as it goes, saying the experience is not perfect. On October 29, Peter Yang used Genie 4 to build 5 worlds in real time, arguing that world models are the next big direction for AI after LLMs.

English translation is being prepared.

Microsoft says China-linked hackers breached about 2,250 organizations using AI agents

Microsoft security researchers say that hackers linked to the Chinese government breached ConnectWise's cloud backup service and stole application credentials used to establish data links, allowing them to log in as authorized users to customers' Microsoft-hosted environments and access email and other files. The hackers used AI agents to break into about 2,250 organizations at a rate of nearly 3,000 login attempts per minute, a pace Microsoft says left defenders with almost no time to detect and respond. The AI framework used was the open-source stress-testing framework ARTEX. ConnectWise said it has suspended access to the cloud backup system and launched an investigation, and as of Thursday it was restoring service.

Andrew Ng proposes a four-element framework for AI-native engineering teams

On October 23, 2026, in the DeepLearning.AI letter, Andrew Ng proposed that engineering teams transforming into AI-native teams should start from four aspects: people, projects, platforms, and processes. Specifically, he suggested encouraging members to improve their AI engineering capabilities and build hands-on themselves, supporting a small number of promising projects from the top down, organizing data, repositories, and internal tools into infrastructure usable by agents, and redesigning processes around the new bottlenecks brought by AI. He believes that the cost of building software with clear specifications has dropped significantly, while other links have instead become bottlenecks, so teams with sufficient context and full authorization should be formed. Currently, this framework was proposed by Andrew Ng himself, and there are no subsequent development facts.

Pichai explains why Gemini 3.5 Pro has not been released

On October 23, 2026, Google CEO Sundar Pichai responded to the question of why Gemini 3.5 Pro was ultimately not released. According to a report by Alex Heath, Pichai said that while training the 3.5 series, the team found gaps in agentic capabilities, especially in coding, and believed that iterating with the smaller and more efficient Flash would be faster, so they released 3.5, 3.6, 3.7, and 3.8 Flash in succession; by the time this set of capabilities could be applied to Pro, Argon pretraining had already been completed. This explanation currently comes from that report, and there are no other developments at this time.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Meta is going all in on Muse, and this AI app is quickly gaining popularity

On September 25, 2026, reports said Meta heavily promoted its AI app Muse at the annual Meta Connect developer conference. According to Sensor Tower estimates, Muse has surpassed 3.4 million downloads since launching on September 8, up from 2.5 million earlier this week; estimates from other firms vary widely, with Apptopia counting 4.3 million and Appfigures about 2.3 million. The report also said Muse topped the U.S. App Store on September 18 and has held the top spot since, and reached the top of Google Play on September 19. On the same day, another report said Meta put the Muse agent at the center of the conference and announced features including Computer Use, a dedicated email address, a digital avatar, and the pocket device Muse Charm. Subsequently, Meta opened early access applications for new Muse features on September 25; launched the Meta Enterprise Platform on September 28, incorporating Muse and others into its enterprise technology stack; announced on September 29 that it would expand Muse to small businesses and add multiple integrations; and on October 7, the Muse iOS app added native iPad support and integrated more third-party tools.

English translation is being prepared.

Li Xinyi criticizes OpenAI and Anthropic for harming the mathematics community by using AI to solve problems

On October 3, 2026, Li Xinyi, associate professor at the Beijing International Center for Mathematical Research at Peking University and an alumnus of Peking University's Mathematics Department (Class of 2007), publicly criticized OpenAI and Anthropic for using AI to "solve" difficult mathematical problems. He believes that such results are "solved but not understood at all," and that the accompanying publicity will scare off young people who want to study mathematics and mislead funding decisions. Li Xinyi also revealed that a postdoc on his team's draft on an obscure conjecture highly overlapped with a cross-disciplinary paper generated by AI on arXiv, ultimately ending in joint authorship; he maintains that academic credit should only go to those who contribute to understanding. In addition, he pointed out that work in applied mathematics of the "combining A with B" type is more easily replaced by AI, whereas in fields such as geometry and topology that rely on spatial intuition, AI currently does not learn well. At present, this criticism was reported by Huxiu, and no other sources have cross-verified it.