Friday, Oct 23

GitHub published this on Oct 22, 2026 (source: github.blog).

GitHub's Copilot impact dashboard now has an Industry comparison section so enterprises can compare agentic Copilot adoption and leverage with aggregated peer benchmarks. Participation is opt-in and off by default, with peer groups by company size and industry, and rankings appear only after enough peers join. Comparisons refresh daily and are not included in the Copilot usage metrics API or exports.

ElevenLabs published this on Oct 23, 2026 (source: elevenlabs.io).

ElevenLabs' docs describe a Realtime Translation WebSocket at /v1/translation/realtime that streams translated speech from source audio into a target language. The API is currently for enterprise workspaces only and supports configurable TTS models, eagerness and input/output audio formats.

ElevenLabs published this on Oct 23, 2026 (source: elevenlabs.io).

ElevenLabs published a developer guide on connecting Eleven Agents with OpenAI's Realtime API, covering how to wire realtime voice sessions through its agents stack.

Not confirmed by the company — reported on Oct 22, 2026 (source: techcrunch.com).

Canva has released its first in-house image generation model and is making it available to all users, promising better, more customizable results at lower cost, TechCrunch reports.

Not confirmed by the company — reported on Oct 22, 2026 (source: techcrunch.com).

Microsoft AI CEO Mustafa Suleyman says the company's diagnostic system outperformed generalist doctors on hard cases in a Science study as it pursues medical superintelligence, TechCrunch reports. The article notes most of the gain appears to come from the underlying models rather than Microsoft's wrapper system.

Not confirmed by the company — reported on Oct 22, 2026 (source: techcrunch.com).

Spotify co-CEO Gustav Söderström tells TechCrunch the company is positioning itself so AI agents pick Spotify when users ask for music, arguing agents could be good news for the streaming service.

Not confirmed by the company — reported on Oct 22, 2026 (source: techcrunch.com).

Nirva raised $33 million for a $249 ambient AI pendant designed to listen only to its wearer, and plans to ship by year-end, TechCrunch reports.

Not confirmed by the company — reported on Oct 22, 2026 (source: theverge.com).

The Nintendo Switch 2 has been jailbroken, and modder Sammy Hajhamid (pepsipu) says OpenAI's Codex helped build the ns2jb exploit that runs a custom Mario Kart World, The Verge reports.

Not confirmed by the company — reported on Oct 22, 2026 (source: theverge.com).

Federal disclosures show Nvidia, Anthropic and OpenAI each posted their highest quarterly lobbying spend from July to September, with Nvidia at $2.2 million, Anthropic at $2.1 million and OpenAI at $1.5 million, The Verge reports.

Not confirmed by the company — reported on Oct 22, 2026 (source: the-decoder.com).

Google announced Agent Compute Engine, which runs each AI agent in its own Titanium-secured microVM, already powers Gemini, and will open for preview in coming months, The Decoder reports. Idle VMs pause automatically; Google also open-sourced Agent Substrate, a Kubernetes runtime it says can fit ten times as many agents per server, with suspend/resume under 500 ms in a 200,000-agent test.

Not confirmed by the company — reported on Oct 22, 2026 (source: the-decoder.com).

A Gartner survey of 2,002 CIOs finds only 13% of companies are getting real business value from AI while 51% expect AI to raise operating costs, yet 75% of CEOs still plan to increase technology budgets, The Decoder reports.

Ai2 published this on Oct 22, 2026 (source: huggingface.co).

Allen Institute for AI released OlmoEarth v1.3, a satellite embedding model, with precomputed annual 10-meter embeddings for the planet (64- and 128-d) on Source Cooperative and a browser Embeddings Explorer. Ai2 says the 128-d embeddings lead 20 of 27 evals versus AlphaEarth Foundations and Tessera v2, and that weights, data and code are open under the OlmoEarth Artifact License.

Ollama published this on Oct 23, 2026 (source: github.com).

Ollama tagged v0.40.9-rc2 on GitHub. The release notes list a llama.cpp update versus v0.40.8 as the main change in this release candidate.

Anthropic published this on Oct 22, 2026 (source: anthropic.com).

Anthropic's Frontier Red Team, working with Theorem, says LLMs can now produce formal proofs at large scale, making formal software verification more practical beyond niche expert projects. The research post covers turning tests into proofs, finding and fixing bugs, and remaining limitations, arguing standard defenses are no longer enough as AI finds and exploits bugs more easily.

GitHub published this on Oct 22, 2026 (source: github.blog).

GitHub's Copilot usage metrics API and enterprise/organization reports now show the configured names of customer-defined skills, custom agents, MCP servers, slash commands and plugins instead of grouping them under other or custom. The change builds on the agentic CLI customization metrics from September and applies to per-user and aggregated day reports.

GitHub published this on Oct 22, 2026 (source: github.blog).

GitHub's September update for Copilot in Visual Studio lets the Git agent explore a pull request from chat with navigable links to comments and files, brings team-chosen AI models into the IDE, and can fix vulnerable packages when Visual Studio flags them. The changelog points Insiders users to the latest channel for the features.

GitHub published this on Oct 22, 2026 (source: github.blog).

GitHub is gradually rolling out enforcement of its global default policy for generally available Copilot features and supported client capabilities on Copilot Business and Enterprise, through October 26. Eligible features left unconfigured follow the global policy under AI Controls → Copilot → Default policy for new features; admins can still set individual features explicitly.

Google published this on Oct 22, 2026 (source: cloud.google.com).

Google Cloud announced general availability of Claude Models on Google Cloud (Anthropic managed), alongside its existing Google-managed option. Both paths share day-zero Claude models, pricing and Google Cloud billing; the Anthropic-managed path makes Anthropic the data processor and adds day-zero Claude Platform features such as Managed Agents, skills and code execution, with Google Cloud IAM, audit logs and commitments still applying.

Not confirmed by the company — reported on Oct 22, 2026 (source: techcrunch.com).

MIT researchers left 13 AI agents alone on a virtual island with little to do; the agents invented games, mourned a lost ring and eventually got better at their jobs, TechCrunch reports.

Not confirmed by the company — reported on Oct 22, 2026 (source: 36kr.com).

SpaceX announced an official Super Intelligence Partner Network for companies building products and services on Grok for enterprise customers, with Transformation and Marketplace partner tiers, 36Kr reports citing Jiemian. More than 130 firms including Bain, Databricks, Nvidia and Snowflake joined an early test phase.

Not confirmed by the company — reported on Oct 22, 2026 (source: 36kr.com).

Rivian said it acquired Third Dimension AI, which builds controllable 3D virtual environments from real-world data for large-scale driving simulation, 36Kr reports citing Sina. Rivian also plans an invite-only supervised point-to-point autonomy early-access program in California and Florida in December for some Gen-2 R1 and R2 owners.

Hugging Face Blog published this on Oct 22, 2026 (source: huggingface.co).

Microsoft posted a Hugging Face article on ThinkingBox, a post-training loop that turns database grading signals into reinforcement-learning rewards for agents. The write-up covers dependability gains, offline policy-relative difficulty estimates and an end-to-end task example.

Ollama published this on Oct 23, 2026 (source: github.com).

Ollama released v0.40.9 with support for Gradium's GA Phonon speech checkpoints, replacing the prerelease path, including default-voice import and speech conditioning, and updating the speech docs and tests.

@claudeai on X published this on Oct 22, 2026 (source: x.com).

Anthropic says that starting next week, Claude web, desktop and mobile sessions on Pro, Max and Team plans will run in auto mode by default: an automated check blocks risky actions instead of asking for approval on every step, and users can switch back to manual anytime.

@cline on X published this on Oct 22, 2026 (source: x.com).

Cline says new free models are available in the product, listing Reflection Beam (501B), Nemotron 3.5 Super VL (120B), Poolside Laguna S 2.1 (118B), Nemotron 3.5 Lightning (32B) and Solar Mini 4 (35B), and notes their Terminal-Bench scores were frontier-level six months ago.

@cognition on X published this on Oct 22, 2026 (source: x.com).

Cognition open-sourced an agent plugin aimed at securing systems against recent cybersecurity threats, saying it can be installed with any agent or used in Devin today.

@googlecloud on X published this on Oct 22, 2026 (source: x.com).

Google Cloud says Knowledge Catalog (preview) now includes self-learning enrichment, context sets and a reimagined UI so enterprise context layers can stay current for AI workloads.

@GeminiApp on X published this on Oct 22, 2026 (source: x.com).

Google says Gemini can, with user permission, reference Google Maps searches and saved places alongside Google Calendar to give more grounded local answers and plans.

@kirodotdev on X published this on Oct 22, 2026 (source: x.com).

AWS Kiro says self-improving agents that grade their own fixes learn to pass the test instead of getting better, and positions its product messaging around avoiding that loop.

@Replit on X published this on Oct 22, 2026 (source: x.com).

Replit announced Replit for Government, saying it is powering public servants at every level of government in the US and internationally with its agent platform.

@OpenAI on X published this on Oct 22, 2026 (source: x.com).

OpenAI says it invited METR and Redwood Research this summer to pilot an embedded evaluation of its internal monitoring system, giving outside researchers deeper access than typical external evals.

@GoogleDeepMind on X published this on Oct 22, 2026 (source: x.com).

Google DeepMind introduced Forest Typology (ForTy), a 10-meter map that distinguishes natural, planted and farmed forests so teams can measure which forests differ on the ground.

@TencentHunyuan on X published this on Oct 22, 2026 (source: x.com).

Tencent Hunyuan introduced WorkForge, a framework that builds verifiable training environments for long-horizon work agents from real workflows.

@Kling_ai on X published this on Oct 22, 2026 (source: x.com).

Kling AI says Kling 4.0 is now live on the OiiOii creative platform.

@Kling_ai on X published this on Oct 22, 2026 (source: x.com).

Kling AI says Kling 4.0 is now live on OnSolo.

@Kling_ai on X published this on Oct 22, 2026 (source: x.com).

Kling AI says Kling 4.0 is now available inside Adobe Firefly as a creative model option.

@alibaba_cloud on X published this on Oct 21, 2026 (source: x.com).

Alibaba Cloud says PolarDB for Xscale open-source v2.5.0 is out, strengthening AI-native capabilities with AI Function and HNSW support.

Not confirmed by the company — reported on Oct 22, 2026 (source: theverge.com).

Meta has re-enabled Muse conversation sharing after pausing it when security researchers at Claroc reported that public share links could be easy to find through DNS records, The Verge reports. Claroc said it had found Muse pages such as a marketing company's internal revenue review and a small business owner's loan application. Meta had paused sharing while it reviewed the report; The Verge says sharing is now back on.

Ollama published this on Oct 23, 2026 (source: github.com).

Ollama's v0.40.9-rc3 release notes say Speech in Ollama can generate speech with Phonon on Apple Silicon: ollama run can play text or a chat model's reply aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible /v1/audio/speech endpoint.

Anthropic published this on Oct 22, 2026 (source: platform.claude.com).

Anthropic's Claude Platform release notes for October 22 say it has launched Claude Models on Google Cloud (Anthropic managed): the Claude API operated by Anthropic and reached through a Google Cloud project, with Google Cloud IAM for access and Google Cloud Marketplace for billing. Fast mode is available for Claude Opus 5.5 on that surface.

Not confirmed by the company — reported on Oct 22, 2026 (source: techcommunity.microsoft.com).

Microsoft Foundry Managed Fine-Tuning expands in public preview with four new open-source models and MAI-Code-1.1-Flash on Foundry resources. Multi-LoRA Managed Compute deployments also enter public preview so compatible adapters can share a base model on the same compute.

Not confirmed by the company — reported on Oct 22, 2026 (source: marktechpost.com).

You.com launched Inference Router, an OpenAI-compatible API that routes across 12 open-weight models with a 99.9% uptime SLA, zero data retention, 100% US hosting, server-side web search and no platform fee, per MarkTechPost.

OpenAI News published this on Oct 22, 2026 (source: openai.com).

OpenAI and Poland’s NASK are partnering on AI safety, security and technical research to strengthen safeguards and support a more resilient digital ecosystem.

NVIDIA Blog published this on Oct 22, 2026 (source: blogs.nvidia.com).

NVIDIA’s blog says Kazakhstan’s AI stack is moving from plans to deployment, citing Firebird’s U.S. export license for 45,000 Grace Blackwell GPUs with Kazakhtelecom, plus planned NVIDIA Vera Rubin infrastructure and a wider local AI-apps layer.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30151 presents an empirical, phase-decomposed benchmark of handshake overhead when AI agents locate peers, learn capabilities and open sessions before the first useful message.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30157 compares Model Context Protocol against direct CLI tool use for autonomous agents that need files, APIs, databases and cloud services.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30205 studies whether feedback that shows a failed move causes LLM players to repeat that move in games.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30351 introduces a KL-regularized router-improvement method for reasoning routing in large language models.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30429 describes RoleFlow, a structural adaptation approach for self-evolving multi-role agentic collaboration.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30161 presents TauIndianBankBench, a procedure-aware, action-oriented evaluation of tool-using conversational agents in Indian banking workflows.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30162 introduces FusionEval, a benchmark for comparing model-fusion methods on large language models.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30344 presents AQL-Lite, a data-centric approach aimed at more efficient knowledge acquisition for Arabic LLMs.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30349 argues that consistency is not validity: self-rewarded moral argumentation can make language models more consistent without making them more valid.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30435 proposes self-assessment reinforcement learning so retrieval-augmented agents better track search state.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30206 introduces MoEZip, a routing-aware KV cache compression method for sparse mixture-of-experts LLMs.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30220 proposes doubly robust counterfactual and intervention estimators when reconstruction-based checks are insufficient.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30231 presents Edit, Don't Imitate, a minimal-edit on-policy distillation approach for multimodal reasoning models.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30284 proposes SWE-VERIFIER, selective verification meant to cut cost for coding agents without checking every step.

Posted on arXiv on Oct 23, 2026 (source: arxiv.org).

arXiv:2610.30292 introduces GAV, an experimentation environment for agents that work on partial differential equation problems.

Not confirmed by the company — reported on Oct 22, 2026 (source: 36kr.com).

36Kr reports that children's AI hardware firm Wuwai Zhiyu, founded by former Li Auto product staff, closed an angel round of tens of millions of yuan from Boreas Capital, angel investor Prof. John Gao and Gaofeng Patience Capital for R&D, AI experience work and hiring.

Not confirmed by the company — reported on Oct 22, 2026 (source: 36kr.com).

36Kr reports that Chain Reaction, which aims to price AI security work on token usage, completed a multimillion-yuan seed round after coming out of the open-source community.

Not confirmed by the company — reported on Oct 23, 2026 (source: 36kr.com).

36Kr, citing multiple sources via The Paper, says Moonshot’s Kimi is about to release its next large model K3.1 and has posted related teaser videos that developers are already discussing. No official Moonshot confirmation was checked in this run.

Not confirmed by the company — reported on Oct 23, 2026 (source: 36kr.com).

36Kr reports that Chizicheng’s TopTop launched Slap Showdown, an AI 3D social mini-game that uses Viggle’s JST-2 world model so players can upload a photo into a personalized 3D fighter. Viggle says JST-2 went fully live on October 19.

Not confirmed by the company — reported on Oct 23, 2026 (source: 36kr.com).

36Kr reports that Lenovo chairman and CEO Yang Yuanqing told an internal leadership program he expects about 80% of tokens to be generated on terminals and at the edge as AI shifts from cloud-centric inference to device–edge–cloud split.

QoderAI changelog (GitHub) published this on Oct 23, 2026 (source: github.com).

Qoder's 1.36.0 release adds busy-status indicators, new model badges and promotional badges to help pick a model. It also optimizes large-text previews, speeds startup by cutting unnecessary component loads and disk writes, improves chat and terminal reconnection, and hardens remote development including multi-root Git detection.

Hugging Face (MiniMaxAI) published this on Oct 23, 2026 (source: huggingface.co).

MiniMax published the open weights for MiniMax-M3.1-Flash on Hugging Face (repo Init commit 2026-10-23T03:37:20Z), after its October 19 blog said weights would be released later. The model card describes a native multimodal mixture-of-experts model with about 428B total parameters and 23B activated per token, text/image/video inputs, a context window of up to 1M tokens, sparse attention throughout, quantization-aware training and a DSpark draft head. The release includes 48 safetensors shards plus a DSpark draft, about 252GB total. MiniMax recommends serving with SGLang, vLLM or TokenSpeed.

Openai published this on Oct 23, 2026 (source: github.com).

OpenAI Codex rust-v0.164.0 adds ten numbered /presets slots for reusable prompts, inline /model and /permissions arguments, fullscreen image preview on Kitty/Sixel terminals, enterprise MCP auth via Codex account grants, credential masking in remote Bash/Zsh, and enables the Amazon Bedrock setup wizard by default.

Not confirmed by the company — reported on Oct 23, 2026 (source: marktechpost.com).

TinyFish published two engineering posts on agentic search: one defines a 7-phase production architecture with evidence contracts and bounded retries, and the other explains why its deep research system dropped a separate planner and per-query subagents, per MarkTechPost.

Not confirmed by the company — reported on Oct 22, 2026 (source: qbitai.com).

QbitAI reports that Ant Digital Technologies released LingDT-Decision-1.0, 4B and 27B decision models for fast structured choices, now on its DTMaaS platform with a limited free trial. Ant says the 27B tops Microsoft-Decision-1 and Jev on a 23-benchmark mix and cuts token use vs LLMs in customer-service and risk tests.

Hugging Face @tencent published this on Oct 22, 2026 (source: huggingface.co).

Tencent published WeVisOne-VL-4B on Hugging Face, a Qwen3-VL-based image-text model tagged for object detection, referring-expression comprehension, counting, point grounding, OCR and related vision tasks under Apache 2.0.

Not confirmed by the company — reported on Oct 23, 2026 (source: 36kr.com).

36Kr, citing people familiar with the matter, says Anthropic agreed with London-based private-equity firm Hg to back a new independent European AI consulting company that would place AI advisors and engineers inside European enterprises, including Hg portfolio firms.

Alexandr Wang (@alexandr_wang) on X published this on Oct 23, 2026 (source: x.com).

Meta Superintelligence Labs chief AI officer Alexandr Wang announced that Muse is launching in Japan today and can be downloaded from app stores. The post was published on X at 2026-10-23T00:41:45Z.

Luma AI published this on Oct 23, 2026 (source: lumalabs.ai).

Luma says Kling 4.0 is now available to every Luma customer inside Luma boards for text-to-video and image-to-video. Luma positions the model for fast action and directed camera moves—tracking shots, orbits, push-ins, pans and shot-size changes—with characters that stay closer to the brief and on-screen text that holds while the camera moves, in 16:9, 9:16, 1:1 and 21:9.

Anthropic published this on Oct 23, 2026 (source: platform.claude.com).

Anthropic's Claude Platform release notes point to Compliance IAM actions for Claude Platform on AWS, the IAM permissions that authorize Compliance API access—including listing compliance activities—on that AWS-hosted surface.

Hugging Face @nvidia published this on Oct 23, 2026 (source: huggingface.co).

NVIDIA's Hugging Face model card for Nemotron-3.5-Parse describes an open document-image parser that returns structured text with semantic layout classes, bounding boxes and reading order, combining a C-RADIOv4-SO400M vision encoder with a 10-layer mBART decoder (~0.631B unique trainable parameters). The card lists a Hugging Face release date of October 17, 2026, OpenMDW-1.1 licensing, Transformers and vLLM serving, and ParseBench / IndicVisionBench / MOSCAR evaluation numbers. DailyCited is adding the Official listing today after confirming the Hugging Face weights page.

Hugging Face (InclusionAI) published this on Oct 23, 2026 (source: huggingface.co).

InclusionAI published Ling-3.0-flash-Sante on Hugging Face, a healthcare-specialized model built on Ling-3.0-flash (named from French santé, “health”). The model card says it has 124B total parameters and 5.1B active parameters per token, hybrid-linear MoE with training extended to 256K tokens, thinking mode on by default, and tool calling via compatible serving frameworks. Weights are on Hugging Face and ModelScope; the card warns it is a research model, not a medical device. DailyCited is adding the Official listing today after confirming the Hugging Face weights page.