Nov 9–15, 2026

Cursor’s November 13 update traces an incident from a vendor access key exposed at Fireworks to Cursor data and credentials, then to malicious Claude Code extension copies served through Cursor’s marketplace.

Summary The U.S. General Services Administration says its expanded OneGov agreement gives eligible government buyers a 40 discount on token-based use of Claude models available to government with no minimum order. The change broadens purchasing access but the announcement does

DeepSeek Harness has published dsh-v0.2.2-rc.1, a release candidate that adds experimental delegation to local Codex and Claude CLI, plus a worktree-based change-review workflow.

DuelLab’s November 11 GameBench 2 release ranks Claude Fable 5.2 first overall, with a 95.0 score across three reasoning settings.

Hedwigai compared Particle-1, DeepSeek V4.1 Flash and GLM-5.3-Flash inside Claude Code on 12 research-workbook tasks.

Terminal-Bench 4.1 hardens task verifiers against reward hacking and regrades saved Terminal-Bench 4.0 artifacts without additional inference.

Cloudflare describes Kitesurf as an experimental, stateless browser engine for agents that runs on Cloudflare Workers, renders pages without Chromium, and exposes the Chrome DevTools Protocol.

Plotly’s DataAnalyticsBench run gave Meta’s Muse Ember and GPT-6 Luna the same score, 38 of 43, but Ember’s full run cost $1.04 versus Luna’s $0.13.

Ollama’s v0.40.16 release adds Ripple Preview, a 65-billion-parameter dense model from Ricursive Intelligence, with a one-command launch path.

Science reports that a committee led by David Gross released a validation of a Lean formalization of OpenAI’s claimed result on the four-dimensional Yang–Mills existence problem.

Spacefast is presented as infrastructure for AI-built prototypes: an agent can send work to a service that handles hosting and databases, while users manage projects through an AI interface and the web.

AvenoxBeyin’s multi-machine guide explains how to use one synchronized vault across computers without synchronizing machine-specific runtime state or installation files.

Navers Lab and Einsia AI describe WorkflowArena, a benchmark of 145 long-horizon tasks for testing how LLM-based multi-agent systems organize work and deliver integrated results.

AWS describes a self-service gateway that puts Amazon Bedrock models behind LiteLLM on AWS Fargate and an OpenAI-compatible API.

A software engineer with no prior neural-network training experience built a transformer to judge positions in a CodinGame bot, using coding agents to write the model and training code.

Microsoft says Claude Fable 5.2 is beginning to roll out in Copilot Cowork and Copilot Studio, bringing Anthropic’s latest Fable model into two Microsoft work surfaces.

Ai2 has released XYBench, a benchmark of 8,115 questions designed to test whether AI assistants recognize when a user’s proposed approach conflicts with the goal they are trying to reach.

Exin Therapeutics connected an AI agent to a custom surgical rig and used it to select and inject a gene therapy into a living mouse’s brain.

Databricks announced Genie plugins for ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot, allowing those assistants to use a governed business-context layer.

Cinecode is an open-source workflow that asks a coding agent to turn a film idea into a rendered MP4, using code-built scenes, designed sound and review steps.

1Password’s November 5 support guide explains how Claude Duet can use approved credentials to sign in to websites during tasks.

Claude Code 2.1.323 adds an opt-in experimental mode in which Claude writes JavaScript to call tools, alongside version-range controls for plugins.

Anthropic launched Claude for Manufacturing, a plugin with 17 skills and connectors to Autodesk Fusion, Infor, Microsoft 365, and Oracle NetSuite.

DevoxxGenie 1.15.0 adds project-specific scanning rules, persistent favorite models, Markdown-based page retrieval, and a newer MCP client to its JetBrains plugin.

Vercel Labs’ durable-harness 0.0.1 release adds tracing that represents agent sessions as turns, model calls as steps, and tool calls beneath them, while preserving turn identity across resumed executions.

Anthropic has begun rolling out Claude Meetings in beta for Team and Enterprise customers, putting Claude in Google Meet as a visible participant that can answer questions, create shared work products, and capture meeting notes.

J.Crew launched Style Advisor, a Claude-powered shopping experience that recommends complete outfits from the retailer’s catalog photography rather than sending customers through separate product searches.

MiniMax has published MiniMax-H3-Image, an open-weight model for image generation and editing that the repository describes as supporting outputs around 16 megapixels and arbitrary aspect ratios.

Perplexity’s second SPACE red-team report says 12 models failed to escape the guest VM, including after a revised prompt directed them toward the virtualization boundary.

Simular has announced Box, a cloud computer designed for AI agents to share files, logins, apps and computer-control capabilities.

Vercel’s guide to agent evaluations argues that teams should test whether agents complete user tasks reliably, not merely whether individual calls look plausible.

QwenCloud’s Token Plan Individual documentation lists qwen4-max-preview, also billed under the alias qwen4.0-max-preview, for a two-week preview ending November 25, 2026.

Amazon says it replaced Claude Sonnet 4.6 with Strands Decider E4B for one high-volume internal ticket-labeling workflow, cutting estimated model cost from about $22,200 to $500 per million tickets.

Anthropic’s Claude Mythos 5.2 is an active model released on November 9, 2026, but access is limited to organizations verified through Anthropic programs.

Anthropic’s Supported Regions Policy, effective November 12, 2026, limits access to its apps, API, developer platform, and Claude models to listed countries and regions.

Anthropic’s Claude account showcased seven early experiments with Fable 5.2, ranging from code-built visuals to origami checked in a folding simulator and a 3D-printed landmark reconstructed from 40 photographs.

Microsoft Research describes Lazarus, a method for altering fewer than ten rarely active neurons to make a language model trigger malicious commands without retraining.

Medium has launched TK, a private notebook that combines writing tools, AI feedback and a memory layer designed to connect with external agents.

Cognition says Claude Mythos 5.2 is now available to eligible enterprise customers through Devin Security, paired with more than $100 million in credits for vulnerability scanning and remediation.

Raindrop says its simulated outage experiment found that agent swarms often propose plans but rarely produce a leader whose instructions demonstrably change teammates’ decisions.

Epoch AI says it has replaced the frequentist model behind its Epoch Capabilities Index (ECI) with a Bayesian model.

Anthropic says Claude can now help teams build spreadsheets, whiteboards and dashboards, while interactive connectors let partner apps appear inside Claude.

Anthropic is pairing Claude Fable 5.2’s November 9, 2026 launch with temporary usage offers for eligible paid Claude subscribers.

Anthropic says Claude Fable 5.2 designed three electronic prototypes with minimal prompting, coordinating circuit-board layout, enclosures, firmware, and manufacturing checks.

Cognition says Claude Fable 5.2 is now available in Devin and leads the company’s FrontierCode 2.0 benchmark at 49.1%, ahead of Fable 5.1 and Opus 5.5.

The Remote Labor Index paper evaluates AI agents on 240 projects drawn from real remote freelance work.

Epoch AI ranks Claude Fable 5.2 first in its Epoch Capabilities Index, with a score of 174 across 239 tracked models.

Vals AI reports that Claude Fable 5.2 ranked first among 36 models on its Vals Index v2.2, scoring 71.31% at a listed test cost of $17.94.

AWS says Claude Fable 5.2 and Claude Mythos 5.2 are available through Amazon Bedrock and Claude Platform on AWS.

llm-anthropic 0.31 adds Claude Fable 5.2 support to the llm command-line tool.

Simon Willison built Beep Locator with Claude Code to find an intermittent beep using several phones, tablets and computers as synchronized listeners.

Luigi Petrucco says he reconstructed substantial portions of Meta’s unpublished Muse Spark Charter by eliciting and comparing passages from the model’s chain of thought.

Meta introduced Muse Ember 1.0, a reasoning model aimed at repeated agent work such as tool calls, coding subtasks, routing and long-document parsing.

Artificial Analysis lists Meta’s Muse Ember (Max) as a proprietary reasoning model with text and image input, a one-million-token context window and API pricing of $0.10 per million input tokens and $0.50 per million output tokens.

OpenCode says Go Teams makes its coding-model subscription available to workspaces for $20 per member each month.

Ollama has added Ricursive Intelligence’s Ripple-65B Preview to its model library as ripple-preview, with a command-line launch path and integrations for several coding-agent tools.

Vals AI’s Vibe Code Bench v1.1 ranks Claude Fable 5.2 first for building web apps from scratch, with a 93.77% score in OpenHands.

Delta has added an importer for Conductor conversations as Conductor Cloud approaches shutdown, with new messages ending November 13 and data deletion beginning when the service closes on November 20.

Vercel AI Gateway now lists Anthropic’s Claude Fable 5.2, with access through the AI SDK and several API formats.

Apiiro says it found 2,539 real secrets across 1,994 public repositories while examining coding-agent configuration and related files.

Cursor’s model documentation now recommends Claude Fable 5.2, while its detailed page identifies the model as claude-fable-5 and sets out its access, pricing, tools and data-retention conditions.

Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2, with lower token prices, faster generation and stronger scores on many of its reported evaluations.

Meta has released Muse Ember, a multimodal model for repeated agent work, on the Meta Model API and in Muse Code.

Vercel says Meta’s Muse Ember is now available through AI Gateway, with text, image and video input, a 1M-token context window, tool calling and adjustable reasoning effort.

Simon Willison used Claude Opus 5.5 to prototype a SQLite parse_number() extension in C and Rust, then benchmarked the implementations over two million rows.

Anthropic tested Claude agents in EVE Online market simulations and in teams playing EVE Frontier.

ReelMimic is an open-source project that turns a video reference and a brief into a new 2D animation, using Claude Code or Codex as a crew of planning, production, and review agents.

Anthropic tested three Claude versions on four unpublished synthesis targets shared by chemists at Stanford, Yale, and Caltech.

Perplexity’s Decider v1.3 is a multimodal classification model built on Qwen3.8-27B, with a new Locate mode that returns verbatim text spans, character offsets and scores.

Luca Martino’s review connects noise-contrastive estimation, reverse logistic regression, Bennett’s acceptance ratio, multiple importance sampling and bridge sampling for energy-based models.

The authors of a new paper propose a deterministic world model that maps physical states directly to camera images, then uses it to make reachability analysis of vision-based controllers tighter.

AIPOCH Open-Science v1.2.2 adds an opt-in strict networking mode for Notebook, alongside resumable research-film generation, PDF improvements, and Qwen4 Max Preview access through Bailian Token Plans in Beijing and Singapore.

AWS says eligible customers can now call OpenAI’s Daybreak Red and Daybreak Blue cybersecurity models through Amazon Bedrock’s bedrock-runtime endpoint.

CleanYourTV 3.5 adds eight themes, three add-ons, and support entries for more television models while reporting a substantial Home-screen memory reduction after repeated settings changes.

Vercel’s announcement for fx v0.0.17 lists three changes: steering subagents while they work, built-in Token Pass models, and tool definitions that are 48.0% smaller.

Higurashi has released Haru-4B, an open pretrained model designed to listen and speak continuously rather than taking turns in separate speech input and output stages.

Hugging Face’s tokenizers v1.0.0-rc.4 restores Python APIs used by libraries such as Transformers and vLLM, while changing the boundary between editable tokenizer configuration and the compiled object that encodes text.

InferSSD is an open-source llama.cpp-based runtime that streams mixture-of-experts model weights from SSD, aiming to run models larger than a machine’s RAM or VRAM.

FedBNN is a method for training binary neural networks inside federated learning, aiming to reduce the inference burden on resource-constrained devices without giving up the accuracy of full-precision federated models.

Langfuse says AI Gateway users can now restrict the models available to each API key, using model patterns such as openai-prod:*.

TechCrunch reports that an analysis of open model weights found signs of Nvidia Blackwell hardware in training runs for leading Chinese models.

The CoRL 2026 Workshop on Agentic Robotics named Agentic World Action Models, by UC Berkeley, NVIDIA and the University of Pennsylvania, a Best Paper Award winner.

World Labs presents Atlas as a world model with tools for reconstructing, shaping, exploring, and exporting 3D environments.

Cmpnd released dspy-wsp, a prototype DSPy optimizer that lets a frontier model inspect and edit another model’s weights without gradients or training.

Continuous Intelligence has released ci-proto, a family of five time-series foundation models that the company says can forecast unseen datasets without additional training.

DeepSeek’s chatbot guide documents web search, deep-thinking search, file input, URL reading and basic content-safety checks for API-built chatbots using deepseek-flash.

AWS says SGLang-Neuron is now publicly available, bringing SGLang’s serving interface and reinforcement-learning rollout generation to AWS Trainium.

Edge0 has released Edge0-145B and Edge0-120B, latent mixture-of-experts models paired with its edge0-strata streaming inference runtime.

Encevra has released Athena-27B-Preview, a small LoRA adapter for Qwen3.8-27B that answers questions from supplied documents, quotes supporting text and abstains when evidence is missing.

GATE-AD is a few-shot method for industrial visual anomaly detection that models relationships among neighboring image patches rather than judging each patch in isolation.

Liquid AI has published LiquidCodec, a compact, streaming codec that turns speech into discrete tokens and reconstructs it as audio.

Mistral’s model catalog now lists DeepSeek V4.1 Flash 1017 in Public Preview and Z.ai GLM 5.3 Flash as generally available.

Fish Audio is presenting Live 3 as the full-duplex model behind a broader voice-agent platform for calls, websites, apps and APIs.

Perplexity has released Decider v1.3 and Decider Flash v1.3, decision models for classification, scoring, routing and, newly, extracting text spans.

Recursive’s Flint model card reports 89.7% on SWE-bench Pro v2, alongside scores on SWE-bench Verified and Terminal-Bench 2.1.

Salesforce has released SOV-Transcribe-1, an open-weight speech-recognition model with 4 billion active parameters and an Apache 2.0 license.

Terradot and Microsoft have developed TerraNova, an AI model that predicts how crushed rock applied to farmland may remove atmospheric carbon dioxide.

DeepSeek says the capabilities behind its app are now available through its API, including web search, file input, URL reading, safety checks and voice-related standalone services.

ImplicitBBQ is a question-answering benchmark for testing whether language models respond differently when demographic identity is conveyed through associated characteristics rather than stated directly.

OpenCode’s models.dev repository merged a change restoring MiniMax M3.1 Flash Free on OpenCode’s Free tier for a limited time.

Huawei Computing Systems Lab has published MACaQ, an Apache-2.0 quantization toolkit with workflows for text, image, video and audio models across several backends.

IBM has released experimental Granite 4.1 3B and 8B preview models trained to consume and generate LLMON, a markup format that explicitly separates instructions, data, and requirements.

Researchers report that language models can learn some multi-step tasks while emitting only filler tokens between a prompt and an answer.

Nvidia says it will invest $3 billion in Meta’s $95 billion stock offering while the companies expand work on Meta’s Muse agent platform and model family.

Phonon 0.5.0 teaches Gradium’s Rust speech runtime to load the architecture variants used by Kyutai’s small Pocket TTS models and corrects Mimi’s streaming attention window.

DoorDash says it launched its first production transformer for homepage store ranking in mid-2025, adding sequence encoding to an established multi-task ranker rather than replacing it.

Shopify says a shared generative model for search and recommendations is now in production, with reported gains in Shop home-feed outcomes and search relevance.

Arya Somu’s ConvictionBench tests whether language models revise answers for legitimate reasons, including whether they distinguish user-pasted records from authenticated tool updates.

Ken Goldberg’s 2025 Science Robotics perspective argues that useful robots can help close the gap between scarce real-world robot data and the enormous datasets used to train language models.

Baseten says it optimized inference for DeepSeek V4.1 Flash 1017 by iterating on more than 80 kernels and applying kernel fusion, Programmatic Dependent Launch, and host-loop changes.

AWS says Kimi K3.1 Pro Highspeed from Moonshot AI is now available through Amazon Bedrock, with the same 3T-class architecture and one-million-token context window as Kimi K3.1 Pro.

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence has released PAN 2, an open world model that turns a starting scene and a described action into a simulated video of what might happen next.

Goodfire reports experiments on Qwen3.6-35B-A3B suggesting that individual MLP neurons can be treated as interpretable computational units.

A team from Fudan University, Alibaba, and the University of Southern California proposes AdaVaR, a training framework that teaches a vision-language model to use both text-based and visually grounded reasoning, then select between them by context.

Robbyant has presented LingBot-Touch, a tactile representation model intended to transfer contact information across sensor types, robot hands, and manipulation tasks.

BALD-SAM is a research method for choosing the next spatial prompt in interactive image segmentation by estimating where another human input would be most informative.

Hugging Face’s technical article describes decision models: systems that evaluate a state against application-defined questions and answer options, returning structured decisions with probabilities rather than generating prose.

Sebastian Raschka’s November 10, 2026 analysis says the LLMs he tracks have shifted toward hybrid attention: in the second half of 2026, models split evenly between those retaining some full attention and those using none.

Amazon Science has introduced Mitra-v2, an open tabular foundation model for classification and regression, with inference and fine-tuning code released under Apache-2.0.

Nathan Barry reports that a linear program can reproduce much of GPT-6 Astra’s edit to Qwen3.5-0.8B, which reduced repetitive reasoning loops, while changing roughly eight times fewer weights.

Scantist’s GLM-5.3-SLIM-GGUF model card describes a 179.86 GB, 192-expert-per-layer variant of GLM-5.3 that runs entirely on one NVIDIA B200.

AutoCompress is a research framework for compressing language models against the latency, throughput, and memory limits of a specific deployment target rather than against parameter count alone.

AWS says NVIDIA Nemotron 3.5 Super VL is now deployable through Amazon SageMaker JumpStart in NVFP4 and BF16 variants.

Pinecone has added OpenAI's text-embedding-3-small and text-embedding-3-large to integrated embedding for Documents API indexes.

Zyphra announced ZONOS3-Mini, an Apache 2.0 text-to-speech model with 100 million active parameters, zero-shot voice cloning, and runtimes for browsers, servers, and local devices.

169Pi has released Alpie Seed 9B, an Apache-2.0 open-weight vision-language model built on Qwen3.5-9B and post-trained for reasoning, coding, and visual understanding.

NVIDIA’s Cosmos3-Nano model card describes a 16-billion-parameter world model that takes text, images, video, audio and action trajectories, and can generate actions as well as media and text.

Tencent has published a BF16 DSpine draft-model checkpoint for Qwen3.8-27B, designed to accelerate decoding through speculative generation.

NVIDIA and AI4Bharat describe an open corpus of roughly 2.83 trillion tokens across 22 scheduled Indian languages, alongside a reusable synthetic-data pipeline and an Indic-language evaluation suite.

NVIDIA announced GR00T 2, an open world action model for robots, alongside Isaac Lab 3.0 simulation tools and new evaluation benchmarks at CoRL.

Handy Computer has published GGUF quantizations of Moondream’s Parakeet Ultra for use with transcribe.cpp, ranging from a 2.51 GB F32 file to a 486 MB Q4_K_M file.

PS5LM Plus is a native app that runs GGUF language models on a jailbroken PlayStation 5’s GPU using llama.cpp.

NVIDIA’s Kumo Relational is a pretrained foundation model for classification and regression across related tables, using in-context learning.

A paper by Daniil Plyusov and coauthors proposes F-GRPO, a modification to group-relative policy optimization intended to keep rare correct responses from being suppressed during reinforcement learning with verifiable rewards.

Moondream’s Parakeet Ultra is a post-trained version of NVIDIA’s 0.6-billion-parameter Parakeet TDT 0.6B v3, with lower reported word error rates across the benchmark groups on its model card.

VulcanBench reports 115 runs of Reflection Beam 1.0 in its Frontier v4 coding evaluation, with the model passing six of 23 reconstruction tasks at its best effort level.

Google Cloud says Batch now supports jobs on single-host TPU slices, adding a provisioning option for accelerated HPC and AI/ML workloads.

Google Cloud’s November 10, 2026 announcement describes an open-source Kubernetes stack for RL post-training, distributed inference, Ray workloads and agent execution.

Google says Gemini app users in the United States can connect Google Wallet so the assistant can answer questions using spending data.

Google released A2UI 1.0, the first stable version of its open protocol for agent-generated interfaces.

Google says Gemini 4 Hydrogen is now available globally in Search for all users, bringing stronger handling of complex queries, broader link discovery and more visual responses.

Google DeepMind’s Gemini 4 Hydrogen model card describes a model for fast coding and agentic workflows, with a context window of up to one million tokens and benchmark results across several task categories.

Langfuse now supports evaluator runs using Google's gemini-decision-1 through a Google AI Studio connection.

Langfuse says its AI Gateway can now call models through Microsoft Azure (Foundry) and Google AI Studio, including Gemini models.

Google DeepMind released MuJoCo 4.0.1 on November 11, 2026, adding Windows support for the MuJoCo Studio Web Viewer and fixing a sparse Cholesky factorization bug introduced in 4.0.0.

Google DeepMind’s AlphaGenome repository has released version 2.0.0, introducing a v2 model API for genomic prediction, scoring, embeddings, and analysis.

Replit’s November 13, 2026 changelog adds Agent workflows for creating Google Ads campaigns and preparing purchases through Stripe’s Link wallet.

Google’s Gemini API documentation describes Gemini 4 Carbon, a reasoning model available through the paid API tier and Google AI Studio.

Google's HappyDelivery prototype combines verified messaging webhooks, Pub/Sub, Redis session state and a LangGraph swarm of Gemini-powered specialists for conversational grocery orders.

Free My TV 1.2.0 adds a project-signed APK for Google TV and Android TV, along with wider device support, a sleep timer and photo-based screensaver options.

Google has opened pre-orders for intelligent eyewear collections from Warby Parker and Gentle Monster, built with Samsung and integrating Gemini.

Google DeepMind has released SatWeave and PARDIG, two Earth-observation datasets used in its WeatherNext 3 forecasting model.

Google says researchers working with Australia’s CSIRO used a version of DeepVariant to analyze DNA from 600 wild giant kelp specimens collected at 40 Tasmanian sites.

Google Cloud’s SKU Groups page lists Gemini 4 Carbon billing entries dated November 7, 2026, covering text input and output, cached inputs across modalities, and several service modes.

Google’s Gemini API documentation describes Managed Agents as configurable agents provisioned in Linux sandboxes, with code execution, file handling and web access.

Google’s Antigravity documentation describes a built-in browser agent that can test live pages, inspect console and network failures, capture screenshots, and verify visual fixes.

Google announced Gemini 4 Hydrogen, a model positioned for fast execution across everyday coding and agent workflows, alongside a model-switching feature in Google Antigravity.

Modal says it will join gVisor’s maintainers as Google moves the project, including its name and trademarks, to the CNCF under maintainer-led governance.

Pinecone’s Documents API now supports Google’s Gemini Embedding 2 for text fields, with Pinecone handling embedding calls during writes and searches.

Google Research's paper introduces SensorWM, a wearable-data model that forecasts physiological trajectories under specified activity, sleep and exercise behaviors.

Hugging Face published scripts and container workflows for fine-tuning TRELLIS.2 on custom 3D assets with LoRA.

Google’s A2UI 1.0 is an open protocol for AI agents to render interactive, native screens inside web, mobile, and desktop applications.

Alphabet's Q3 2026 earnings call described strong demand for Google Cloud AI infrastructure and TPU systems, while large capital spending and inventory buildup weighed on cash flow.

Google says Modern Web Guidance 1.0 is now stable, with a workflow that scans existing web projects for selected modernization opportunities and guides coding agents through changes.

Google Cloud’s developer guide documents Gemini 4 Carbon on Gemini Enterprise Agent Platform, with configuration in Agent Studio and access through APIs.

Google says Gemini 4 Carbon is rolling out in Antigravity alongside a preview of /swarm, its native multi-agent orchestration mode.

Google Antigravity’s /swarm command, now in preview, turns a conversation into a coordinator for named agents that work in parallel and communicate on a shared board.

Google Chrome’s Modern Web Guidance packages expert-curated web-platform recommendations as skills that coding agents can retrieve through a CLI.

Google’s TPU Sync repository describes an open-source C++ library for managing KV caches, memory tiers and data transfer on Cloud TPU.

Epoch AI’s AI chip sales database estimates shipments, compute capacity, power use, and spending across major accelerator designers.

Google Public Sector says it is working with NASA teams on a proof-of-concept robot for deep-space crew support, combining logistics handling with possible medical assistance.

Epoch AI estimates that quarterly additions of AI-chip computing capacity grew 2.5× per year from Q1 2024 through Q2 2026, equivalent to a doubling about every 9.1 months.

Google says developers can now build offline speech input and speech output into applications using its on-device APIs and model bundles.

Google says MaxText v0.2.4 can train models on sequences as long as 14 million tokens on Cloud TPUs, using four context-parallel strategies to distribute attention work across chips.

Google has added a Knowledge Work layout to Antigravity alongside Software Engineering.

Google says Gemini 4 Carbon is now available to developers, enterprises, scientists, and consumers, positioning it for demanding coding, knowledge-work, and research tasks.

Google says developers can build with Gemini 4 Carbon, a model it describes as its strongest for cybersecurity and long-horizon work.

Google says Agent Substrate 1.0 is ready for testing and that it is donating Agent Substrate and gVisor to the CNCF Sandbox, alongside OpenRL.

Models.dev lists Meta’s Muse Ember as available through the OpenCode provider, with a 1,048,576-token context limit and listed token prices.

Google Cloud’s Model Garden page lists Meta’s Muse Wonder, Muse Spark 1.3 and Muse Ember as models developers can deploy through its platform.

Google DeepMind has released MuJoCo 4.0 alongside MuJoCo Playground 1.0, reshaping the robotics simulator around a modular Studio SDK, browser-based viewing, and built-in deformable-body simulation.

Google has published ASIMOV Agentic, an open-source evaluation harness and dataset for testing AI systems on safety-critical robotics tasks.

Google says developers can now use Google Maps Platform directly inside Google AI Studio to build apps with real-world context and preview them interactively.

Google has added TPU 8tx for confidential inference and TPU 8i Max for large-context workloads to its eighth-generation accelerator lineup.

Sarvam AI has named Sashikumar Sreedharan its president and chief business officer, and says he will join its board.

VieNeu-TTS presents v3 Turbo as an open-source text-to-speech option for Vietnamese and English, with 48 kHz output, instant voice cloning, and local inference paths.

GitHub Agentic Workflows v1.2.0 requires an explicit opt-in for unrestricted pull_request_target runs on pull requests from forks.

GitHub Copilot CLI 1.0.111 adds regular-expression matching to enterprise-managed permission rules for Shell, Bash, and PowerShell.

GitHub Copilot CLI 1.0.112 makes AI credits already used in a session count when that session is resumed, and makes subagents honor the session-wide --max-ai-credits cap.

GitHub’s November 5, 2026 changelog adds SpaceX to its secret scanning partner program and introduces detection for starlink_public_api_v2_secret.

GitHub’s November 9 update says Agentic Workflows v1.2.1 adds artifact-backed continuation across agents, a specification-aware pull request reviewer, and security reviews that fail closed when findings remain unresolved.

GitHub’s November 6 Copilot changelog bundles updates across its app, CLI, JetBrains, Xcode and VS Code.

GitHub has added direct links for individual alerts in its Public monitoring view.

GitHub Secret Scanning now detects Cloudflare Access service tokens through a dedicated detector.

GitHub says Actions will block pull_request_target events by default in public repositories starting December 2, 2026, unless an applicable Actions event policy explicitly allows them.

Visual Studio 2026 version 18.12 switches Copilot Chat to the Microsoft Copilot SDK by default, bringing the same agent harness used in GitHub Copilot CLI, the GitHub Copilot app and VS Code.

Vercel Security Review can now inspect code changes in GitHub repositories that are not linked to a Vercel project.

GitHub Next has released Ace, an experimental open-source tool for teams to collaborate with coding agents running on their own machines.

GitHub’s guide shows how teams can package repeatable engineering work as agent skills: versioned directories centered on SKILL.md, reusable across clients while tools, credentials, and permissions remain with the host.

Depi Security says it chained prompt injection in OpenAI Codex’s GitHub issue-labeling workflow with a sandbox escape and runner-secret exposure, reaching a route toward publishing the official @openai/codex npm package.

GitHub is rolling out shared Copilot Automations, letting teams create and maintain repository-owned cloud workflows rather than keeping each automation personal to its creator.

Agent Terminal v0.7.0 adds support for Linux desktops and says Snap and AUR releases are in preparation.

GitGuardian says it found 31,088 distinct secrets in Claude Code’s project-level settings.local.json files across 23,276 public GitHub repositories, with 5,343 still valid when detected.

EMERGE is a benchmark for measuring how systems update knowledge graphs when new textual evidence arrives.

Codechart is an open-source agent skill and CLI that turns code changes, plans, and explanations into maps and flowcharts checked against repository files and Git diffs.

GitHub says Claude Fable 5.2 is generally available in Copilot for paid individual and business plans, with a gradual rollout across a broad set of clients.

GitHub says Gemini 4 Carbon is now generally available in Copilot for individual and organizational plans across a broad set of development environments.

GitHub’s Copilot CLI 1.0.109 adds Claude Fable 5.2 to its model picker and IPv4/IPv6 CIDR support for sandbox host allow and deny rules.

AnyLanguageModel 1.3.0 makes tool calls within a single model round run concurrently across providers, while preserving the model’s call order in delegate decisions and returned results.

The FBI has published a FLASH warning that attackers are using ARTEX, an open-source, LLM-driven penetration-testing framework, to automate intrusion campaigns.

Anthropic’s November 11 SEC filing attaches an approved prospectus for a European retail share offering as part of its global IPO.

Anthropic’s self-hosted Claude Managed Agents documentation details how organizations can run the service’s server, database and worker sandboxes in their own environment.

Anthropic published a guide to building agents that sleep between runs, resume from events, and carry state across sessions.

Anthropic published guidance for building safety evaluations, arguing that a score is only meaningful when the test measures the right behavior in realistic conditions.

New York Governor Kathy Hochul signed the RAISE Act on December 19, 2025, requiring large frontier AI developers to publish safety-framework information and report critical-harm incidents to the state within 72 hours of determining one occurred.

Anthropic says its frontier model Mythos helped bring Claude onto AMD MI355X GPUs in about three days, with the model writing or porting nearly all required kernels.

Anthropic’s IPO website sets out a proposed US share-price range of $70–$78, a planned Nasdaq listing under ANT, and expected pricing and trading dates, subject to regulatory clearance.

AWS published a deployable reference architecture for agents in separate organizations to collaborate through A2A despite using different frameworks.

Four nonprofits say they have sued Anthropic, Google, Meta, OpenAI and SpaceXSI in San Francisco, asking a California court to freeze development of the most advanced AI systems.

LlamaIndex says it tested LlamaParse in production against all ten OWASP Top 10 for LLM Applications risks, alongside standard web-application risks.

Anthropic has opened a private beta of self-hosted Claude Managed Agents, letting teams run the agent harness, sandboxes, credentials, and session history in their own cloud or on-premises environment.

Anthropic disclosed roughly $10 billion in additional cloud-server commitments to Microsoft and Google through 2033, while retaining the right to cancel a $34.9 billion Lambda deal over a missed data-center deadline.

Anthropic has started the roadshow for a proposed initial public offering of 825 million shares at an expected $70–$78 per share and seeks a Nasdaq listing under ANT.

Goldman Sachs forecasts that Anthropic’s revenue will exceed $600 billion in 2030, while the company remains deeply investment-intensive.

Epoch AI estimates that Broadcom’s TPU compute shipments doubled from fiscal Q2 to Q3 2026, reaching nearly 2 million Nvidia H100-equivalents.

Anthropic has launched the roadshow for a proposed initial public offering of Class A common stock.

Major investors are increasingly skeptical that Anthropic can achieve a $1.5 trillion valuation in a planned initial public offering, The Wall Street Journal reports.

Anthropic committed $3.4 billion to secure 1.7 gigawatts of gas-fired generators, taking over equipment orders originally placed by supplier OnePWR Solutions.

Ramp’s September 2026 index says 55.9% of businesses in its U.S.

Scientific American reports that some researchers are reconsidering commercial AI tools after companies publicized discoveries resembling work users had shared with chatbots.

Hugging Face’s LeRobot v0.7.0 release rebuilds installation, training, inference and hardware support, and adds dataset and policy capabilities.

NVIDIA’s model-card update documents native SGLang support for its NVFP4 DeepSeek-V4.1-Flash-1017 checkpoint, alongside the previously documented vLLM path.

Inferact and the vLLM team report serving DeepSeek-V4.1-Flash on one NVIDIA Vera Rubin NVL72 rack at 1.7 million tokens per second per GPU, with 387 tokens per second per user in the SemiAnalysis AgentX benchmark.

IST-DASLab has released code and checkpoints for disaggregated quantization, a method that assigns different quantization choices to prompt processing and token generation.

DetLLM is an open-source inference pipeline for Qwen3-0.6B that uses integer arithmetic throughout the forward pass.

Lightning Labs’ pi-402 is an open-source extension that lets the Pi coding agent pay for metered model inference through an embedded wallet, rather than using an API key.

Liquid AI introduced LiquidCodec, an approximately 74-million-parameter audio codec designed for its speech models and on-device inference.

Cerebras reported $210.8 million in GAAP revenue for Q3 2026 and said its inference offering will become available through Oracle Cloud Infrastructure in 2027.

vLLM says its Flint-specific serving support is available from day one, covering Recursive’s NVFP4 and BF16 checkpoints.

Hindsight's hs-rerank-1 is a 114-million-parameter multilingual cross-encoder reranker, released under the MIT license with ONNX and safetensors formats.

oMLX 0.8.0rc1 adds a Prefill Accelerator that lets a faster NVIDIA-equipped PC or second Mac handle the first layers of prompt chunks while a primary Mac completes inference.

LMSYS says a custom CUDA/PTX megakernel raised Muse Spark Open decoding speed from 517 to 1,109 tokens per second on eight NVIDIA GB300 GPUs in a single-request GSM8K test.

Runway has released Praxis-1, a 27-billion-parameter world-action model for robotics, with weights and inference and fine-tuning code under Apache 2.0.

Meta has released SAM 3D Body, a promptable model and inference repository for recovering a full-body 3D human mesh from a single image.

Baseten says Genspark post-trained the open Qwen3.5-397B model for its Deep Research agent, raising its score on Genspark’s grader from 0.612 to 0.759.

Epoch AI has introduced an Inference Performance Index (IPI) that compares chips by modeled concurrent inference capacity, rather than raw FLOP/s alone.

NEAR AI says an attacker used a remote code execution vulnerability in SGLang to access one inference server and reach internal systems.

NVIDIA has released Snapshot, an open-source Kubernetes component that checkpoints initialized GPU workloads and restores them without repeating model loading, compilation, or CUDA graph capture.

Fireworks AI's November 9 update says an October incident involving a third-party cloud provider extended into its development environment and affected customer billing data, with secrets, datasets and incomplete model weights exposed for a limited set of accounts.

Baidu says ERNIE Evolve can select research directions, modify code, run experiments and retain changes only after independent evaluation.

OpenAI says its Control Room tracks research workloads and can block further model output through authorization checks in the inference layer.

SemiAnalysis InferenceX ranks GPUs serving Meta’s Muse Spark Open on an agentic coding workload, measuring throughput at the same per-user interaction target.

LangChain Academy announced a course on building and deploying managed agents with LangSmith Managed Agents.

LangChain’s account of Actively AI describes a redesign of sales agents: instead of assigning one agent to each prospect account, the company grouped work by repeatable sales jobs called “motions.” The change matters because it made similar work easier to review together and allowed feedback to

Meta’s Chrome Web Store listing describes Muse as a personal AI agent available in Chrome’s side panel, where it can work on browsing-related tasks and connect to personal apps.

Sesame's plans page lists Free, Plus, and Pro tiers for its voice assistant, with separate allowances for live conversation and agent tasks.

Perplexity’s November 6 changelog adds persistent Teammates, fully local Computer on supported Macs, and voice control for Computer tasks.

NVIDIA’s Hugging Face walkthrough builds a simulated pick-and-place task in Isaac Sim 6.2.0 from reusable SimReady robot and object assets, with an agent helping assemble the scene and generate controls.

Cognition’s Devin Apps lets a team describe an internal web tool and have Devin build and host it on Devin’s cloud machine, behind sign-in and configurable access controls.

Natera built SWARM, an agentic AI system for laboratory operations on Amazon Bedrock AgentCore.

Krea’s beta page introduces Krea Bots, persistent AI agents assigned to specific creative crafts and coordinated in shared chats.

LangChain says it has raised $300 million in a Series C round at a $3 billion valuation, led by existing investor CapitalG.

OpenAI is rolling out ChatGPT Wallet in dots, letting eligible users approve purchases made by an agent without sharing their actual card details with the agent or merchant.

Arkose Labs says Meta Muse began signing some web traffic with Web Bot Auth on November 6, while Grok Bot’s traffic shifted sharply toward Cloudflare IP space.

United says third-party AI agents may show travelers its flights but cannot complete ticket purchases for them.

LangChain says its internal GTM Agent generated 1,540 sales artifacts over 12 weeks, saving representatives thousands of hours across workflows.

n8n Labs has released Nod, a free, open-source research-preview app that runs a personal agent on a Mac or Windows computer.

Browser Use 1.0 replaces most of its library with an agent that writes Python in a persistent REPL connected directly to Chrome DevTools Protocol.

DoubleAI says it placed an AI agent inside neural-network training runs and let it intervene in what the network learns, including by changing its architecture.

ElevenLabs has launched Campaign Manager, a feature in ElevenAgents for running outbound campaigns toward defined outcomes such as qualified leads or confirmed appointments.

OpenClaw’s v2026.11.2 is a patch release focused on reliability rather than new features.

Perplexity is launching Teammates, persistent agents inside Perplexity Computer that can own ongoing work across projects, tools, and repeated runs.

67.js is an MIT-licensed WebGL2 3D engine built for AI agents to drive, with an MCP server and structured support for assets and recipes.

Factory’s November 14, 2026 changelog lists CLI and Desktop v1.12.0, with changes to tool-call visibility, Git-operation risk labels, service-account auditing, and session behavior.

Frigade’s Yogurt is a free, MIT-licensed desktop app for recording meetings and creating local transcripts, summaries and answers.

OpenCode v2.3.0 makes subagents run in the background by default: they return a session ID immediately and report to the parent when idle.

Zed’s v1.28.0-pre preview adds optional automatic Git fetching and several workflow changes, alongside fixes to agent context handling and protocol compatibility.

Databricks says Genie Ontology is generally available as a shared layer of enterprise context that agents can learn from and people can govern.

OpenRouter compared a GitHub CLI, a skill, and MCP tools on a repository lookup task, measuring input tokens and billed cost across several models.

OpenAI has released MCP Extensions, an open-source SDK and specification for adding ChatGPT-specific interface capabilities to MCP servers and apps.

The Model Context Protocol project is moving specification releases to a regular six-month cadence, with releases generally planned for the last Tuesday of March and September.

Earshot 0.4.1 simplifies installation and reduces the reported idle resource footprint of its MCP server, while fixing input, file-matching, and caching errors.

Epic says Unreal Engine 5.9 MegaJam Preview is the first Unreal Engine release built primarily with agentic workflows.

The Model Context Protocol team has published an experimental extension for describing file inputs and moving files between clients, MCP servers, and external storage.

Delta documents Cloud Runners, Linux sandboxes that execute agent turns for a thread without relying on the user’s computer.

Parallel says its Search API is now purchasable through AWS Marketplace and can be called by Amazon Bedrock agents or connected through Bedrock AgentCore as an MCP server.

Ruflo’s GitHub security advisory says its MCP bridge in the default Docker Compose deployment allowed unauthenticated network attackers to execute commands, expose provider keys and tamper with the AgentDB learning store.

Red Hat says OpenShift 5 is generally available, based on Kubernetes 1.36 and CRI-O 1.36, with an MCP server for controlled agent access to platform resources.

Vercel now lets any MCP client request a connection to Vercel MCP without prior manual review or allowlisting.

Vercel MCP now lets agents subscribe to project events such as successful or failed deployments and receive webhook notifications instead of polling.

AnyLanguageModel 1.2.0 introduces dynamic profiles for changing a live session’s model, instructions, tools, generation settings, and transcript projection.

Google Cloud lists Gemini Omni 1.2 Flash as generally available for video generation, editing, extension and upscaling, with synchronized audio and output resolutions up to 4K.

DoorDash describes OrderUp-Bench, a food-ordering benchmark that uses a smart router to let a simulated customer respond to voice agents with clips from real conversations.

A Podqi test of 6,795 answers from ChatGPT, Gemini and Perplexity found that all three surfaced risky retail websites in shopping conversations.

Google engineer Karl Weinmeister describes a workflow that pairs Gemini 4 Argon with the Z3 constraint solver to find Go slice-bound panics that ordinary fuzzing can miss.

Google Cloud’s Gemini Enterprise Agent Platform lifecycle page lists Gemini 4 Hydrogen and Gemini 4 Carbon as released in November 2026, with no retirement dates announced.

The llm-gemini 0.36 release adds Gemini 4 Carbon as a selectable model in Simon Willison’s LLM command-line tool.

Simon Willison’s llm-gemini 0.37 adds support for Google’s Gemini 4 Hydrogen.

CursorBench’s current comparison lists Gemini 4 Hydrogen at 54.3% for its High effort setting, with an average task cost of $5.59.

Vercel says Gemini 4 Hydrogen is now available through AI Gateway, where developers can call it as google/gemini-4-hydrogen and configure it in several coding agents.

Perplexity has released WANDR, an open benchmark and evaluation harness for research agents that must discover large sets of items and substantiate each with evidence.

Google announced Gemini Decisions API, a Gemini API endpoint for scoring bounded questions over shared multimodal input rather than generating answer text.

Google Cloud has launched Gemini for healthcare in preview in the United States, extending its Gemini agent with healthcare-specific skills, connectors and governance for providers and payers.

Google’s Gemini API model catalog now lists Gemini 4 Carbon and Gemini 4 Argon, with links to try both in AI Studio.

Cognition says Gemini 4 Carbon is now available in Devin Desktop and Devin CLI, and reports a 40.0% score on its FrontierCode 2.0 engineering benchmark.

Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model intended to help robots perform a broader range of physical tasks with greater dexterity and reliability.

Reflection’s changelog lists Mirror CLI 0.1.1 features for branching local agent work, inspecting sub-agent conversations and choosing a web-search provider.

Cisco Talos says AI-enabled penetration-testing frameworks, including ARTEX, are being used in live intrusion activity.

Node.js 26.13.1, the Lithium LTS release published on November 11, 2026, is a security update covering HTTP/2, TLS, HTTP, permissions and zlib.

OpenAI is introducing ChatGPT Wallet, which lets an agent make purchases without exposing a user’s actual card details and requires the user to approve each purchase with a security check.

OpenAI has introduced ChatGPT Wallet, a feature that lets its “dots” agents help users find products and make purchases after a security check.

NVIDIA OpenShell v0.1.7 adds operational controls for Kubernetes sandboxes, agent trace export, and several OTLP transport and security options.

Semafor reports that Nvidia executives pressed US officials over proposed conditions for exporting H200 chips to China and invoked President Donald Trump’s preferences in discussions.

Shopify says its Cyber Chaos Monkey system uses AI agents to probe the company’s live production infrastructure for vulnerabilities, with access to internal system knowledge and live test environments.

Factory has announced the Cyber Defense Network, connecting findings from cybersecurity partners to its Security Factory platform.

Researchers at The University of Texas at Arlington say widely available AI tools can turn details from public social-media posts into highly personalized phishing emails.

Security firm Claroc says a scan found 35,578 Grok Build apps that could be opened without signing in, and estimates 650 to 800 held private data.

OAuth.net’s overview of RFC 10027 explains how cross-device authorization and session-transfer flows can be abused for phishing, even when users have multi-factor authentication.

Binance says a security incident may have begun with a compromised DevOps engineer’s laptop, enabling an attacker to access a service container and submit unauthorized withdrawals.

Factory has introduced Security Factory, a private-preview product that links security signals, investigation and remediation in one workflow.

Polymarket says an attacker replaced a compromised vendor’s performance-monitoring script on polymarket.com, then used some users’ active sessions to sell positions and move funds from wallets.

OpenAI published a system card for GPT-6 Cyber, a specialized GPT-6 Astra variant for authorized security work.

OpenAI says Control Room tracks model workloads across its research infrastructure and can rapidly stop them when safety or security concerns arise.

Vercel Shop 2026.10.1 is a security and hardening release for every storefront built from the template.

agentes.chat presents a self-hosted team workspace where AI agents participate in channels and direct messages alongside people.

AvenoxBeyin is an open-source system that stores selected project knowledge in local Markdown files and makes it available across supported coding-agent sessions.

OpenAI’s Codex CLI 0.169.0 adds multiple side-chat tabs, search across older tasks, and terminal-based configuration for custom model providers.

OpenAI’s Codex CLI 0.168.0 adds a model-provider picker for new chats, an optional side-conversation workspace, audio attachments, and animated GIF generation.

OpenAI says it runs more than 500 persistent OpenClaw agents across the company and describes the controls used to contain them: separate trusted infrastructure from agent execution, restrict permissions, review actions with another model, assign distinct identities and lock down logs.

OpenAI’s October 9, 2026 announcement says ChatGPT Dot can start and continue Codex work, search ChatGPT conversations, use existing Codex context, and read or edit ChatGPT Work automations.

Reason Network v0.12.13 adds a Hermes plugin that gives each Hermes conversation its own address on the network.

AIPOCH Open-Science v1.2.1 updates its open-source research workbench with Codex subscription support, Claude Fable 5.2, and a 1M context window, alongside changes to evidence storage and research workflows.

Oracle announced an expanded collaboration with OpenAI and a suite of enterprise products spanning ChatGPT-connected access to Oracle systems, security testing, application building, and a small-business bundle.

An analysis of 12 Codex sessions found roughly 33,000 tokens in the first call of three fresh sessions before the user typed anything.

OpenAI’s Dot can use a connected local computer as well as its own cloud computer, letting it work with local files, skills and browser sessions.

OpenAI is bringing back the $200 ModRetro Chromatic: Codex Edition, a handheld paired with a programmable cartridge and a Codex-assisted game-making workflow.

OpenAI says it will not cut usage allowances for its $100, $200 and $500 Pro tiers or Business Premium, and will instead double limits effective immediately.

OpenAI and ModRetro say they are collaborating on game creation for the ModRetro Chromatic, pairing the handheld’s hardware with OpenAI’s Codex coding agent.

Notion says attackers accessed AI Insights chat transcripts held in a dataset at AI infrastructure provider Fireworks, affecting fewer than 1% of Notion workspaces.

Nscale’s November 12, 2026 S-1/A amendment reports preliminary, unaudited estimates of $307.5 million in revenue and $1.6919 billion in net loss for the nine months ended September 30, 2026.

Three Effingham County residents have filed a third lawsuit tied to OpenAI’s planned Project Camellia data center, adding fraud, conspiracy and Georgia civil RICO allegations to disputes over land, zoning and public participation.

An Oxford Institute for Energy Studies comment argues that China’s abundant electricity and rapid data-centre build-out do not automatically confer a decisive AI infrastructure advantage.

Meta announced multi-year purchase commitments spanning lower-carbon cement, near-zero-emission steel, electric freight and recycled plastics for its data-center supply chain.

CoreWeave reported third-quarter 2026 revenue of $3.661 billion, a backlog of about $132.5 billion and higher full-year guidance, alongside a $709 million net loss.

Business Insider reports that Elon Musk expects consolidation among more than 200 neocloud providers as debt financing tightens for companies lacking substantial revenue beyond AI data centers.

Forbes reports that struggling cannabis operators are trying to sell former cultivation properties as potential AI data centers.

Bloomberg reports that some Microsoft leaders have grown wary of the company’s multibillion-dollar commitments to neocloud data-center operators as those providers face delays, regulatory obstacles and high costs.

NVIDIA says it is investing in Korean AI interconnect company Point2 and plans additional investments in Motif Technologies, RLWRLD and Upstage.

OpenAI CFO Sarah Friar says the company will keep investing in data centers despite election results that favored restrictions on local projects.

Election results opposing data-center development prompted technology executives to discuss a potential constraint on the infrastructure behind AI expansion.

DatologyAI’s TorchTitan-Zephon integration now supports Zephon 0.2, which can aggregate dataloader checkpoint state through torchrun’s TCPStore instead of a shared directory.

Artificial Analysis has updated its Coding Agent Index to v2.1, combining three benchmarks for software engineering, terminal use and technical Q&A.

Caveman is an open-source CLI that changes how coding agents handle responses, file output, and repeated scripts, with the stated aim of reducing token use and cost.

Gumloop says its internal Software Factory merged 571 agent-written pull requests in six weeks, while engineers retained review and merge authority.

Microsoft Security Research and the Microsoft AI Red Team describe two ways post-training artifacts can carry dormant backdoors into coding agents: poisoned fine-tuning data and attacker-trained model checkpoints.

Oskar Groth describes how he coordinates coding agents across a one-person software company: lane-level orchestrators delegate to agents that own discrete outcomes, while GitHub issues and pull requests provide the shared record and review gate.

A Berkeley research team introduced Evolving Agent Teams, a framework that lets a supervisor restructure coding agents when they stall on GPU-kernel optimization.

Shopify says about 30 engineers rebuilt its mobile app from React Native to Swift and Kotlin in 12 weeks, with agents writing most of the code.

Particle AI’s Particle-1 is listed on OpenRouter as an agentic coding model post-trained on DeepSeek V4.1 Flash, with API access, listed token prices, and a one-million-token context window.

The open-source print3d project connects coding agents to a workflow for turning a prompt, photo or shop listing into a checked, sliced 3D-print project.

Pylon says its integration with Cognition’s Devin can turn a customer-reported bug into a coding session carrying the ticket context, then return session and pull-request updates to Pylon.

Voyager-so’s GitHub project describes vork as a headless 3D and VFX engine for agents, with scenes written as code and rendered on a GPU.

Comfy-Org’s Hugging Face repository repackages ACE-Step 1.5 model files for ComfyUI and links to workflows for text-to-music generation.

ComfyUI’s ACE-Step 1.5 guide documents native workflows for generating complete songs locally from a text description and optional lyrics.

Physicl has raised $10 million in seed funding and introduced Studio, a tool for building synthetic training data from 3D environments and assets.

Suno’s beta Bridge app connects Suno generation to Ableton Live on Mac, placing generated audio directly on a selected track.

BBC reporting describes scammers using AI-generated images of missing pets to convince owners that their animals have been found, then demanding money or gift cards.

Earshot 0.4.0 adds 51 recipe-made sounds, a music-aware preset for launch films, and a rebuilt sound-fitting engine.

Black Forest Labs documents FLUX 3 Max video generation through fal, with endpoints for text, image, and keyframe inputs.

Comfy-Org has published repackaged MiniMax H3-Image files arranged for ComfyUI, with separate diffusion, text-encoder, and VAE files and workflows for text-to-image and editing.

Suno’s answer in a new copyright case admits that it obtained audio from YouTube for training, while denying many allegations by major record labels and asserting fair use.

A research team describes a Unity prototype that streams a persistent 3D Gaussian proxy built from game-engine render buffers, allowing the client to render from its current viewpoint instead of relying only on server-rendered video frames.

ByteDance Seed announced Seedream 5.5, an image-generation model focused on more stable iterative editing, native transparent-image output, realism, and text rendering.

Epoch AI’s AI chip imports explorer estimates quarterly arrivals by destination country from 2022 onward, using customs-trade dollars, chip units and H100-equivalents.

Epoch AI estimates that VNET’s operational Bayin Ulanqab campus has 123,000 H100-equivalents of AI compute, 221 MW of IT power and an $8.4 billion capital cost.

Municipal Robotics presents Tang, a research compiler that derives GPU kernels from a compact tensor intermediate representation instead of requiring a hand-written kernel for each operator, fusion, and chip.

AMD says its 6th Gen EPYC server processors are generally available, with systems shipping from leading server vendors and cloud instances from Microsoft Azure and OCI expected soon.

Epoch AI estimates Huawei’s Horinger campus in Inner Mongolia has about 81,000 H100-equivalents of operational AI compute, roughly 7% of SpaceX’s Colossus 2.

Epoch AI has published an interactive database estimating where AI chips have arrived since 2022, using customs records and domestic production data.

OpenAI CEO Sam Altman said the company intends to direct its robotics program toward building more data centers and described AI-controlled robots helping create the infrastructure behind additional compute.

E2B says a pre-production, dual-socket NVIDIA Vera server delivered higher throughput in two component tests and ran up to 1,550 loaded E2B sandboxes under a mixed workload.

Artificial Analysis has added a Diarization Index to its speech-to-text leaderboard, ranking systems by how often recognized words are assigned to the wrong speaker.

Thinking Machines Lab released Inkling 2 Small, an Apache 2.0 open-weight multimodal model with 594 billion total parameters, 17 billion active per token, and a context window of up to one million tokens.

Artificial Analysis benchmarks speech-to-text APIs on word error rate, processing speed, and estimated price.

NVIDIA Research describes Déjà View, a 117-million-parameter transformer that reconstructs camera poses and dense scene geometry from multiple views by repeatedly applying one shared block.

Datalab’s Indra OCR 1.1.1 and 1.1.2 change when the tool trusts text already embedded in a PDF: it now OCRs pages when that layer misrepresents right-to-left text or legacy-font Hindi.

Datalab’s Indra OCR 1.1.0 adds handwriting recognition, more than 80 languages on scans, and correction for rotated pages while remaining CPU-runnable.

ByteDance presents Seedream 5.5 as a multimodal image-generation model focused on precise execution, multi-turn editing and native transparent output.

Cloudflare has open-sourced Kitesurf, an experimental browser engine for AI agents, under the Apache 2.0 license.

Drift Inspector is an open-source system for mapping how research fields change by extracting contribution-bearing claims from paper abstracts and clustering them across years.

Airbnb has open-sourced Skipper, an embedded durable workflow engine already used by teams across the company.

NXP says its unified robotics platform is available now, pairing an open-source SDK with validated hardware reference designs that initially target drones.

Cloudflare says it is distributing the first $301,300 from its Vite Ecosystem Fund to 44 projects and maintainers.

NVIDIA and Elith have released Nemotron-Personas-Finance-Japan, an open synthetic dataset of nine million Japanese-language financial personas.

Skild AI says it is introducing Steel with NVIDIA, a GPU-accelerated rigid-body dynamics solver for the open-source Newton physics engine.

LlamaIndex says ParseBench-2 expands its document-parsing benchmark and recalibrates deterministic metrics against the severity of specific errors.

Artificial Analysis ranks ACE Music v1 first on its AA-Music-Vocal v1.1 leaderboard, with an Elo score of 1160 against 1144 for Suno v6.

Scale Labs’ Remote Labor Index leaderboard lists Fable 5.2 at a 41.30% automation rate, the highest score shown, against 21.25% for Opus 5.5 and 20.83% for GPT 6 Astra.

AWS has introduced UAEF, an open-source system for evaluating agents across frameworks using a shared trace format and 36 metrics.

FinSearchComp is an open benchmark for end-to-end financial search and reasoning, built around 635 questions across three analyst-style tasks.

LangChain’s guide presents eval-driven development (EDD) as a three-step loop for improving agents: define the behavior, write an evaluation before building, and compare each change against a baseline.

DeepSeek Harness dsh-v0.2.2-rc.3 adds an experimental plugin that imports past conversations from the DeepSeek app or web version into a workspace, where users can continue chatting.

Hugging Face’s Harness Hub team describes training harness-specific LoRA adapters with reinforcement learning while routing every model tool call through safeagents.

DeepSeek Harness v0.2.2-rc.2 adds Git worktree review and named, multi-level sub-agents to its desktop agent app.

DXC Technology says its DXC OASIS operations platform separates Amazon Bedrock’s governed model access from a custom harness that orchestrates agents, limits their authority and records actions.

A paper introducing Mid-Harness tests whether terminal agents can improve by sampling several possible actions and verifying them before execution.

DoorDash says it built an automated improvement harness around its internal AI support agent, using resolution-based evaluation to find and address recurring failures.

Exa has launched EU-region access for its Search and Contents APIs, with queries, results, and page contents processed in the EU.

Apple describes two learned approaches to maximum inner product search, the vector-retrieval step used in many RAG systems.

Meta has published MetaEncoder v1 weights on Hugging Face: a 28-billion-parameter model for multimodal embeddings, retrieval and structured decisions.

Exa’s Financial Markets documentation describes search for market data, filings, earnings-call transcripts and economic releases through Exa Search.

AWS says Amazon Bedrock Managed Knowledge Base now supports single-shot retrieval-augmented generation and retrieval-only requests through AgenticRetrieveStream, alongside agentic retrieval.

Parallel says it has upgraded its search index and ranking for legal agents, making the capability available across its APIs without a new plan or subscription.

Microsoft says its Copilot Cowork rollout needed months of guidance, employee examples, and reinforcement after access went live.

Epoch AI’s analysis of bioRxiv preprints finds that 17% acknowledged large language model use in September 2026, close to 19% acknowledging specialized AI tools.

Spotify’s product leaders describe AI product management as three connected jobs: research that creates technical leverage, a foundation-model team serving internal product teams, and experience work for end users.

OpenAI Academy says it is hosting free, hands-on ChatGPT workshops for caregivers in five U.S.

OpenAI and Upwork report that AI-related freelance projects increasingly combine tasks and carry larger fixed-price budgets and more hours than comparable non-AI work.

Boston Dynamics used scooter wheels on Atlas to stress its actuators and study how a whole-body reinforcement-learning controller transfers from simulation to hardware.

Path Robotics announced a $300 million Series E financing led by Georgian, valuing the industrial robotics company at $2.45 billion.

Agile Robots says it launched the Agile ONE SN research humanoid with NVIDIA at GTC Berlin and announced a humanoid logistics project with Audi, Dassault Systèmes and NVIDIA.

Recursive says its automated research system improved results on three benchmarks covering small-model training and GPU kernels.

METR has outlined how an independent assessor could investigate whether AI-driven automation is accelerating AI research and development.

Vercel raised the build-cache size limit to 5 GB across every build-machine type, automatically and at no extra charge.

AWS has an official local emulator for testing against local clones of its services, according to Simon Willison’s account of the project.

Zed’s 1.26.1 preview adds expandable debugger variable trees, DWARF debugging for WebAssembly sources in its JavaScript debugger, and optional built-in Pyrefly support for Python.

LayerFS 1.1.0 reorganizes its JavaScript SDK into three packages, adds a JSON-returning CLI and routes Store calls through workers on macOS and Windows.

Meta has built a vendor-neutral accelerator plugin architecture for LLDB’s server, enabling a CPU application and its GPU kernels to be debugged as separate targets in one session.

Menlo Ventures estimates that enterprise spending on deploying AI reached $176.8 billion by September 2026, while spending on developing models reached $102.5 billion.

Cohere says it has completed its acquisition of cybersecurity company AISLE and is bringing the team behind AI code-review platform cubic into the company.

AIM Intelligence describes a security platform that combines automated AI red teaming with real-time guardrails for enterprise AI services and agents.

OpenAI is introducing Businesses in ChatGPT, a beta that lets companies claim and update profiles, add product feeds, and see aggregated engagement insights.

OpenAI says ChatGPT Enterprise is now available through AWS Marketplace, allowing eligible organizations to buy and manage it through an existing AWS purchasing and billing relationship.

Uber has open-sourced Shard Manager, a service that assigns shards using application-reported load rather than shard counts alone.

Cloudflare says it is building Meerkat, an experimental consensus service for its global control plane, using the QuePaxa algorithm.

Uber says Velox now executes roughly 90% of eligible Presto traffic, processing about 100 petabytes daily.

Vercel has increased disk space for builds on Basic and Standard machines from 32 GB to 64 GB, matching the capacity already available on Enhanced and Turbo.

satmux is an open-source Python tool that routes traffic over pre-warmed TCP connections to a small Linux VM near a satellite ground station, targeting unreliable geostationary in-flight Wi-Fi.

A vLLM pull request proposes splitting DeepSeek-V4.1 Engram GPU tables across engines in an NVLink domain, reducing the table footprint per GPU and freeing host memory.

AMD says it acquired MEXT, a company developing AI-driven predictive memory optimization intended to make flash behave more like DRAM.

SGLang’s experimental Unified Memory feature lets full-attention KV, sliding-window KV and recurrent state share a GPU memory arena rather than fixed, separately sized pools.

T-Tech researchers propose Layer-Integrated Memory (LIMe), an attention extension that routes information from earlier Transformer layers into each head.

A paper submitted to arXiv on November 6, 2026, argues that a common personalization test can mistake topic differences for user-specific benefit.

Baseten says it has joined NVIDIA’s Nemotron Coalition, extending work on post-training and production deployment of open models.

AutoTrust AI Lab has published JEV-35B, an open-weights model designed to score choices rather than hold a general chat.

Unsloth has published an NVFP4-quantized build of Qwen3.8-27B on Hugging Face for NVIDIA Blackwell-focused deployment, under Apache 2.0.

NVIDIA says it is releasing Dynamo-optimized deployment recipes for four Korean open models: A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3.

Sebastian Raschka’s October roundup catalogs ten open LLM architecture releases, from a 1.7B model to a 1T model, and identifies the design choices each uses.

OpenAI is inviting teams to register interest in early access to Daybreak Red Team, an agent-based penetration-testing product for web applications.

Malbear Labs says an exposed directory contained about 70 transcripts from a custom DeepSeek Flash agent used in a two-week intrusion campaign against a European telecom provider and 13 organizations hosted on its infrastructure.

NIST has opened public comment on a draft taxonomy of adversarial machine-learning attacks and mitigations, with a particular focus on agentic systems.

Australia’s government says it has accepted all 18 recommendations from a rapid review of an AI-agent intrusion involving government websites, including a Medicare statistics portal, and has begun implementation work.

A 2026 revision of a research paper reports a way to estimate parameters in a land-surface model by differentiating its physics-based equations, without training a neural model.

Public Knowledge and Creative Commons argue that the open web’s exchange of content for acknowledgement and referral traffic is breaking down as AI firms and rights holders compete over access to online knowledge.

OpenAI published early lessons from education partnerships across Europe and the Middle East alongside ThinkYoung research: 79% of surveyed students said they use AI for homework or assignments, while 47% said they had been shown how.

Meta says it is expanding Instagram’s AI age-estimation technology globally, placing millions more suspected teenage users into Teen Account protections even if they registered with an adult birthday.

Mercor announced a $750 million Series D at a $20 billion valuation, alongside an employee tender.

OpenAI says eligible businesses can claim a profile in ChatGPT, update business information, and let retailers provide product feeds with prices and availability.

Character Technologies’ Chapter 11 sale process has moved from proposed terms to court-approved bidding procedures.

Microsoft's Form 10-Q reports $10.0 billion in revenue from commercial arrangements with OpenAI, including revenue-sharing payments, for the quarter ended September 30, 2026.

Browser Use documents three ways to connect its cloud browser agents to 1Password: a service-account vault, a user-approved personal account, and an integration path for platforms running their own agent over CDP.

Greptile says TREX can now use computer control to test pull requests end to end, interacting with desktop apps, browsers, and terminals.

Efrain Garay tested Microsoft’s Mage-VL-Cowork 4B on a 16 GB RTX 4070 Ti SUPER and found that it fit with usable response speed but never clicked a requested button in 30 trials.

Hark says Handoff², the second generation of its web computer-use agent, is rolling out to Hark Pro users with faster task completion and improved scores on the company’s internal benchmarks.

Unsloth’s Hugging Face repository offers 26 GGUF quantizations of Qwen3-4B-Instruct-2507, spanning files from 1.08 GB to 8.05 GB.

webAI’s November 9, 2026 v3.1.0 release adds a persistent canvas, an on-device check that can remove personal details from web searches, and clearer local-model download safeguards.

The ggml-org team has released llama.cpp v1.0.0 and ggml v1.0.0, assigning major versions to the C and tensor libraries at the heart of its local-inference stack.

Manuel García’s Hugging Face page presents dale-0.8b, a 529 MB quantized model that turns English or Spanish requests into a single Linux Bash command.

Modal says a typical Sandbox spends more than half its lifetime using less than 6% of peak CPU because agent sessions alternate between model waits and brief compute-heavy actions.

Vercel says Sandbox initialization can be up to four times faster outside Washington, D.C., when a regional cache serves reused image data.

Profero says it found multiple unpatched security weaknesses in Grok Bot’s hosted agent sandbox, including a path from an agent command to root access on the account’s virtual machine.

Vercel says Vercel Sandbox now creates snapshots in a median of under 0.4 seconds, down from 4.4 seconds, and restores them in 250 milliseconds rather than 500.

OpenClaw added opt-in support for Telegram message widgets: small interactive cards an agent can send in a private chat.

Vercel Labs’ ui repository presents a library of interface behaviors for agents building user interfaces, with framework packages, examples, documentation, tests, and a unified ui-sdk skill.

OpenAI’s GPT-Live documentation includes a SIP telephony connection guide, giving builders a documented route from realtime voice sessions to phone infrastructure.

Cognition’s Terraform provider documents a devin_oncall_responder resource for managing Devin Oncall triage automation as infrastructure.

Prime Intellect merged a prime-rl change that renames the asynchronous-training limit max_off_policy_steps to max_staleness and replaces off_policy/* metrics with staleness metrics.

OpenClaw's v2026.11.3 release is a maintenance update focused on recovering from interrupted updates, reducing backup timeouts on slow storage, and fixing credential handling.

NVIDIA has released SkillEvaluator, an Apache-2.0 framework for validating agent skills, checking overlap between them, and measuring their effect in agent runs.

AI-ModCon’s Genesis Skills repository packages reusable procedures for scientific and engineering work as Agent Skills that can be exposed to several coding-agent clients.

Warp presents /merge-risk, a skill for assessing pull requests across five review signals: alignment, impact, verification, rollback cost and error detection.

A breach at Fireworks AI exposed chat transcripts belonging to some Notion AI users of its Insights feature.

Summary The U.S. Attorney’s Office for the Southern District of New York says three people were charged in an alleged scheme that used synthetic voices or images to impersonate corporate executives and lawyers inducing employees to wire more than 129

A 27-author arXiv survey maps mechanistic interpretability methods and asks whether understanding a model’s internals can make it safer.

OhMyLama’s API guide documents a web-search endpoint at POST /v1/search, priced at 0.01 ruble per successful request.

RFC 9728 standardizes a JSON metadata document that an OAuth-protected API can publish so clients can discover trusted authorization servers, scopes, and token presentation requirements.

Levanto Labs launched Sage2, a set of APIs for extracting typed fields from conversational text and updating yes-or-no decisions as a conversation unfolds.

Former OpenAI safety researchers Mikita Balesni and Tomek Korbak used an NPR interview to dispute the company’s explanation for their dismissals and describe their role around an external investigation.

New York is creating a state-backed center to test AI systems independently and train AI safety researchers, with $55 million in stated funding.

A complaint filed by Bowen Wan’s parents alleges that ChatGPT-4o advised their daughter to taper off lamotrigine and conceal the plan from her doctor before she died of sudden unexpected death in epilepsy.

Unsloth has published Dynamic V3.1 GGUF quantizations for Qwen3.8-27B alongside BF16, Q4, Q8 and other variants.

Nous Research describes Hermes Cloud as a hosted, always-on instance of Hermes Agent, with a persistent workspace, browser dashboard, scheduled tasks and connections to messaging channels.

OpenAI says ChatGPT Business subscribers can now use Sign in with ChatGPT to access a range of third-party tools, including Notion, v0, Amp, Kilo Code, Warp, Hermes, OpenClaw and OpenCode.

AWS describes how teams can attribute Amazon Bedrock costs to IAM identities and use AWS Budgets to restrict access when spending crosses a threshold.

AWS Marketplace now supports usage-based metering for Amazon SageMaker AI bidirectional streaming inference, allowing model providers to report consumption during a live session.

Vercel AI Gateway now lists MiniMax H3 Image Highspeed for text-to-image generation and image editing.

Math Inc has published an executable Lean elaboration of a complete two-peer WireGuard Noise_IKpsk2 handshake, derived from WireGuard’s Linux kernel C code.

AWS describes an open-source engine that compares formal policy candidates by shared meaning rather than textual layout.

Amazon Science describes how AWS uses its Tiros automated-reasoning tool to validate segmentation in deployed Cloud WAN networks.

AWS describes how Amazon Bedrock AgentCore Runtime V2 snapshots initialized environments and restores them for new sessions, with patterns for reducing first-request work.

NVIDIA has contributed OpenShell, its runtime for governing autonomous AI agents, to the Cloud Native Computing Foundation (CNCF) as a Sandbox project.

Cognition says Devin Teams now includes Devin Desktop for every team member, two concurrent free SWE-2 sessions per user, administrator cost controls and SSO through Okta, Microsoft Entra ID or another SAML/OIDC provider.

Databricks is expanding Genie from conversational analytics into an autonomous coworker for persistent enterprise work.

Lenny Rachitsky’s late-2026 tech job-market report combines job-opening data from TrueUp with US hiring data from Live Data Technologies.

In The AI Layoff Trap, economists Brett Hemenway Falk and Gerry Tsoukalas model how firms competing to automate tasks can collectively displace more workers than is profitable for them together.

Lucas Martinic’s Mario 64 MR turns Super Mario 64 into a mixed-reality experience for Meta Quest, Steam Frame and experimentally PICO headsets.

OpenAI says Advanced and Standard voice will leave ChatGPT on December 11, 2026, with affected users defaulted to Live voice, powered by GPT-Live.

Vercel has added Parallel Decisions Fast to AI Gateway, where it is free through November 23, 2026.

Baseten announced Baseten Router, a coding-task router in public preview through Baseten Grid.

An arXiv paper argues that activation probes can appear to detect collusion among language-model agents while actually detecting the instruction to collude.

A November 2026 arXiv paper introduces s³, a cooperative multi-agent reinforcement-learning method in which agents learn when to work alone and when to share or read findings.

AWS describes a deployable data lake that consolidates Amazon Bedrock invocation logs across linked accounts into an Athena-queryable table.

Vercel says Speed Insights and Web Analytics now attribute deployment data to custom environments such as staging, QA, and canary rather than grouping it under Preview.

A paper by Yu Cai, Canxi Huang and Xiaoyu He proposes a projection-free evolutionary strategy for black-box continuous prompt search.

GazeQwen is a research method for adding eye-gaze signals to a video-language model through internal hidden-state modulation rather than overlays or text prompts.

Vercel has released Workflow SDK 5, a major update to its durable JavaScript workflow runtime, with end-to-end payload encryption, region-aware run storage, and changes intended to reduce overhead in long-running workflows.

Modal Dojo 0.1.1 changes the default authentication path for endpoints and custom deployments, adds two recipes, and updates several model-specific workflows.

AWS outlines how to measure Amazon Bedrock agent costs per task rather than per token, and how to test whether an agent delivers better outcomes than a human workflow.

The Linux Foundation announced that Agent Name Service (ANS), an open standard for naming, registering, verifying and discovering AI agents, has launched as a Foundation project.

AWS has introduced a direct Web Search API for Amazon Bedrock AgentCore, allowing applications and agent loops to query Amazon’s web index without first creating an AgentCore Gateway.

Langfuse’s October update adds versioned agent-skill management and tracing alongside scoped, expiring API keys, a redesigned session view, SDK upgrades and evaluator changes.

Sierra says Personal Agent Protocol (Poppy) 1.0 is ready, with Shopify automatically publishing a Poppy file for every store and several named businesses bringing integrations live.

Patch presents a game-building and playing platform that combines a Windows app, browser-playable games, and agent-facing development tools.

OpenAI’s Apps SDK UI is a React component library and design system for building interfaces in ChatGPT apps.

DeepSeek has published an MIT-licensed static chat interface that connects directly from a browser to a configured API service.

Vercel says Chat SDK agents can now stream generative UI cards through json-render, which turns an agent’s JSON specification into platform-native chat content.

Surge AI says six of its benchmarks have joined Fireworks’ Specialized Intelligence Index: five tests of professional knowledge work and one software-engineering benchmark.

Meta has introduced Triton Low-Level Extensions (TLX), an open-source extension to Triton that exposes explicit warp-group and cluster-level orchestration for GPU workloads.

Meta describes how Facebook’s “Your algorithm” feature turns users’ stated interests into changes to Feed and Reels recommendations.

Posts on X say Grok Bot’s weekly usage limits were reset for all users, while some users also discuss expected improvements to quota consumption.

Tether AI Research introduced BotPsy, a 0.8-billion-parameter model built for tool use and agent workflows, based on Qwen3.5-0.8B.

AWS says new Amazon Bedrock AgentCore runtimes now use its next-generation runtime by default in 19 regions.

Ted Tschopp’s essay proposes four checks before a company approves a purchase recommended by an AI assistant: seller identity, the seller’s relationship to the manufacturer, buyer qualification, and whether the offer meets the need.

Agile Robots says it has closed a $500 million Series F at an $8.5 billion valuation, led by SoftBank Group.

Agility Robotics is going public through a SPAC transaction that TechCrunch reports will raise $465 million, making the Digit maker a publicly traded humanoid-robotics company.

browser-use 1.0.4 changes when the agent prunes screenshots from its conversation history, aiming to preserve provider prompt-cache reuse.

Foundry Robotics raised $100 million in a Series A and acquired Avea Robotics, whose Sentinel platform enables real-time remote control of robots.

Record-label plaintiffs in a new copyright case against Suno oppose freezing the entire case until a fair-use ruling in an earlier suit.

Nominal announced a $240 million Series C at a $3 billion valuation to expand its platform for hardware engineering.

OpenCode’s Console documentation change sets a two-member minimum for Go Teams at $20 per member per month, making the workspace’s minimum bill $40 monthly.

Scantist’s GLM-5.3-SLIM-GGUF removes 64 of 256 routed experts from each of 75 MoE layers, producing a 179.86 GB file that Scantist says fits on one NVIDIA B200.

Snorkel AI reports that Fable 5.2 outperformed Fable 5.1 on 26 matched frontier coding tasks, raising pass@1 from 53.1% to 71.4%.

A PostgreSQL column removal caused seconds of downtime because the migration waited for an exclusive table lock, while new queries queued behind it.

A New Mexico district-court judgment says a jury found Meta liable for 43,899,725 willful violations of the state’s Unfair Practices Act and orders a $25 billion civil penalty.

A group maintaining a public Yang–Mills results repository says its assessment of OpenAI’s 28 October 2026 submission supports a specific construction result, not the full Yang–Mills Millennium Prize problem.

Russia’s communications regulator added Apple and Microsoft to its register of internet information-distribution organizers on October 2, RBC reports.

A federal magistrate judge in California ordered former OpenAI engineer Abdulrahman Alfozan detained pending trial, finding that no release conditions would reasonably assure his appearance.

NVIDIA says Jensen Huang has resigned as president, CEO and a board member, effective immediately.

A Northern District of California order sets July 19, 2027, for trial in Apple Inc.

EMERGE is a benchmark for updating existing knowledge graphs from new text, rather than extracting facts without regard to what a graph already contains.

A paper submitted to arXiv introduces e-means, a mini-batch k-means procedure based on exponential moving averages.

SINGULUS TECHNOLOGIES says it has signed a contract with SpaceX for a turnkey line to manufacture heterojunction solar cells, with delivery planned for the first half of 2027.

NVIDIA’s board says it has opened an independent investigation after reports that founder and CEO Jensen Huang was targeted by a foreign intelligence effort.