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 DeepMind has released Gemini 3.8 Live with Live Avatar, providing near-real-time audio and video generation capabilities for Gemini Enterprise. It supports lip-syncing and facial expression adaptation across 97 languages, features asynchronous tool calling and brand-customizable virtual avatars, and employs SynthID watermarking to ensure content identifiability.
The AMD ROCm team released UltraQuant—a 4-bit KV cache compression method designed specifically for long-context agentic services, compatible with the Qwen3.8-2.4T-MXFP4 model and AMD Instinct MI355X GPUs.
The official AMD blog announced AMD Quark's capabilities for local quantization and multi-backend deployment on the Strix Halo platform, quantizing the Qwen3.6-35B-A3B model from BF16 to W4A16.
The official AMD ROCm blog introduces the Model Weight Profiles analysis method, demonstrating how to decompose parameter distributions by component (embeddings, attention, dense layers, etc.) and numerical format (FP32/BF16/FP8); this method can be used to accurately estimate model disk size, quantization benefits, and KV cache memory overhead.
At the Accelerate seller conference, Amazon announced that US sellers can now manage stores on eBay, Shopify, TikTok Shop, and Walmart directly within Seller Central. The new multi-channel selling tool is available for free and supports product listing synchronization, order aggregation, and cross-platform profit analysis. This feature marks a shift in Amazon's role from restricting third-party sales to becoming a hub for multi-channel operations, although it does not directly generate revenue.
Liquid AI released the experimental DSpark draft model LFM2.5-VL-3B-DSpark for LFM2.5-VL-3B, achieving up to 3.13x decoding acceleration (on M5 Max) and 2.66x (on H100) by adding 280M parameters (+8.9%).
NVIDIA, in collaboration with institutions including Google DeepMind and EMBL-EBI, has publicly released predicted 3D structure data for protein complexes from over 2,800 viruses in the AlphaFold database. This was achieved using the AlphaFold2 model optimized via the NVIDIA BioNeMo Inference Runtime. Additionally, the BioNeMo structural prediction pipeline has been open-sourced, enabling researchers to batch-generate predicted structures from protein sequences.
Shopify has integrated Meta's newly released personal AI agent, Muse, into its Agentic Storefronts program as an AI commerce channel, sharing merchant product catalogs by default and supporting direct checkout. In contrast, Amazon blocked Muse from accessing its platform, citing violations of its Terms of Use and stating that Meta lacked authorization. This divergence highlights the struggle for control over product discovery, advertising reach, and customer relationships in agent-driven e-commerce.
At Meta Connect 2026, Meta unveiled the Muse personal agent, which supports voice and real-time video interaction, Muse Mail (a dedicated email address), computer use on Mac, and integrates over 1,500 app connectors.
Anthropic released the Project Swap study, which involved having 201 employees engage in brief conversations (median length of 216 characters) with Claude agents to generate their preference rankings for 10 books, followed by completing book swaps in a digital trading floor.
This paper reports on applying the AutoResearch paradigm to optimize production-grade embedding systems within a book recommendation pipeline. Over 12 weeks and more than 220 experiments, it identified five new failure modes: infrastructure fragility, agent memory decay, search direction stagnation, asymmetric iteration costs, and metric fixation. It proposes a 'Prevention-Persistence-Redirection' three-principle framework, achieving a 1.82x improvement in Recall@6, a 2.1x improvement in coherence, and autonomously designing a text fallback solution that expanded catalog coverage by 5.8x.
NVIDIA has released NV-Reason-CT, an open-source vision-language model (VLM) designed specifically for full 3D CT analysis, supporting structured report generation, radiologist-style chain-of-thought reasoning, and multi-turn conversational follow-ups.
The real-world asset (RWA) perpetual futures market recorded $701 billion in trading volume in August, a year-on-year increase of 1,148 times; open interest reached $9.3 billion, an increase of more than 700 times. Of this, 85% of trading volume occurred on centralized exchanges, with the share of equities rising to 62% and commodities falling to 22%. Under US regulation, only one centralized platform is authorized to offer true perpetual contracts to US users, while all others restrict access.
The trading volume for real-world asset (RWA) perpetual futures reached $701 billion in August, an annual increase of 1,148 times; open interest reached $9.3 billion, an annual increase of over 700 times; among these, 85% of trading volume occurred on centralized exchanges, with the share of equity-based assets rising to 62%, commodity-based assets falling to 22%, and index-based assets accounting for 15%; the author notes that this growth indicates that gaining exposure to traditional markets via on-chain mechanisms has become a realistic trend.
NVIDIA has released an open-source Kubernetes controller called the Cluster Readiness Engine (NVCRE), which validates GPU clusters by running real distributed AI workloads (such as NCCL all-reduce).
The GitHub Copilot app has refactored the Pull Request view to support rendering massive Pull Requests containing 2,200 files, over one million lines of changes, and more than 400 inline comments. Core improvements include a dual-geometry model (deterministic code height + dynamic block height), a lazy measurement scheduling mechanism, scroll anchoring corrections, and a data pipeline that prioritizes streaming structure. This capability is now live in the GitHub Copilot app.
NVIDIA officially released the open-source project NodeWright as part of DSX OS, designed for declaratively configuring and securely updating Kubernetes node operating systems without interrupting training tasks.
Simon Willison developed the Gemini 3.8 TTS Playground tool, which supports trying out Google's newly released gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts models using your own API key.
Microsoft Research published a study titled "Offloaded inference for real-world physical AI robotics," which systematically evaluates mobile manipulation robot tasks (such as clearing kitchen waste).
Google DeepMind announced the introduction of private, persistent server-side memory functionality within its Private AI Compute platform. By leveraging hardware-isolated secure enclaves, end-to-end encrypted channels, and device-derived encryption keys, it enables long-term storage of cross-device AI contexts with strict privacy protection. Additionally, a technical white paper, software hash attestation, and third-party cybersecurity audit results were released.
NVIDIA's technical blog released SWE-Serve, a new benchmark evaluating the capabilities of AI coding agents in large model inference service engineering. It builds 53 executable tasks based on 83 merged SGLang PRs, covering six major engineering families: model enablement, decoding, caching, scheduling, API, and runtime performance; among them, 19 tasks launch real servers for end-to-end validation.
Google DeepMind has released two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, which support customizing character voices via natural language prompts.
Practical Ecommerce compiled over a dozen new e-commerce tools released on September 23, 2026: OpenAI launched Sponsored Agents ad functionality within ChatGPT, integrating with HubSpot and Shopify; Pinterest released Visual Search Ads supporting Performance+; The European Central Bank initiated a digital euro pilot; Maersk partnered with Shipstore to expand North American e-commerce delivery; Tulip launched an AI-assisted Integration Builder and expanded its Shopify connectors; Salesforce released Koa, a CRM reasoning model; Google Ads is testing Product Value Optimization; Alchemy AgentCard will support Mastercard payment credentials; FedEx introduced Authenticated Delivery; MikMak released AI commerce capabilities; Worldline integrated the Universal Commerce Protocol; Marvin launched AI-hosted Live Intercept interviews; Noibu released six types of AI agents for conversion optimization; Euka AI added Instagram integration; Snipp released plugins for ChatGPT and Claude; AIQ launched the Astro AI operating system.
Cursor's official blog announced multiple optimizations to its agent harness, including streamlining system prompts (reducing them by approximately 66%), on-demand tool loading, improving cache reuse rates (reducing cold cache miss rates by 20%), compressing line number annotations in file reads (reducing tokens by 1.6%), and strategically using sub-agents. These changes collectively reduced user token costs by 7% without compromising agent quality.