Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
The Berkeley Artificial Intelligence Research (BAIR) lab released the ABBEL framework, which compresses interaction history into supervised natural language belief states, improving summary quality through belief grading.
SGLang v0.5.16 has been officially released, with core updates including: the DSpark confidence-driven speculative decoding algorithm (achieving 383.7 tok/s on DeepSeek-V4-Pro).
Playwright v1.62.0 has been released, introducing a new component testing model based on stories and galleries, supporting operation cancellation via AbortSignal, and adding WebP format screenshots (with lossless/lossy options).
NVIDIA has released ModelExpress (MX), a system designed to accelerate large model weight distribution and inference startup. It supports multiple paths including GPU-to-GPU direct transfer (P2P RDMA), streaming loading from object storage (ModelStreamer), and GPUDirect Storage (GDS). Measured results show that it reduced the cold start time of the DeepSeek-V4 Pro model from 8 minutes to 1 minute and 44 seconds, and supports JIT kernel cache reuse and dynamic weight updates after RL training.
In its blog, ElevenLabs elaborates on how voice AI can be used to strengthen democratic participation in the 2026 global elections, including collaborating with Brazil's Superior Electoral Court (TSE) to add a voice interface to the 'Election Assistant' and supporting the Comprova alliance for fact-checking.
An analysis by Sequoia Capital indicates that between January 2024 and February 2026, Qwen's share in new open-source model fine-tuning and adaptation rose from 1% to 69%. A large number of U.S. AI startups and labs are using Chinese open-source models such as Kimi, Qwen, GLM, and DeepSeek as teacher models and synthetic data sources to circumvent legal restrictions on using outputs from GPT or Claude. While this dependency pathway is legal and efficient, it carries fundamental risks, including unauditable model backdoors and constrained access to frontier capabilities dependent on China's release policies. The report proposes establishing a domestically controlled 'teacher access' mechanism in the United States as an alternative pathway.
Ruff 0.16.0 has been officially released. Key changes include: the number of rules enabled by default increased from 59 to 413 (while removing 18 opinionated pycodestyle and pyflakes rules); added default formatting support for Python code blocks in Markdown files; support for `ruff: ignore` trailing or preceding comments; the `check` and `format --check` commands now display specific fixes and multi-format output (including GitHub/GitLab CI annotations); some fields in JSON output are allowed to be `null`; multiple rules and behaviors have been stabilized; new preview features, performance optimizations, documentation corrections, and bug fixes were added.
The NVIDIA OptiX Toolkit (OTK) is a set of open-source debugging tools offering two core capabilities: consistent OptiX/CUDA API error-checking macros (OTK_ERROR_CHECK) and a targeted device-side debug printing mechanism based on the DebugLocation structure. It supports one-click enablement of validation, visual localization of debug pixels, single-trigger output, and includes a DemandPbrtScene example demonstrating an interactive debugging workflow. The code is hosted on GitHub under the BSD 3-clause license.
NVIDIA 技术博客发布教程，演示如何通过 Prime Intellect Lab 平台对 NVIDIA Nemotron 3 Nano 进行强化学习定制：第一步获取基线性能（准确率21.9%），第二步配置并运行100步RLVR训练生成LoRA适配器，第三步部署后重测，准确率提升至90.6%，总花费低于5美元；教程同样适用于 Nemotron 3 Super 和 Ultra，并提供全部CLI命令、环境配置和验证方法。
The official Hugging Face blog announced that Diffusers now natively supports Nunchaku Lite, a 4-bit weight and activation (W4A4) quantization scheme based on SVDQuant.
ChatGPT has introduced a Health feature that allows eligible US users to securely connect their medical records and Apple Health to gain more personalized health insights and better understand their own health conditions.
Google Research released a research paper titled "SymptomAI: Towards a conversational AI agent for everyday symptom assessment".
Google Research published the paper "Reinforcement learning control of quantum error correction" in Nature.
An official NVIDIA technical blog post introduces a tutorial on how to make the TensorRT engine build process observable (nested progress bars) and cancellable (responding to Ctrl-C or programmatic signals) via the IProgressMonitor interface.
At the DOE Genesis Mission Summit 2026, Google DeepMind announced an investment of $40 million in AI tokens and cloud resources to provide AlphaEvolve to U.S. Department of Energy National Laboratories.
In a technical blog post, NVIDIA announced that its GB300 NVL72 system achieved a world record of 1,648 TFLOPs per GPU during the pre-training of the DeepSeek-V3 671B Mixture-of-Experts (MoE) model. This performance improvement stems from the NVLink 5.0 all-to-all interconnect architecture, infrastructure offloading via BlueField DPU, and continuous software optimizations in Megatron Core/TorchTitan/JAX, with software optimizations alone yielding performance gains ranging from 1.5x to 10x within six months.
NVIDIA has released the Rubin GPU architecture, designed specifically for Agentic AI workloads, claiming up to 10x higher agentic inference throughput per watt compared to Blackwell.
NVIDIA has released the Vera CPU, featuring the Olympus core designed specifically for agentic AI. It emphasizes high single-thread performance, low latency, strong branch prediction, deep out-of-order execution, graph-aware prefetching, and Spatial Multithreading technology. The architecture employs a monolithic compute die, Scalable Coherency Fabric, SOCAMM2 LPDDR5X memory, and NVLink-C2C dual-socket scaling. SPEC CPU® 2026 tests show up to a 1.8x performance improvement on agentic workloads.
Google DeepMind has released three new models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The 3.6 Flash model reduces output token consumption by 17% compared to the 3.5 Flash.
The Tongyi Qianwen team released Qwen-Image-3.0, the third-generation image generation foundation model, emphasizing "substance"—rich content (supports up to 4.5k token input, capable of generating a nine-grid containing 9 complex infographics in one go), realistic details (precise rendering of 10px small text, complete LaTeX formula typesetting, simulation of paper texture and handwritten annotations), and substantial knowledge (native support for 12 languages, simulation of various UI interfaces, and retrieval of real-time weather and IP characters via internet connectivity).
Hugging Face has released Grabette, an open-source, low-cost handheld system for collecting robot operation data. It consists of a gripper equipped with dual cameras (fisheye + RGBD) and an IMU, capable of synchronously recording the hand's 6-DoF trajectory and grip state. The data is processed in the browser and automatically converted into the standard LeRobot format before being uploaded to the Hugging Face Hub. The accompanying hardware, Gripette, is a motorized gripper deployable on real or simulated robotic arms. The entire system is fully open-source and supports any robot platform.
Meta AI announced that its open-source vision foundation models, SAM 3 and DINOv3, are being used in SYNAPS-I, a flagship project under the U.S. Department of Energy's Genesis Mission. The team fine-tuned both models on scientific images collected at DOE light source facilities (such as ALS).
Analysis published by a16z crypto indicates that as of the end of June, the market capitalization of tokenized stocks reached $1.7 billion, representing more than a 5-fold increase compared to one year ago. Among these, the share of AI and chip-related tokenized assets surged from 0.3% to 15.5%, while the market share of crypto-linked products dropped from 79% to 21%. Institutions such as DTCC, NYSE, Robinhood, and Coinbase have recently intensified efforts to advance on-chain stock trading, launch proprietary chains, and facilitate cross-border issuance.
NVIDIA has released the sixth-generation NVLink interconnect technology, including the NVLink 6 Switch, which provides 3.6 TB/s bidirectional bandwidth per GPU, 260 TB/s rack-level bandwidth, 3x lower GPU-to-GPU latency, 10x higher packet rates, and 130 TFLOPS FP8 in-network compute; it supports a 72-GPU full-mesh topology for the Vera Rubin NVL72 platform and integrates intelligent resilience features such as hot-swappable switch modules, dynamic traffic rerouting, and fine-grained link telemetry; simultaneously, NVLink-C2C was launched to achieve 1.8 TB/s coherent bandwidth between CPU and GPU, while NVLink Fusion enables semi-custom XPU access to unified AI infrastructure.