Search “Reasoning”

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.

SenseTime Open-Sources SenseNova-RoboRSI Framework and 27B Embodied Model

SenseTime has open-sourced the SenseNova-RoboRSI framework and a 27B embodied model, while also releasing an empirical study on embodied intelligence for GPT-6-Astra. The study covers 12 simulation benchmarks and 8 real robot platforms. Under a unified protocol (ICL plus up to 10x interaction budget), GPT-6-Astra achieved SOTA on all 12 benchmarks, with RoboDojo's success rate rising from 27.6% at 1x budget to 81.9% at 10x, and DexJoCo reaching 87.3%, compared with the previous SOTA of 40.3%. The study also notes that simple tasks such as putting a pen into a pen holder cost about $55 in inference, and that real robots overheat and wait too long; the conclusion is that GPT-6-Astra has essentially crossed the threshold of embodied capability, but practical-grade embodied intelligence has not yet arrived. The open-source framework packages each benchmark as an independent Docker image and provides environment interaction and scoring through a unified HTTP API; the 27B model raises the success rate on RoboCasa365 from 26.0% to 46.5%, reduces task token overhead by 73%, and costs about 1/50 of GPT-6-Astra per completed task.

i14 migrates its custom architecture to AWS Trainium2

i14 migrated its custom architecture to AWS Trainium2. According to information released by AWS Machine Learning on October 21, 2026, i14 collaborated with AWS to train and deploy the Gated Linear Attention-based Laplace 3B model on AWS Trainium2 using TorchNeuron Native. Training throughput on a single trn2.48xlarge is approximately 120,000 tokens/s; the model's recurrent state is fixed at 29.5 MB per sequence and does not grow with context length; inference reaches 11,544 output tokens/s at 1,280 concurrent sequences. The team also reworked Muon's all_to_all communication, reducing per-card communication volume by 24x and increasing training throughput by 1.63x. There are currently no subsequent developments.

NVIDIA releases surgical generative simulation model Cosmos3-H-Dreams

NVIDIA has released the surgical generative simulation model Cosmos3-H-Dreams. The model is specialized from the 16-billion-parameter Cosmos3-H-Embodied-Nano for action-conditioned video generation of cholecystectomy. Official information shows that the model was trained with 28 action dimensions, 832 x 480 video, and 10 Hz action alignment, and was distilled into two student checkpoints, four-step and two-step; the two-step version was trained separately and offers faster inference but lower image quality. NVIDIA also provides the FlashDreams streaming runtime, as well as methods for adapting the model to proprietary data and controllers, and states that the model is a research simulator, not a validated clinical system. There are currently no subsequent development facts.

Google launches Gemini personal intelligence feature on Android Auto

Google has launched the Gemini personal intelligence feature on Android Auto. According to a Thurrott report on October 21, 2026, the feature has begun rolling out to Android Auto; once users enable it, it can connect to personal data in Gmail, Google Calendar, Google Photos, and Google Search to provide more personalized answers. The feature is entirely opt-in in Gemini settings. Karan Daryanani, a senior product manager at Google, said its two core capabilities are extracting specific details from emails, photos, or saved places in Google Maps, and reasoning across complex sources. The report also said that in the coming months the feature will arrive on built-in Google cars that run Google's automotive system directly. There are currently no subsequent developments.

Nebius executive proposes that open models should be evaluated using internal benchmarks

Ryan Hanrui Wang, Senior Vice President of Applied AI at Nebius, proposed that evaluating open models should combine industry benchmarks, internal benchmarks, and testing frameworks. He noted that public benchmarks only reflect a model's best-case performance and cannot show whether agents can accurately complete tasks in production environments, so the entire inference stack should be evaluated. He also suggested using large models as references to scan multiple models and find the scale sweet spot; for example, 10B parameters may already be sufficient for customer service tasks, while 100B or even 1T parameters bring only marginal accuracy gains. The related report also mentioned that Nebius acquired inference optimization startup Eigen AI for approximately $643 million and provides inference services through the Token Factory platform.

Evan Conrad says NVIDIA cannot build its own compute trading market

On October 21, 2026, SF Compute CEO Evan Conrad argued on the Latent Space podcast that NVIDIA cannot build its own compute trading market. His reasoning is that companies with large hardware or inference businesses find it difficult to operate a neutral exchange. In the same appearance, Conrad also said that the compute market must first achieve physical delivery—that is, actually delivering usable GPU compute—before discussing financialization; otherwise, market manipulation and basis risk are likely to arise. He explained that SF Compute is expanding from subleasing GPU clusters to building its own cloud operations capability in order to achieve physical delivery of GPU contracts, so that buyers receive usable compute rather than cash settled against an index; the company does not hold hardware in order to remain neutral, and it already supports customers in partitioning InfiniBand fabric online themselves. He also said that SF Compute has placed 5% of its capacity on the order book, that customers who spend $100 to buy capacity can on average recover $25 by reselling it, and that hundreds of millions of dollars in volume has already been traded. There are currently no subsequent developments.

Stanford releases public video and lecture notes for CME 295 Lecture 4

Stanford course CME 295 Transformers & Large Language Models has been updated to Lecture 4, with both the video and lecture notes now publicly available. The topic of this lecture is Reinforcement learning with LLMs, covering everything from reinforcement learning and reward design to PPO, GRPO, RLVR, and DAPO, explaining each of the notations one by one. The course is taught by Afshine and Shervine Amidi, with 9 lectures in total, covering the Transformer architecture, attention variants, RoPE, LoRA fine-tuning, RLHF/GRPO training, Flash Attention inference optimization, MCP and AI Agents, and diffusion language models. Currently, Lecture 4 is public, and no evidence of updates for subsequent lectures has been seen yet.

Rohan Paul relays that Anthropic has an unreleased model

On October 20, 2026, Rohan Paul relayed a report from The Information stating that Anthropic has an unreleased model capable of inferring user goals from sparse instructions. The relay also stated that this model outperforms Anthropic's current models in reasoning and intuitive judgment, including in cases where users have never explicitly stated their needs. There are currently no follow-up developments, no official evidence, and no additional cross-sources; the status of the event remains at the level of this relay.

AMD executive argues physical AI evaluation should prioritize real-time control and reliability

AMD executive Salil Raje proposed that the performance evaluation of physical AI should shift from raw compute power to real-time control and system-level reliability. The claim was made around October 19, 2026, and related reports were published on October 20. According to the reports, AMD builds heterogeneous architectures with Kria AI SOM, Versal, Zynq, and Spartan FPGAs; in third-party tests, the Kria AI SOM achieved inference times under 0.1 seconds and supported 8,000 control decisions per second; in open navigation benchmarks, the Ryzen AI Embedded configuration achieved control timing reliability up to 3.4 times that of Nvidia Jetson T5000, with 1.6 times more available compute headroom. There have been no subsequent developments so far.

dreamlabs_ai releases robot foundation model Dream-1

On October 20, 2026, dreamlabs_ai emerged from stealth and released its robot foundation model Dream-1. According to the same report, the model can control previously unseen dual-arm robots, perform untrained tasks, and enter locations it has never been to before, starting at $3 per hour. On the same day, the Dream API opened to everyone, with its inference powered by fal Serverless. There are currently no follow-up developments.

Ant Ling thanks SGLang-Omni for providing day-one support for Ming-3.0

On October 20, 2026, Ant Ling publicly thanked the SGLang project, expressing gratitude for providing day-one SGLang-Omni support for Ming-3.0 and releasing a practical cookbook to help developers get started. According to the original SGLang post, SGLang-Omni provides day-one support for three versions of Ming-3.0: Ming-3.0-asr-tiny supports Chinese, English, and dialect ASR; Ming-3.0-tts-tiny supports voice cloning in 7 languages; and Ming-3.0-audio-flash is a 126B-A6.5B MoE model for audio understanding and reasoning. Currently, this acknowledgment was published by Ant Ling, and there are no subsequent developments.

The Ministry of Commerce and the Ministry of Science and Technology solicit public comments on the revision of the technology export catalog

On October 16, 2026, the Ministry of Commerce and the Ministry of Science and Technology issued an announcement soliciting public comments on the revision of the Catalog of Technologies Prohibited or Restricted from Export in China. According to The Decoder, the draft revision covers 23 key AI technologies including large models, agents, computing clusters and data processing; AI tools capable of autonomously writing or rewriting malware to attack critical infrastructure, military or government agencies are proposed to be fully prohibited from export, and some vulnerability-discovery AI tools face new export restrictions. The report states that large models are regulated when trained on and providing services based on China's important data, or when involving core technologies such as weights, checkpoints, source code, architecture and training and inference methods, with trigger thresholds of training compute reaching 10^26 FLOPs, at least 2 trillion parameters, or entering the global top ten; releasing restricted AI technologies and weights as open source can obtain automatic licensing, use through APIs, applications, websites, SaaS or lightweight on-device versions is not regulated, while deployment on local servers outside China requires applying for a standard export license. Other reports state that Joint Announcement No. 48 of the two departments proposes adding a total of 40 technologies in the three major fields of artificial intelligence, biomedicine and photovoltaics to the catalog, including 23 in artificial intelligence, and immediately implementing interim registration management before the catalog takes effect. The catalog is currently still in the public comment stage.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.