Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
Google Research published the paper "Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion" at ICML 2026.
Google DeepMind has released two new voice dialogue models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
An NVIDIA technical blog details the architectural differences between dense and MoE models, noting that MoE models (such as Nemotron 3.5 Lightning) can activate only 3B parameters per token (out of 30B total parameters), improving throughput via routing mechanisms; whereas dense models involve all parameters in inference for each token.
NVIDIA has released NVLink 6 interconnect technology as a core component of the Vera Rubin AI Factory platform. It builds a full-stack lossless network through lightweight FEC + PLR + UPHY recovery at the physical layer, credit-based flow control (CBFC) at the link layer, Shadow Engine Recovery at the application layer (reducing inference interruption recovery time from 283 seconds to 7.3 seconds), and rack-level maintainability at the system layer. This architecture enables GPU clusters and third-party XPUs (via NVLink Fusion) to uniformly inherit the same fault-tolerance capabilities.
NVIDIA integrates the Groq 3 LPX low-latency accelerator into the Vera Rubin platform, achieving clock-cycle-level scheduling of data movement and computation via a cycle-exact deterministic execution model.
IBM Research and Hugging Face introduced the Consistency Analyzer and consistency guidelines in a blog post to diagnose and improve the result consistency of AI agents in repetitive tasks.
Agility Robotics has unveiled the Digit 5 humanoid robot, which stands 1.8 meters tall and weighs 129 kilograms. It can lift 23 kilograms to a height of 2.1 meters per cycle, supports close-proximity collaboration with humans without physical barriers, and offers over 20 hours of daily operational endurance. Its safety mechanisms include autonomous avoidance, stopping, or actively assuming a stable seated position to prevent tipping and injury. According to SEC filings, the target mass-production cost is under $50,000 per unit at an annual volume of 10,000 units, with service pricing quoted at approximately $8,500 per month.
NVIDIA FLARE 2.9 has been officially released, adding native support for the Slurm batch scheduling system; version 2.8 previously supported Docker and Kubernetes.
NVIDIA's official technical blog announced that Transformer Engine has added deep optimizations for Dropless MoE training within the JAX framework, including grouped GEMM.
On August 25, 2026, OpenAI released its first self-developed AI acceleration chip, Jalapeño. It provides up to 13.4 petaflops of 4-bit computing power, features 232 GB of advanced memory with a bandwidth of 15.4 TB/s, and offers up to 3.6 times lower end-to-end latency and lower power consumption compared to Nvidia's GB300.
Miles Jennings, a researcher at a16z crypto, noted that the US Senate will begin debate on the Digital Asset Market Clarity Act (CLARITY Act) on September 15, 2026. The bill aims to bring digital commodity brokers, dealers, and exchanges under a federal regulatory framework, requiring client asset segregation, qualified custody, restrictions on conflicts of interest involving related parties, mandatory disclosures, listing standards, insider trading prohibitions, and the designation of compliance officers. The bill also clarifies jurisdiction between the SEC and CFTC and replaces vague claims of "sufficient decentralization" with a statutory control test. The article refutes three main objections: labeling it as 'deregulation,' concerns about officials holding coins creating conflicts of interest, and fears that stablecoin yields would impact bank deposits, pointing out that the current lack of rules is the true source of risk.
Andon Labs opened an Andon Market store managed by AI in San Francisco and operates an Andon Café in Stockholm, where agents based on Anthropic, Google Gemini, and OpenAI GPT models are actually responsible for procurement.
In a blog post, Cohere co-founder and CEO Aidan Gomez questioned AI regulatory proposals pushed by top Silicon Valley labs such as Anthropic, arguing that seeking antitrust exemptions under the guise of safety effectively constitutes a technology cartel. He proposed a four-pillar alternative based on evidence-driven risk frameworks, mandatory transparency, capability-based rather than entity-based tiered testing, and truly independent safeguards. He emphasized that rules should be developed through multi-party participation and public verification, rather than being drafted behind closed doors by a few commercial companies.
Tomoko Tanaka, Head of Japan and Korea Markets at GitHub, introduces how to use GitHub Copilot as an agent to automate the entire marketing campaign process from planning to post-event follow-up: defining events via GitHub Issue forms.
Cohere has released North Small Translate, an open-source MoE machine translation model supporting over 50 languages, achieving a score of 83.6 on the WMT26 all-language benchmark.
Google Research published the ToolGrad paper at ACL 2026, proposing an answer-first paradigm: first generating real tool-call chains, then reverse-engineering user prompts.
The GitHub Copilot app has added three built-in panels: diff, terminal, and browser. These support completing code change reviews within a single application (the diff panel highlights additions, deletions, and modifications), local execution (the terminal panel runs commands such as npm run dev), and interface previewing and iteration (the browser panel works with the Pick & Polish tool). Finally, users can directly create pull requests, achieving a closed loop for AI coding.
Academia is accelerating the construction of high-quality tactile datasets to break through bottlenecks in dexterous manipulation: The Berkeley team trained a dedicated sub-model using 100 hours of high-quality tactile data, achieving a 65% task success rate by combining it with VLA; The Tsinghua team aggregated over 3,000 hours of data from 21 types of sensors, designing a hardware-agnostic model based on hand template mapping; Fudan University and NeoteAI collected over 30,000 hours of synchronized visual-tactile data to train models capable of tactile prediction; The USC team attempted to infer tactile signals from vision.
NVIDIA released the optimized NIM 2.0.12 service stack, achieving up to a 2.5x throughput increase for Nemotron 3 Ultra on 4xB200 systems (rising from 718 tok/s to 1997 tok/s), suitable for agentic workloads; it offers pre-validated profiles with optimizations including MTP speculative decoding, prefix reuse, and autotuned kernels, and supports benchmark reproduction via NVIDIA AIPerf.
NVIDIA has released the BioNeMo Inference Runtime (BioIR) to accelerate biomolecular structure prediction model inference on GPUs, supporting two usage methods: end-to-end processors and direct PyTorch integration.
ChatGPT Work has launched a Data Agent that supports connecting to company data, extracting insights using natural language, and building interactive dashboards.
a16z crypto published a paper titled "The Compatibility of Permissionless Networks and Financial Integrity: A Practical Guide for Financial Institutions," authored by Rebecca Rettig et al., which indicates that financial institutions can use permissionless networks such as Ethereum and Solana without violating the Bank Secrecy Act (BSA) and U.S. sanctions regulations; the core argument is that compliance obligations should fall on the application layer rather than the infrastructure layer, and institutions can fulfill these obligations through a risk-adapted tiered control framework comprising nine components, including customer-level due diligence, on-chain transaction monitoring, Travel Rule compliance, and protocol-agnostic sanctions controls; the text cites OCC Interpretive Letter 1186 and the GENIUS Act to support the legal feasibility of this approach.
a16z crypto published a paper titled "Permissionless Networks and Financial Compliance: A Practical Guide for Financial Institutions," stating that banks and other institutions mistakenly believe that permissionless blockchains (such as Ethereum and Solana) cannot meet Bank Secrecy Act (BSA) and sanctions regulatory requirements. In reality, existing laws allow them to build risk-adapted compliance frameworks at the application layer.
Cursor has introduced the Projects feature, supporting long-term context maintenance, delegation of thousands of sub-agents, and automatic execution of periodic tasks. It features three core capabilities: default cloud execution with on-demand local switching, cross-environment context sharing, and subscription-based proactive responses (such as monitoring Slack or PRs). It has been used internally for feature development, framework migration, and design system maintenance, resulting in a 30% increase in merged PRs for new users and up to six times for heavy users. The feature is now in beta and being gradually rolled out to all users.