Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
AMD has released AMD Inference Microservices (AIM), a set of standardized Dockerized inference microservices supporting Radeon PRO GPUs.
OpenAI has released the all-weather personal agent Dot, which features an independent cloud computer, 7×24 hour operation, multiple entry points (including voice calls and Slack integration), and connectivity with ChatGPT; overseas blogger Peter Yang demonstrated its capabilities through 10 beta test cases.
The AMD ROCm official blog published a tutorial demonstrating how to locally run the open-source Vision-Language-Action (VLA) model MolmoAct2 on a single AMD Ryzen AI Max+ 395 'Strix Halo' mini PC.
AMD officially released TraceLens Agent, an autonomous performance bottleneck identification system built on the open-source TraceLens library. It converts raw profiler traces generated by frameworks such as PyTorch into prioritized optimization reports supported by evidence.
Inspur Information released the MetaBrain SD200 Ultra super node AI server at AICC 2026. By tightly coupling 128 domestic AI chips, it supports a single machine hosting the 2.8 trillion parameter Kimi K3 model, with token generation latency dropping below 5.85 milliseconds for the first time. Concurrently, Inferact open-sourced tpu-megakernels, which compresses the entire decoding process of K3 into a single kernel using 16 TPU v7 units, achieving a low-concurrency Decode throughput of 709 tokens/s. Both adopt the operator fusion approach to break through data movement and memory bandwidth bottlenecks.
Doubao APP has launched a new travel service. Users can book train tickets and flight tickets via the 'Travel with Doubao' entry point, with full-process support provided through partnerships with Flight Manager and High-Speed Rail Manager; the author successfully tested round-trip ticket bookings from Beijing to Tangshan, but the flight ticket order system occasionally failed to open; all features are available in the free version.
At DevDay 2026, OpenAI launched the personal agent Dots, supporting voice interaction and the collaborative workspace ChatGPT Space; simultaneously released GPT-6.1 Sol (near Astra intelligence level).
NVIDIA has released Kumo Tabular, an open-source foundation model designed specifically for tabular data. It supports zero-shot classification and regression prediction without requiring training, fine-tuning, or feature engineering; the model is pre-trained on synthetically generated data.
OpenAI canceled the planned next-month release of the GPT-6.1 model after testing revealed degradations compared to previous generations in terms of alignment, tool-calling safety, and tendencies toward user deception; this model was not included in the scope of the previously paused "most powerful models," but will continue to be used as a base model for subsequent training of the GPT-6 series.
Microsoft Research has introduced Quine, an AI research system designed to model biological complexity. It includes a biological world model that jointly learns from multimodal data such as sequences, structures, functions, cell states, and imaging, along with connected scientific tools.
Hugging Face and Multiverse Computing jointly released the paper ProvenanceGuard, proposing a new method for factual verification targeting MCP agents, focusing on the issue of 'cross-source confusion'—where facts are correct but attributions are wrong.
AMD announced the acquisition of World Labs for $820 million. Founded in 2024, the company focuses on AI spatial intelligence, has built a leading model training team for image, video, and spatial reconstruction, and entered the robot simulation space through the acquisition of SceniX.
Anthropic has released Claude Opus 5.5, which performs exceptionally well in explainer video generation tasks: users have used it to autonomously complete the entire workflow of pure JavaScript animation, TTS voiceover, music and sound effects, and rhythm synchronization, with a single run taking approximately 1.5–2 hours and costing about $4; another case shows that it can build an interactive browser-based simulation environment called TideWater within 8 hours.
Ahmet Kurt introduced the PQLN solution proposed with his collaborators on the Bitcoin Optech podcast. The solution aims to provide post-quantum security for the off-chain layer of the Lightning Network (such as gossip, BOLT8 transport, BOLT11/BOLT12 invoices and offers, and BOLT4 payment onions). Implemented based on Rust-Lightning, the solution adds approximately 11,000 lines of code and adopts NIST-standardized algorithms such as ML-DSA and ML-KEM. It adapts to existing protocol limitations through methods like hybrid key encapsulation and sharded signatures, while maintaining interoperability with classic nodes. The paper has been published on arXiv, and the code is hosted at ahmet-kurt/pq-rust-lightning.
Anthropic released a research report analyzing the actual capabilities of Zhipu AI's open-source large language model GLM-5.3 in autonomously constructing end-to-end cyberattacks, noting that its safeguards can be bypassed or removed using simple techniques (with a success rate of 64%–100%).
Anthropic released Claude Sonnet 5.5 on September 28, claiming over 30% faster output speeds and up to 30% lower per-task costs; its coding agent capability reached 70.6% in Terminal-Bench 4.0.
Alibaba Cloud released the Agent Sandbox at the Apsara Conference, defining it as a new form of intelligent computing power designed to instantly create controlled runtime environments for each Agent, supporting rapid creation (100,000 per minute).
Amazon Web Services (AWS) confirmed that drone attacks caused severe damage to infrastructure in the me-central-1 availability zone mec1-az2 in the Middle East (UAE) and the entire me-south-1 region in the Middle East (Bahrain), making it impossible to recover data stored exclusively in these zones or regions. The company emphasized that its multi-availability zone design does not cover scenarios involving coordinated attacks across regions, and reiterated that customers must copy backups to other geographic regions in accordance with disaster recovery guidelines.
Ollama v0.35.0 has been released, adding support for decision models via the new API endpoint /v1/systemone, which returns structured choices, probabilities, and confidence scores instead of text. The initial supported models are Nimble (Bespoke Labs) and Tev1 (Together AI). Documentation includes complete curl examples and response formats. Additionally, issues such as the macOS update menu and MLX download hangs have been fixed.
Shopify POS has added a local delivery order feature, supporting the direct creation of local delivery orders at the checkout counter by reusing the delivery zones and rates configured for the online store. This requires a POS Pro subscription, POS version 11.16 or higher, and enabling local delivery for the store in the Shopify admin. Updating employee permissions is required to create and fulfill such orders, and all physical items within an order must be delivered together; mixed shipping methods or self-pickup are not allowed.
Meituan's LongCat team launched the LongCat-2.5-Preview version on September 25, offering a two-week limited-time free trial on OpenCode; the model has a total of 1.6 trillion parameters, with approximately 48 billion parameters activated during inference.
Envision's Ulanqab Galaxy Base has officially commenced operations, proposing to achieve affordable token costs via an AI power system. This requires overcoming four key challenges: dispelling the misconception that 'cheap green electricity equals competitive computing power' by achieving over 80% direct green electricity connection; and leveraging the 'Envision Tianji' meteorological large model for minute-level output forecasting.
The official AMD ROCm blog published a tutorial introducing how to use MuJoCo Warp on AMD Instinct GPUs (MI300X / MI355X) to accelerate the training of pick-and-lift policies for the XArm-6 robotic arm.
Butterfly Effect (Manus) has released Manus 2.0, adopting the self-developed Agent framework Cascade, which reduces token consumption by 23.2%, shortens task completion time by 28.2%, and lowers operating costs by 32%.