VicOne LAB R7 analysis shows that if inputs relied upon by robot safety responses (such as distance readings, position estimates, and stop signals) are accidentally faulty or intentionally manipulated, behaviors that should trigger protection may not be triggered. Examples include text posters misleading robot dogs' movement and inaudible audio interfering with hospital service robots' simulated behavior. The article points out that safety assessments need to re-examine assumptions based on accidental failures and proposes five key questions covering input provenance, transmission verification, manipulation impact testing, identification of common vulnerabilities across multiple inputs, and timing for evidence updates. It emphasizes the need to verify that protections remain effective even when information is tampered with, and that redundant fallbacks exist.
On October 27, 2026, Google DeepMind released Genie 4, a native multimodal world model that supports real-time interactive simulation, synchronously generates environmental sound effects and character dialogue, and integrates models such as Gemini Omni and Gemini 3.8 Flash to achieve physical simulation, causal reasoning, and task decomposition.
The official AMD ROCm blog published complete reinforcement learning training measurements for DeepSeek-V4 Flash on 24 AMD Instinct MI308X GPUs (three-node OAM configuration): utilizing an RL stack supported by AITER, separating FP8 vLLM rolling generation from BF16 Megatron Actor training, supporting online RDMA weight updates, distributed checkpoints, and long-context GRPO; within 75 steps, AIME-2024 accuracy improved from 30.0% to 49.2%, average rollout throughput was 311 tokens/s, and reward accuracy rose from 45.3% to 59.2%.
AMD announced a $1 billion subscription to newly issued Meta shares, with the first rack-scale Helios AI compute systems installed at Meta for testing. Helios integrates GPUs, server CPUs, networking equipment, and software for large model training and inference. Meta is building four AI supercomputing clusters exceeding 1 GW each, planning to nearly double its total infrastructure capacity compared to last year within the year. This round of equity financing reached $97.5 billion, the largest in Meta's history, funding data center construction, compute expansion, and businesses such as the Muse personal agent. Muse has acquired over 10 million users in the US and Canada within a month and a half of launch, but global rollout is paused due to high compute costs.
Diffusion Intelligence has released and open-sourced the diffusion language model Clover-35B-A3B-Preview. It adopts a MoE architecture with 35 billion total parameters and 3 billion active parameters, supporting thinking mode. It ranks first in 6 Agent evaluations including SWE-bench Verified, surpassing Qwen3.6-35B-A3B. On a single GB300 card, the generation speed reaches 680 tokens/s. A joint mode decision-making model was also launched simultaneously, achieving a 95.0% accuracy rate on 500 sets of related questions and defeating Jev in Chinese Chess matches.
NVIDIA's technical blog published a tutorial introducing how to adapt the Nemotron 3.5 Lightning model to public service scenarios in regions such as France, using open-source persona datasets like Nemotron-Personas-France and Nemotron-Personas-Belgium through synthetic data generation, LoRA fine-tuning, and multi-dimensional evaluation; the tutorial includes complete command-line steps, configuration file examples, behavior contract templates, evaluation metric tables, and local/cloud deployment instructions, noting that this workflow can be extended to other languages and domains.
An NVIDIA technical blog introduces three methods for optimizing Ling-3.0-flash (124B total parameters, 5.1B active parameters) MoE inference on DGX Spark: replacing the FlashInfer backend with the Humming MXFP4 MoE kernel reduces TPOT by 9.3%; online FP8 quantization of the LM head further improves throughput by 11.4%; combining NEXTN's Multi-Token Prediction (MTP) speculative decoding increases throughput by 185.1% under synthetic loads and by 42.5% in actual GSM8K tests; it also notes that using all three together requires sharing the complete LM-head module, otherwise silent degradation will occur.
Rohit Patel, former director of Meta Superintelligence Labs, left to found Ndoc, positioning it as an AI-native alternative to Microsoft Office and Google Workspace.
In an environment with unstable network connectivity at the foothills of the Himalayas, the author used the open-source coding agent OpenCode paired with a locally running Qwen3.5 9B model (via Ollama) to successfully fix an accumulation logic error in expense_total.py while completely offline: changing total = Decimal(amount) to total += Decimal(amount), and passing all 5 unittest cases; the article details practical implementation steps such as local model download, configuring cloud features to be disabled, permission review mechanisms, and aligning test commands, and provides two extended exercises: a CSV report and a log analyzer.
Apple has introduced a built-in AI assistant for Pages, Keynote, and Numbers, based on Apple's proprietary models and OpenAI models. It supports designing animations and optimizing slides in Keynote, drafting and restructuring text in Pages, and calculating data and generating charts in Numbers. This feature requires macOS 27 or higher and enables the Model Context Protocol to support third-party AI agent integration. Meanwhile, Pixelmator Pro adds a Generate Fill image editing feature, and Logic Pro and Final Cut Pro also receive synchronized updates with AI assistant and content library features.
Benjamin Skuse used free versions of GPT-5.3/5.2 Mini and Gemini 3 Flash to generate a forged paper on optical solitons via segmented prompting, successfully producing a manuscript with complete structure and professional language but containing numerous hallucinated references (12 detected by STM Integrity Hub); subsequently, he used the AI research system FARS to generate another paper based on real simulations, which was rejected due to excessive reliance on preprints and chart anomalies despite having no hallucinated citations; the experiment exposed the limitations of current automated detection and peer review, and sparked fundamental questioning of 'what constitutes legitimate scientific research'.
Opera has introduced the new WebAI web standard, enabling users to run AI models locally within the Opera One browser (such as Gemma and Apple's on-device foundation models) without relying on the cloud; this standard targets open-weight models and supports text generation, content classification, and understanding of information relationships; users must actively enable it, with functionality planned to expand to Opera GX and Opera Neon in the future; Opera emphasizes the privacy benefits offered by this approach alongside hardware performance trade-offs.
RoboParty has released the full-stack open-source humanoid robot RP1, opening up mechanical drawings, electrical systems, proprietary ROMOMO joint modules, URDF/MJCF models, the hhtools human motion retargeting tool, and some components of PartyOS; it offers body configurations with 24–30 degrees of freedom, supporting continuous walking on real hardware, disturbance recovery, and getting up after falling; resources are hosted on GitHub, and internal testing applications have opened.
David Konitzny (Peec AI) found that content from OpenAI's licensed partners (such as Condé Nast, Reuters, Allrecipes, and other domains totaling 43) is integrated into ChatGPT via dedicated pipelines, presented as pre-generated summaries, and placed before all web results in prompts; whereas general web pages enter only as raw text snippets.
Qianli Technology has released an AI 4D millimeter-wave radar equipped with a built-in 10 TOPS NPU, achieving a fully model-driven pipeline from antenna design and waveform generation to point cloud detection, angle measurement, enhancement, target tracking, and classification. It supports three working modes: edge-side mode, satellite architecture mode, and compatibility mode. Compared to the previous DSP solution, point cloud density has increased by 10 times, detection range by 40%, ghost elimination rate reaches 95%, and spatial angular resolution has improved by 12 times. This radar was developed by Qianli Technology, led by Yin Qi, with collaboration from StepStar's foundation models and Geely's vehicle scenarios.
The U.S. Office of the Comptroller of the Currency (OCC) has granted Ripple final approval to establish Ripple National Trust Bank, which will manage segregated reserves for the RLUSD stablecoin, serve as trustee for holders' collateral, and provide fiduciary digital asset custody services to enterprises; RLUSD was launched in December 2024 with a current market capitalization of approximately $2.6 billion, and is still issued by Standard Custody & Trust Company, regulated by New York State, thus subject to both state and federal regulation.
A report by The Verge indicates that community opposition to data center construction is deeply intervening in the 2026 U.S. midterm elections, becoming a core issue for gubernatorial and senatorial campaigns in multiple states; candidates from both the Democratic and Republican parties are generally adjusting their positions, proposing policies such as pausing construction, eliminating tax incentives, and mandating the bearing of energy and water resource costs, with some supporting state-level or federal moratoriums while others emphasize local autonomy; the report covers specific campaign dynamics and policy proposals in 10 key states: Texas, Florida, Pennsylvania, Ohio, Michigan, Wisconsin, Nevada, Georgia, Arizona, and Iowa.
Zhiyuan has released the embodied agent framework Genie Master, which enables the Genie G2 robot to autonomously diagnose failure causes, adjust strategies (such as using wrist cameras to supplement viewing angles for stocking shelves), accumulate experience, and update action policies in a real convenience store environment, achieving recursive self-improvement (Physical RSI) without relying on manual debugging. The framework includes dual pathways: Harness RSI (experience accumulation) and Policy RSI (policy updating), along with the Embodied Execution Protocol (EEP). A live test involving 6 robots operating in a convenience store for 24 hours will be conducted on November 2.
A seller in the Electronics category conducted three consecutive rounds of Woot BD activities during the new product period when ACOS approached 70%. Daily orders rose from 400 to over 900, and ACOS eventually dropped to 28.19%. The core mechanism involves multiple activity rounds progressively raising the daily sales baseline and pushing the organic ranking of core keywords from Top 9–11 to Top 1–2, driving CR up from 4.25% to 7.63% and CPC down from $2.56 to $1.82. The ad structure used SP Exact (49.2%) to secure keyword positions, while Auto (24.2%) and PT (26.6%) expanded entry points; the number of indexed traffic keywords increased from 1554 to 3444.
A Zhejiang seller had $45,000 in funds frozen after a US company initiated a TRO against them for using "Organic Honey" in an Amazon title, claiming it was a registered trademark. The author cites Chinese judicial precedents involving Sichuan peppercorns and honeysuckle, as well as the US descriptive fair use principle, to point out that the term is a generic description in the food industry. Proper usage by sellers (making their own brand more prominent and using "Organic" only as an adjective) constitutes fair use. The article proposes a three-step self-check method before listing products: check the trademark database, review how terms are used, and distinguish between description and free-riding.
This article details the legal and operational pitfalls that cross-border e-commerce brands are prone to encounter in licensing, agency operations, and co-branding collaborations. Citing real cases such as Laoshan Cola, Wanglaoji, and JDB, it highlights key risk points including quality control clauses, filing obligations, ownership agreements, store attribution, post-termination inventory clearance in co-branding, and ownership of deliverables. It also provides templates for maintaining authorization lists and references to court judgments.