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 AMD ROCm Blog published a tutorial demonstrating how to run VLM-based physical AI agents on the AMD Ryzen AI Max+ (Strix Halo) platform using the Lemonade local inference server and the open-source RAI agent framework.
Researchers from Delft University of Technology (TU Delft) released a preprint on arXiv proposing a method that parses natural language requests (e.g., "I'm in a hurry, drive faster") via GPT-4o-mini and translates them into motion planner parameter adjustments. The system does not directly control the vehicle but dynamically adjusts weights for criteria such as speed and smoothness in path evaluation, explaining proposed adjustments to passengers in natural language and seeking confirmation. The effectiveness of eight types of instructions has been validated in the nuPlan simulator.
NVIDIA has released the Spectrum-X Ethernet architecture, designed specifically for thousand-GPU generative AI. It features three hardware-accelerated control loops: adaptive routing, end-to-end congestion control, and a Plane Load Balancer integrated into SuperNICs. Supporting bandwidths of 800 Gbps and above, it achieves near-zero performance degradation under multi-tenancy in DeepSeek-V3 training simulations and maintains proportional capacity degradation during link failures.
NVIDIA's technical blog released measured results for the Vera Rubin NVL72 and Blackwell GB300 NVL72 under the SemiAnalysis AgentX benchmark.
NVIDIA has released the Groq 3 LPX inference accelerator, designed specifically for the NVIDIA Vera Rubin platform, which works in synergy with Vera Rubin NVL72 to deliver high-interactivity long-context services.
NVIDIA released the BlueField-4 DPU as the core of the newly proposed Scale-In network infrastructure, providing line-rate networking for agentic AI factories by combining the DOCA software framework and Spectrum-X Ethernet.
NVIDIA has released the DSX MaxLPS technology suite, which includes Dynamic Power Allocation Software (DPS), advanced performance-per-watt optimization techniques, and a 45°C warm-water liquid cooling infrastructure design, aiming to enhance GPU throughput in AI factories under fixed power budgets.
NVIDIA released the Vera CPU in a technical blog post. The CPU adopts the Olympus core architecture and is optimized for the coexistence of long-sequence inference and bursty parallel calls characteristic of agentic workloads.
ElevenLabs officially released CLI v1, integrating the ElevenLabs API directly into the terminal. Designed specifically for coding agents, it offers structured JSON output, a --dry-run preview mode, agents-as-code workflows (supporting local configuration file management, diff previews, and pushes), branch isolation for development/production environments, and an agent-oriented --schema machine-readable interface description with parseable JSON error responses.
The Ethereum Foundation announced that the Glamsterdam upgrade will introduce gas repricing, including EIP-8037 (increasing and unifying the cost of new state creation) and EIP-8038 (increasing the cost of state access operations such as SSTORE, SLOAD, etc.), to more accurately reflect actual resource consumption under current state sizes. Historical transaction replays show that most contracts are unaffected, while a minority relying on hardcoded gas assumptions may fail. Developers can check their contracts via ethereum.github.io/repricing-impact/ and verify fixes on the Platåberget testnet (glam-devnet-8).
SGLang v0.5.18 has been officially released, adding support for Muse Glimmer (Autoregressive multimodal), Intern-S2-Mobius, and SANA-Video.
Google Research proposes the Biomarker Discovery Framework, a multi-agent system that iteratively performs hypothesis generation, statistical analysis, adversarial validation, and literature reasoning under human supervision.
An NVIDIA technical blog post released a GPU-accelerated financial instrument clustering workflow based on SymNMF, proposing a memory-efficient implementation and the adaptive solver AdaptGrow. It supports scales from single-GPU (100,000 instruments) to 16-node distributed systems (1 million instruments). The workflow integrates rolling correlation and tail dependence matrices, outputting hard cluster labels, soft factor loadings, and structural break signals, accompanied by complete runnable notebooks and NGC container environments.
Based on practices such as OpenShell, NVIDIA's security team proposed a layered security model for the AI agent stack. It emphasizes that the harness layer is responsible for behavioral guidance, while the runtime layer (such as OpenShell) handles authoritative infrastructure controls. The team noted that enhanced model capabilities amplify the risk of boundary violations, advocating for placing policy enforcement in underlying environments that agents cannot bypass. They also provided five design rules and four security configuration profiles.
NVIDIA's research project Agentic Variation Operators (AVO) is a general agent architecture designed for long-horizon autonomous tasks, featuring persistent memory and supervision mechanisms. It ran continuously for seven days on GPU kernel optimization tasks, producing multi-head attention kernels with performance superior to cuDNN and FlashAttention-4.
Google DeepMind published a long-form article reviewing the journey of game-driven AI research from Atari DQN to AlphaStar and AlphaFold, highlighting the Gemini-based general-purpose game agent SIMA 2.
Google Research proposes the Mobility-Embedded POIs (ME-POIs) framework by integrating textual descriptions with anonymized aggregated mobility patterns from public benchmarks (e.g., arrival times.
Hugging Face built a hybrid search system for Papers with Code based on PostgreSQL full-text search and pgvector vector search, using Jobs to batch generate 256-dimensional embedding vectors from Qwen/Qwen3-Embedding-0.6B.
Hugging Face released new research proposing three probe tests (consensus divergence, masked entity retrieval, and orthographic switching) to quantify the benchmark optimization phenomenon in speech recognition models. Evaluation of 11 open-source ASR models found that some high-scoring models reproduce erroneous reference texts from VoxPopuli or LibriSpeech, even outputting benchmark answers when audio is muted or contradicts the reference. The study also open-sourced detection scripts and unnormalized model outputs.
Liquid AI released LFM2.5-DSpark model checkpoints covering three models: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B.
An NVIDIA technical blog explains how Generative Recommenders (GR) replace traditional embedding similarity methods with sequence modeling to address production challenges such as data scale, long-tail items, cold starts, and low latency; it introduces two mainstream GR architectures: HSTU and Semantic IDs.
Microsoft Research has released Skala-1.1, a deep learning density functional theory (DFT) exchange-correlation functional trained on 2.5 times more data than its predecessor, significantly improving accuracy in key tasks such as thermochemistry, reaction kinetics, and molecular structure prediction.
DeepSeek officially released the experimental multimodal visual understanding model DeepSeek-V4-Flash-Vision-Exp, which can be invoked via model='deepseek-v4-flash-vision-exp'. The model matches the official DeepSeek-V4-Flash version in pure text capabilities but shows a significant leap in Agent benchmarks requiring visual understanding, with multimodal Agent capabilities approaching those of Opus-4.8.
Mistral has released Agentic Search, a new retrieval layer that supports multi-step retrieval loops (search / open / navigate / read / grep).