Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
AllenAI has open-sourced the AstaBrief 8B model, designed specifically for generating cited reports from scientific literature reviews. It supports usage as Fast mode within the Asta platform; the model is based on Qwen3-8B.
Google Research released a new federated learning system that uses Trusted Execution Environments (TEE) to achieve fully verifiable and auditable data anonymization, ensuring that only differentially private model weights and metrics are released; Gboard has deployed this system for next-word prediction in English and Japanese, significantly reducing training time while avoiding the impact of device availability fluctuations.
NVIDIA has released a new configuration of the DGX Spark 64GB, offered by Acer, ASUS, Dell, Gigabyte, HP, and MSI, featuring the GB10 Grace Blackwell Superchip.
The openJiuwen team has released X-Router self-evolving model routing technology as a core capability of its open-source AI Agent platform, supporting dynamic decision-making, multi-objective optimization, and feedback-driven strategy evolution; empirical tests show that it reduces costs by 44.6% compared to the full-cloud baseline on PinchBench while maintaining nearly equivalent scores.
Pi has launched version 1.0, featuring native Codemode support for MCP, Jev, and image models, virtual model extensions, Anthropic model cache pre-warming, and dynamic system messages within sessions. Pi Durable migrates Pi to TypeScript, achieving crash resilience, portability across JavaScript runtimes, concurrent multi-branch sessions, pluggable storage backends, hot-updatable tools and extensions, automatic context compression, and multi-user collaborative state synchronization. Both projects appeared on the Hacker News front page.
ServiceNow CoreAI introduces the AutoSynthData method, which automatically generates agent training tasks with feasibility, authenticity, and targeted difficulty by analyzing failure modes of target models in enterprise environments and successful trajectories from stronger teacher models. Each task includes system specifications, user prompts, and verifiers.
SGLang v0.5.21 has been officially released, adding support for DeepSeek-V4.1 Flash, GigaChat 3.5, IQuest-Q1, MiMo-V2.6 / MiMo-V2.6-Pro.
The paper proposes DuplexJev, which inputs ASR encoder hidden states into a frozen LLM via a small connector and outputs decision results as a single-token distribution without any decoding; it completes 80 speech-related decisions within 0.1 seconds on an 8-GPU node; using the last-layer connector achieves 90% accuracy in spoken question answering, while cross-attention connectors achieve 90% accuracy for gender and emotion recognition (up from 55% and 28%, respectively), with spoken question answering accuracy dropping by only 1 point to 82%; training employs cross-entropy loss on answer tokens rather than traditional transcription distillation, and open-sources weights, training schemes, batch inference pipelines, and a bilingual spoken question answering dataset.
Sail has launched an inference platform designed specifically for AI agents, featuring a restructured inference stack (supporting cross-vendor scheduling, utilizing idle data center resources, and rapidly adapting to new accelerators) and Sailboxes workspaces (with elastic scaling of CPU and memory, supporting continuous operation for months), enabling agents to consume billions of tokens to complete single tasks. Core team members come from NVIDIA, Apple, and Together AI.
Periscope is a training-free inference method that arranges long text chunks in a K×K grid, generates an evidence graph by posing questions through local and strided sampling to locate key paragraphs; evaluated on LongBench v2 and InfiniteBench.
Bitcoin Optech Newsletter #425 disclosed two denial-of-service vulnerabilities affecting Eclair v0.13.1 and earlier versions: first, parsing feature bits in init messages caused approximately 300 MB of memory allocation per message; second, accepting deprecated zlib-encoded gossip queries allowed 64 kB inputs to expand to 64 MB, exhausting node resources. Both vulnerabilities have been fixed in v0.14.0.
Graphify is an open-source tool released in April 2026 that automatically converts codebases, documentation, and unstructured data into queryable multimodal knowledge graphs, supporting the preservation of Terraform block attributes.
AWS Professional Services released a four-agent pattern for cloud migration scenarios, running on the Amazon Bedrock AgentCore platform, built with the Strands Agents SDK, and connecting to organizational systems via MCP tools.
Amazon Quick has introduced the Live Data in Apps feature, enabling AI-built Quick Apps to query governed QuickSight datasets (SPICE or Direct Query) in real time each time a user opens them, rather than relying on static snapshots from build time; queries are executed as the viewer, automatically applying existing Row-Level Security (RLS) and Column-Level Security (CLS) rules, and requiring users to authorize access to each dataset individually.
The Apache Iceberg community has added a read-restrictions field to the REST Catalog specification, supporting standard column masking (required-column-projections) and row filtering (required-row-filter).
NVIDIA announced the launch of DOCA AI Agent Skills in its official technical blog to address API misuse by general-purpose AI agents in BlueField DPU development due to a lack of domain knowledge.
NVIDIA has released the Do Inference Now (DIN) Deploy open-source C++ sample library, which integrates ONNX Runtime with the TensorRT RTX execution provider and supports Windows/Linux/x86-64/Arm64 platforms.
This article demonstrates in detail how to integrate Amazon S3 Vectors as a persistent memory backend into the NVIDIA NeMo Agent Toolkit (NAT), including a three-step implementation.
Amazon Payments built a multi-objective contextual multi-armed bandit (MAB) using LinUCB on AWS to jointly optimize content recommendations across the 3 stages of the product application funnel: launch, submission, and approval.
AWS published a reference implementation of an ambient agent based on Amazon Bedrock AgentCore in its Machine Learning Blog.
An official AWS tutorial guiding users to implement unified subscription and workspace-level isolated access for Claude Platform on AWS across three scenarios: production workloads (via cross-account SigV4), developer local (via workspace-scoped long-term API keys), and external environments (via OIDC federated identity and short-lived bearer tokens).
A research team from Carnegie Mellon University, MIT, New York University, and Stanford University developed an AI named Ataraxos, which defeated the recognized historically strongest player Pim Niemeijer in the game of Stratego with a record of 15 wins, 1 loss, and 4 draws; the model overcame the long-standing difficult problem of imperfect information games using only 16 GPUs and training costs of several thousand dollars.
Boston Dynamics announced a newly designed hand for its Atlas robot, increasing the degrees of freedom from 7 to 13. The design eliminates the pinky finger and adds a fourth finger, supporting finger abduction to accommodate trigger-operated tools. It employs embedded direct-drive actuators without fragile tendons, emphasizing mass producibility, reliability, and low-cost maintenance.
Shopify has released Canvas, a new design interface for online stores that offers an infinite canvas, real-time editing previews, Sidekick chat command editing, and true interactive previews across multiple devices. It supports creating or copying from existing themes but currently does not support third-party themes, app blocks, translations, or automatic theme updates.