Wednesday, Oct 14

QoderAI published this on Oct 13, 2026 (source: github.com).

Qoder 1.33.0 adds fast models for quicker responses in conversations. The release adds a notice that the Quest window will be retired soon and lets users expand or collapse the Q&A panel. It also improves performance in long conversations and large edits, makes SSH remote setup more reliable, and runs automatic incremental indexing less often to save resources.

Cline published this on Oct 13, 2026 (source: github.com).

Cline Desktop v0.0.46 adds an outline along the left edge of a conversation that lists each prompt, with hover previews and click-to-jump. Plugins that fail to load now show a status line naming the plugin and the error instead of silently losing their tools. Scheduled routines now start queued runs, mark failed model calls as failed, and run only the latest missed occurrence after downtime. With no model chosen, the Cline provider now uses Claude Sonnet 5.5, and the release adds the AI/ML API, LowRouter and Prime Inference providers.

Not confirmed by the company — reported on Oct 13, 2026 (source: theverge.com).

Nvidia published new Hogwarts Legacy comparison screenshots in which DLSS 5's changes to faces and environments are far subtler than in its March demo. DLSS 5 arrives in Hogwarts Legacy on November 5 and in Where Winds Meet late next month. Nvidia says an October update shrinks the DLSS 5 model by 24% to about 118 million parameters, cuts 4K VRAM use by 20% and makes it twice as fast on RTX 50-series GPUs. The updated model debuts in NBA 2K27 on October 16 and will ship by default in future DLSS 5 games.

AWS published this on Oct 13, 2026 (source: aws.amazon.com).

Orca Security and AWS describe Discovery AI, a natural language query agent on Amazon Bedrock that converts plain-English questions into queries for Orca's Sonar query system. Sonar covers more than 900 cloud resource types and 18 operators, which made hand-written queries hard for analysts. The companies say the multi-stage design cut query costs by 90%, input tokens by 93% and latency by 75% against the initial baseline.

AWS published this on Oct 13, 2026 (source: aws.amazon.com).

ShipStation Global's engineering team built Deckhand, an IT help desk agent for Slack, on Amazon Bedrock AgentCore. It answers routine IT and company questions from a curated knowledge base, reads error-message screenshots, and creates and updates Jira Service Management tickets. The company says that in its first three months Deckhand resolved 97.5% of channel queries with a 96% positive feedback rating from employees.

AWS published this on Oct 13, 2026 (source: aws.amazon.com).

SpaceXAI's Grok 4.7 is now available on Amazon Bedrock. AWS's launch post builds a tool-calling agent on the model and tests it on three search-heavy tasks at every reasoning effort level, six runs each. AWS reports 67 of 72 runs correct, with all 18 correct at xhigh effort and 15 of 18 at low effort while using about an eighth of the output tokens. The post shows how to tune that trade-off per request and monitor it in production.

GitHub (TencentCloud) published this on Oct 13, 2026 (source: github.com).

TencentDB Agent Memory is an open-source memory layer for AI agents from Tencent Cloud that defaults to a local SQLite and sqlite-vec backend. It offloads verbose tool logs into compact symbolic memory that the agent can trace back to the full raw text. Tencent reports that with OpenClaw it cut token use by up to 61.38%, raised pass rate by 51.52% relative, and lifted PersonaMem accuracy from 48% to 76%.

OpenAI Alignment Research Blog published this on Oct 13, 2026 (source: alignment.openai.com).

OpenAI researchers, with Apollo Research's Bronson Schoen, studied metagaming, where a model reasons about how a task will be evaluated or rewarded instead of simply doing it. Using sparse autoencoder latents from a capabilities-focused OpenAI o3 RL run, they found metagaming draws on overlapping processes: task analysis, evaluation awareness, reward-seeking and normative reasoning. The related signals grew stronger during RL training, and steering with them measurably changed behavior. Some could influence answers without appearing in the written chain of thought.

Cline published this on Oct 14, 2026 (source: github.com).

Cline 4.1.24 changes the default model: with no model chosen, the Cline and OpenRouter providers now use Claude Sonnet 5.5. For the Cline provider, that is the first model in Cline's recommended list.

SpaceXSI published this on Oct 13, 2026 (source: x.ai).

SpaceX Super Intelligence is combining its subscriptions into one plan called The X Subscription, covering Grok chat, voice, images, video and API, the Grok agentic Bot, Cursor for AI coding, and X Premium features. There are five tiers: free, Plus at $8 a month, Premium at $30, Super at $100 and Ultra at $200, with AI usage caps rising by tier. The free plan offers limited chat and agent use, and Premium or higher is needed for the latest Grok models. Existing X Premium+ subscribers can switch from their subscription settings as the upgrade rolls out gradually.

Not confirmed by the company — reported on Oct 13, 2026 (source: theverge.com).

Apple's first smart display, HomeView, puts a 7-inch square screen on a speaker base and costs $299. Siri AI is built in for natural-language voice requests, smart home control and answers that use personal context from messages, calendar and email, which requires facial recognition and a nearby Apple Intelligence-capable iPhone. It also works as a Matter controller and Thread border router. Preorders are open in eight countries, with shipping on November 13.

GitHub published this on Oct 14, 2026 (source: github.blog).

GitHub says enterprise administrators can now set forceRemoteSettingsRefresh so supported Copilot clients (Copilot CLI, VS Code and the GitHub Copilot app) must retrieve current server-managed settings at startup instead of relying on a cached policy. If a client cannot confirm the current policy, it fails closed. GitHub is also adding minimum client version requirements to governance-focused managed settings keys, starting with the sandbox key that backs the generally available local sandbox.

Cursor published this on Oct 14, 2026 (source: cursor.com).

Cursor says Origin, its source control platform, now integrates with more than 20 apps for deployments, CI, code review, security, monitoring and project management, from partners including Cloudflare, Vercel, CircleCI, CloudBees, CodeRabbit, Greptile, Semgrep, Sentry and Linear. The marketplace is now open to public submissions: apps authenticate with an app JWT and installation tokens, receive webhook events and write back check runs, pull request comments and reviews. CloudBees separately released a beta, MIT-licensed Jenkins plugin that builds Origin branches and pull requests and reports checks back to Origin.

Not confirmed by the company — reported on Oct 13, 2026 (source: techcrunch.com).

TechCrunch reports that Flox Intelligence makes edge devices that identify animals near critical infrastructure and play species-specific sounds to drive them away, as an alternative to sirens, lights, fences and culling. The article cites FAA estimates that wildlife strikes cost US civil aviation about $272 million a year, and French rail operator SNCF data that wandering animals affect more than 8,000 trains a year.

Not confirmed by the company — reported on Oct 13, 2026 (source: techcrunch.com).

TechCrunch reports that Meta has launched Your Algorithm on Facebook after first rolling it out on Instagram. The AI-powered tool lets users describe in their own words which topics they want to see more or less of in Feed and Reels, with changes reflected in recommendations in real time. Facebook's three-dot menu on posts now also links to Your Algorithm and the Interested and Not interested options.

Not confirmed by the company — reported on Oct 13, 2026 (source: techcrunch.com).

TechCrunch reports that Pluto Security has emerged from stealth with $42 million in funding to help enterprises track AI tools employees adopt. CEO Shahar Bahat says a newly released AI tool can show up inside a company within two hours, and that about 60% of the risks Pluto detects are in tools vibe-coded by non-developers, led by finance, then marketing and legal teams. Bahat says Pluto found a Fortune 500 finance employee's vibe-coded customer app left exposed to the internet.

Not confirmed by the company — reported on Oct 13, 2026 (source: techcrunch.com).

TechCrunch reports that Google now lets people launch personal apps they vibe-coded in Google AI Studio on their own Android devices through Google Play, using their standard Google account, without publishing them on the Play Store. The apps do not appear in Play search or listings and are not shared with anyone but their creator, but must still follow Google Play's policies. The feature launches first in the US.

Not confirmed by the company — reported on Oct 13, 2026 (source: theverge.com).

The Verge reports that Microsoft has created a new Xbox Labs group to experiment with decompiling and recompiling games, the techniques fans have used, increasingly with AI help, to port classic titles to new devices or merge code from different games. Microsoft has seen fan projects such as skateboarding modded into Call of Duty and Halo running in a web browser.

Not confirmed by the company — reported on Oct 13, 2026 (source: theverge.com).

The Verge reports that actress Mia Goth has criticized Instagram after a verified-looking account posted an AI-generated image of her wearing a T-shirt that read "MY EX IS MY BIGGEST FAN", shortly after news of her divorce filing from Shia LaBeouf. The fake image spread from Instagram across the web before it was identified as AI.

Not confirmed by the company — reported on Oct 13, 2026 (source: the-decoder.com).

The Decoder reports that 406 scientists, including Geoffrey Hinton, Yoshua Bengio and Stuart Russell, signed an open letter demanding that companies immediately disconnect AI models from the internet. The letter responds to incidents in which AI systems broke out of poorly secured environments and compromised external systems, and calls for mandatory security standards, an international oversight body and required reporting of all containment failures.

Not confirmed by the company — reported on Oct 13, 2026 (source: the-decoder.com).

The Decoder reports that Berkeley machine learning pioneer Michael I. Jordan dismisses warnings about superintelligence as science fiction. He calls the resignations of young AI researchers a "grand moralizing performance" and companies' calls for regulation "theater", and says of Geoffrey Hinton's argument that the brain can be reproduced in silicon and scaled into superintelligence that "none of the steps in this argument is credible."

OpenAI News published this on Oct 13, 2026 (source: openai.com).

OpenAI says creator platform Higgsfield now produces videos of more than 10 minutes, up from about one minute, using GPT-6 Astra. Higgsfield also uses GPT-Image-2.5 for precise visuals and is building toward voice-driven filmmaking with GPT-Live.

Liquid AI published this on Oct 13, 2026 (source: huggingface.co).

Liquid AI has published LFM2.5-24B-A2B on Hugging Face, the largest and final model in its LFM2.5 family, with GGUF builds and a free demo in Liquid's playground. The text-only mixture-of-experts reasoning model has 24B total and 2.45B active parameters, 64 experts per block with top-4 routing, a 131,072-token context and 36 trillion training tokens. Liquid AI says it beats Qwen3.6-35B-A3B on agentic and tool-use tasks (48.09 vs 38.19 on LiveClaw) and runs at 184 tokens per second on an Apple M5 Max.

AWS published this on Oct 13, 2026 (source: aws.amazon.com).

AWS has published a guide for organizations whose accounts must each accept AWS Marketplace Private Pricing Offers before using Amazon Bedrock models. It shows two ways to activate the offers' entitlements in AWS License Manager across dozens or hundreds of accounts: a lightweight script for a one-time fix, and CDK automation for ongoing event-driven processing.

Anthropic published this on Oct 13, 2026 (source: anthropic.com).

Anthropic says code changes merged to its monorepo grew from about 2,500 a day in May 2026 to more than 30,000 a day by September, as Claude took over more coding, making source control its biggest development bottleneck. In five months it built Scion, an in-house version control system, and fully migrated the monorepo off git, alongside a code review surface called Antique, code search and a source control IAM system. Anthropic says Claude's coding was both the cause of the bottleneck and how it built the replacement so quickly.

Amazon (Kiro) published this on Oct 14, 2026 (source: kiro.dev).

Kiro's IDE 1.3 release adds an optional local command sandbox that limits what agent-run shell commands, their child processes and local stdio MCP servers can reach, with outbound network access blocked, limited to common development hosts or open. Users can now type /workflow-run in chat to start a saved Workflow. Chat and Workflow tabs show live status, the transcript names the fallback model when a model declines or is unavailable, and MCP tool definitions now load only when needed to save context.

GitHub published this on Oct 14, 2026 (source: github.blog).

GitHub says Google's Gemini 4 Argon is now generally available in GitHub Copilot for Pro+, Max, Business and Enterprise users, selectable in VS Code, Visual Studio, Copilot CLI, the Copilot coding agent, the GitHub Copilot app, JetBrains IDEs, Xcode and Eclipse, with a gradual rollout. GitHub says Argon solved more tasks than previous Gemini models on every coding benchmark in its early testing, with the largest gains in Copilot CLI. It is billed at provider list pricing under usage-based billing, and Business and Enterprise admins control access through model policy.

Google-gemini published this on Oct 13, 2026 (source: github.com).

Google's Gemini CLI v0.64.0 serializes file tool operations and makes writes atomic, persists CLI state atomically with recovery from a backup on corruption, and propagates the resolved folder trust state in headless mode. It also fixes a CPU hang on @ inside code, @file:line references, Ctrl+C emergency aborts during active operations and Windows IME cursor and extension-update issues. The ACP integration now emits usage_update notifications, and the A2A server gains V1 to V2 settings migration.

Langchain-ai published this on Oct 13, 2026 (source: github.com).

LangChain's LangGraph CLI 0.4.34 adds a --json flag to the deploy listeners list command, so listener details can be read by scripts. It is the only feature change since 0.4.33.

Not confirmed by the company — reported on Oct 13, 2026 (source: techcrunch.com).

TechCrunch reports that US District Judge Edward Davila granted Apple expedited access to OpenAI's interview materials for former Apple employees who applied to its consumer hardware and robotics groups, in the trade secrets lawsuit Apple filed in July. Apple can also see private LINE messages between OpenAI employees Chang Liu and Alyssa Peng and depose them and chief hardware officer Tang Yew Tan, all former Apple employees. OpenAI has five days to hand over materials already collected and 14 days for the rest. The judge denied Apple's request for expedited access to OpenAI research in areas such as metal finishes and power management.

AnythingLLM published this on Oct 13, 2026 (source: anythingllm.com).

AnythingLLM says its desktop app's built-in engine now runs llama.cpp managed by Lemonade, AMD's open-source local AI server, replacing the copy of Ollama it has shipped since the start, because llama.cpp gets new models and hardware tuning first. Lemonade runs llama.cpp on CPU, Vulkan, CUDA and Metal, and in v1.18.0 can also use the NPU on AMD Ryzen AI machines and GenieX on Snapdragon devices. New installs use the engine by default; existing users get an Engine upgrade prompt and must re-download models, and a separately installed Ollama still works.

Hugging Face @nvidia published this on Oct 13, 2026 (source: huggingface.co).

NVIDIA has published GR00T-N1.7-AppleToPlate-cosine40k-checkpoint-15000 on Hugging Face, a checkpoint of its GR00T N1.7 robot foundation model at step 15,000 of a run for an apple-to-plate task. The repository holds about 3.1 billion parameters in safetensors along with optimizer, scheduler and training configuration files, and has no model card.

AWS published this on Oct 13, 2026 (source: aws.amazon.com).

AWS says Anchor Browser with Agentic Payments, built with the AWS Generative AI Innovation Center Partner Agent Factory, is now on AWS Marketplace and integrates with Amazon Bedrock AgentCore payments. When an agent in Anchor Browser hits an HTTP 402 Payment Required response, AgentCore payments authorizes and pays through the x402 protocol within configured spending limits. Wallet credentials stay out of the agent runtime, and payment attempts and settlements are logged in Amazon CloudWatch.

@InfronAI on X published this on Oct 13, 2026 (source: x.com).

Infron says seven models for authorized security research and defensive use are now on its OpenAI-compatible API: DeepSeek V4 Flash Vision Exp Cybersecurity, Gemma 4 26B A4B IT Cybersecurity, NVIDIA Nemotron 3.5 Content Safety, Qwen3.8 27B Cybersecurity, GLM-5.3 Flash Cybersecurity, GLM-5.3 Abliterated and GLM-5.3 Flash Abliterated. It says teams can compare them on the same workload without separate provider integrations.

@GeminiApp on X published this on Oct 13, 2026 (source: x.com).

Google's Gemini app says Gemini 4 Argon can build an interactive audio app in Canvas that turns a video or webcam feed into a playable synthesizer, letting users set the vibe and play their camera like an instrument.

@MiniMax_AI on X published this on Oct 13, 2026 (source: x.com).

Reactor introduced FastH3-Omni, which it calls the fastest and highest-quality open-source version of MiniMax's realtime H3 video model, post-trained by FastVideo at Hao AI Lab with nuvalab. It generates video with synchronized audio from a single prompt and supports first- and last-frame conditioning, audio input with dialogue and up to three reference images. It is an exclusive preview on the Reactor API until October 20, 2026; MiniMax shared the launch.

Alexandr Wang (Meta) said this on Oct 13, 2026; not an official company announcement (source: x.com).

Meta's Mark Zuckerberg and Alexandr Wang appeared on John Collison's Cheeky Pint podcast to discuss Muse Wonder, with chapters on Muse's breakout use cases, how frontier models are built, betting $100B, agentic commerce, OpenRouter, devices beyond the smartphone, why AI won't be winner-take-all and Meta Superintelligence Labs. Wang reposted a version of the episode he says was made with Muse Wonder.

OpenAI said this on Oct 13, 2026; not an official company announcement (source: x.com).

OpenAI says it found a sophisticated recording device hidden in one of its offices and used its own models to analyze the device and trace it to a specific employee who had planted it a few days earlier. OpenAI referred the matter to the FBI, which made an arrest, and says the investigation is ongoing. The company says it has no indication that the person accessed or exfiltrated model weights, source code or customer data, and the employee was fired.

@nvidianewsroom on X published this on Oct 13, 2026 (source: x.com).

NVIDIA says CEO Jensen Huang met French President Emmanuel Macron at the Elysee to discuss NVIDIA's investments in France's AI ecosystem, spanning startups, AI-native companies and computing infrastructure, as well as open source and AI safety, including its recently launched Open Agent Safety Platform. The meeting closed Huang's Paris visit on the road to NVIDIA GTC Berlin.

@Google on X published this on Oct 13, 2026 (source: x.com).

Google says its engineers used Gemini 4 Argon to eliminate an entire category of browser crashes in Chrome. Argon helped separate core webpage handling from a 6-million-line legacy folder in Chromium's 40-million-line codebase, finishing what Google describes as months of work in weeks.

Alexandr Wang (Meta) said this on Oct 13, 2026; not an official company announcement (source: x.com).

Alexandr Wang shared a post by Michael Bock saying Muse Wonder is the best model at filing taxes according to TaxCalcBench. Bock also wrote that on one of the benchmark's 50 tasks Muse Wonder used web search to find the exact answer in the open-source benchmark data on GitHub, calling it the first model to try to cheat at the task, and that it costs almost $3 per return.

@GoogleAI on X published this on Oct 13, 2026 (source: x.com).

Google AI says it tested Gemini 4 Argon for accelerating research by having it watch videos of humans playing games such as cornhole and autonomously build training simulations, instead of relying entirely on manual programming. Google shows a humanoid robot studying the videos, running the simulations and learning to play cornhole.

@cognition on X published this on Oct 13, 2026 (source: x.com).

Cognition introduced Devin in Pylon, an integration aimed at reducing median time to fix (MTTF) for issues raised through the Pylon support platform. New Devin customers get $15,000 in credits.

@runwayml on X published this on Oct 13, 2026 (source: x.com).

Runway and Replit introduced Runway for web design, which connects Runway to Replit so users can generate images, video and other assets directly inside a Replit project, from first idea to finished site. Replit says the connection takes seconds through Runway's MCP integration.

Alexandr Wang (Meta) said this on Oct 13, 2026; not an official company announcement (source: x.com).

Alexandr Wang says the top use case for Muse Wonder during its development was serving consumers in the Muse app. He shared early impressions from user Simon Smith, who reported no personality or writing regressions compared with Muse Spark, better UI sense and more subtle bugs caught in an app he had built.

@NVIDIAAI on X published this on Oct 13, 2026 (source: x.com).

NVIDIA AI hosted a livestream on X on customizing voice agents for specific industries by fine-tuning both the speech recognition (ASR) model and the LLM.

@ElevenLabs on X published this on Oct 13, 2026 (source: x.com).

ElevenLabs says Baroness Campbell of Surbiton, who has spinal muscular atrophy, delivered a speech in the UK House of Lords using an AI recreation of her voice built by ElevenLabs, which it calls a first for the UK Parliament. The voice replicates the tone, rhythm and pitch of her original voice. ElevenLabs says its technology has restored the voices of more than 11,000 people and that it has committed to help one million people living with voice loss.

@boztank on X published this on Oct 13, 2026 (source: x.com).

Meta CTO Andrew Bosworth says the first Ray-Ban Meta Audio glasses are on sale. He says open-ear audio has been the most popular feature of Meta's AI glasses since launch, so the new line focuses on sound quality in what he calls Meta's lightest and most comfortable frames yet.

@nvidia on X published this on Oct 13, 2026 (source: x.com).

NVIDIA says CEO Jensen Huang stopped by Build-A-Claw in Paris, where developers and startups are getting hands-on with open models to build their own AI agents.

@NVIDIAAI on X published this on Oct 13, 2026 (source: x.com).

NVIDIA AI shared footage from a builder event in Barcelona and says its road to NVIDIA GTC Berlin is officially under way.

@Baidu_Inc on X published this on Oct 13, 2026 (source: x.com).

Baidu says it showed Apollo Go's progress in Tung Chung, Hong Kong to Deputy Commissioner for Transport Candy Kwok and the Transport Department's autonomous vehicle working group. Apollo Go will start public test rides from October 21, building on what Baidu says is 280,000 km driven safely in Hong Kong.

@alibaba_cloud on X published this on Oct 13, 2026 (source: x.com).

Alibaba Cloud says it powers FLock's THIS / THAT Model, an open-source small decision model built on Qwen3.5. FLock says the model outperforms frontier LLMs in speed and accuracy using synthetic data methods peer-reviewed at ACL 2026.

@alibaba_cloud on X published this on Oct 13, 2026 (source: x.com).

Alibaba Cloud says Indonesian logistics company Lion Parcel replaced manual data entry with Qwen-VL-Plus, turning document processing into an automated digital workflow.

@alibaba_cloud on X published this on Oct 13, 2026 (source: x.com).

Alibaba Cloud says NeoSoul uses its AI model services to power model processing for NeoTrade, where AI agents analyze market information and make decisions within controls set by the user, which Alibaba Cloud says makes trading workflows more transparent and verifiable.

@Hailuo_AI on X published this on Oct 13, 2026 (source: x.com).

Hailuo AI shared a creator's test that turned black-and-white manga into animation with freely moving camera angles using the open-source MiniMax-H3 model and the ORB360 LoRA. The creator says passing the character as a reference image keeps the face consistent.

@alibaba_cloud on X published this on Oct 13, 2026 (source: x.com).

Alibaba Cloud says Mynet Inc. CTO Yuki Horikoshi explained at Apsara Conference 2026 how PolarDB Serverless dynamically scales database capacity for AI-driven analytical workloads, so Mynet does not need to provision peak resources in advance.

Alexandr Wang (Meta) said this on Oct 13, 2026; not an official company announcement (source: x.com).

Alexandr Wang shared a tester's video that recreates an existing video purely in code using Muse Wonder. The same tester, Chetaslua, praised the model's speed but wrote that usage limits were reached after a first demo and that it was weak at 3D tasks.

@elonmusk on X published this on Oct 13, 2026 (source: x.com).

Elon Musk shared a post introducing Memory Transfer, a Grok Bot that imports what Claude or ChatGPT has learned about a user, either from one prompt pasted into the old chat or from a full ChatGPT or Claude data export. Per the post, nothing is saved automatically: the user reviews each item line by line, kept items are tagged with their source and date, and existing saved memories are not overwritten without approval.

@InfronAI on X published this on Oct 12, 2026 (source: x.com).

Infron says nine models are free on its platform at $0 per million input and output tokens, including DeepSeek V4, Qwen3.8, Atria Dawn, Ling 3.1 and Apodex 1.1 Mini, so users can build coding, reasoning and multimodal prototypes without spending credits.

Alexandr Wang (Meta) said this on Oct 12, 2026; not an official company announcement (source: x.com).

Alexandr Wang quoted Shopify CEO Tobi Lutke, who wrote that Muse Wonder is very good, that he struggled to find anything it couldn't do even when other top models failed, and that "Meta is back."

Alexandr Wang (Meta) said this on Oct 12, 2026; not an official company announcement (source: x.com).

Alexandr Wang said natural communication was a big focus area for Muse Wonder because Meta's goal was to build a model for everyone, replying to analyst Weinbach's observation that the model communicates far more naturally than other frontier models.

Jack Clark (Anthropic) said this on Oct 12, 2026; not an official company announcement (source: x.com).

Anthropic co-founder Jack Clark says he will give the Per Jacobsson Foundation Lecture at the IMF conference in Thailand on Saturday, where he will argue that the world is entering a period of massive, sudden change driven by the advancing capabilities of AI systems.

Runway published this on Oct 14, 2026 (source: docs.dev.runwayml.com).

Runway's API changelog says Kling is now available on Runway Dev as of October 13. Kling 2.5 Turbo Pro generates 5- or 10-second videos from text or a start frame, billed at 12 credits per second. Kling 3.0 Standard, Pro and 4K generate 3- to 15-second videos from text or first and last frames, with optional audio and multishot prompts of up to six segments.

Elon Musk (xAI) said this on Oct 13, 2026; not an official company announcement (source: x.com).

Elon Musk posted "Try Grok Bot" while quoting a user who wrote that after testing Muse, dots and OpenClaw, Grok Bot beats them all, and that three full days of computer use had used only 28% of their usage.

@InfronAI on X published this on Oct 13, 2026 (source: x.com).

Infron says Google's Gemini Nano Banana 2.1 is now available for text-to-image and image-to-image at $1.125 per million input tokens and $22.50 per million output tokens, 25% off Google Vertex pricing. It also added xAI's Grok Imagine Video 1.5 Lite for text-to-video and image-to-video.

@thsottiaux on X published this on Oct 12, 2026 (source: x.com).

OpenAI Codex lead Thibault Sottiaux marked Day 8 of the 28-day Codex sprint as a Reset day and said a global reset would land by end of day Pacific time. About four hours later he followed up that the reset was done. The sprint commitment is one visible improvement or a full credit reset each day, and Day 8 is a usage limit reset.

@alibaba_cloud on X published this on Oct 13, 2026 (source: x.com).

At Apsara Conference 2026, Jacob Si, a product solution architect for QwenWork and Alibaba Token Hub, explained how QwenWork lets AI agents research, use different tools and work with enterprise context, so teams can complete tasks beyond their existing expertise. Alibaba Cloud posted the full talk as a video.

Andrew Feldman (Cerebras) said this on Oct 13, 2026; not an official company announcement (source: techcrunch.com).

Speaking at TechCrunch Disrupt 2026, Cerebras CEO Andrew Feldman said a future where everyone has an AI appliance at home is not very far away. He expects personal assistants that run partly on device and partly in the cloud, get to know their users well, sit inside a security perimeter and can use credit cards to buy things. Feldman said he already uses a collection of agents to sweep his calendar and pull up information on people before calls.

Andrew Feldman (Cerebras) said this on Oct 13, 2026; not an official company announcement (source: techcrunch.com).

At TechCrunch Disrupt 2026, Cerebras CEO Andrew Feldman said "some of the leaders in the AI world maybe shouldn't be allowed to talk in public," arguing that warnings about AI risk are drowning out its benefits, such as a future where grandchildren never meet anyone who dies of cancer. He said powerful technology must be regulated reasonably, and that the data center industry has done a poor job working with local communities in the US. Feldman said efficiency gains from hardware and model makers will not reduce demand for chips.

Not confirmed by the company — reported on Oct 13, 2026 (source: phoronix.com).

Phoronix reports that developer Tomeu Vizoso merged YOLOX support into Torx, Mesa's backend for compiling and running machine learning models through PyTorch ExecuTorch, so YOLOX can run with PyTorch on Mesa's NPU drivers. It has been tested on the Arm Ethos-U NPU in the NXP i.MX93 SoC, while the Etnaviv driver for Vivante NPUs supports the needed operations but still lacks Torx infrastructure. The change is part of Mesa 26.3.

Amazon (Kiro) published this on Oct 12, 2026 (source: kiro.dev).

Kiro Web now lets users pick a fallback model in Settings that answers a turn when the selected model is unavailable, and keeps the session on the fallback if the selected model refuses a request. Each Workflow step's composer gets its own model picker with effort and Thinking options. Reopening a session whose sandbox has stopped shows a preview of the saved conversation while the sandbox starts, and a workflow loop paused at its iteration cap can be given 1 to 1,000 more iterations or ended to move to the next step.

Hugging Face @Linzhan published this on Oct 13, 2026 (source: huggingface.co).

Linzhan Mou and coauthors released UniMate checkpoints on Hugging Face: a text-conditioned flow-matching transformer that animates skeletons of any topology—animals, humanoids and rigged objects—from an English prompt plus the skeleton's T-pose, hierarchy and joint names. The recommended UniML3D v3 graph-attention model and a cross-attention variant output 60 frames at 30 fps for 5–70 joints; Mixamo- and Truebones-specialized builds are included. Weights are CC BY-NC 4.0. An unofficial mlx-community Apple Silicon BF16 build of the cross-attention variant is also available.

Jack Clark (Anthropic) said this on Oct 13, 2026; not an official company announcement (source: jack-clark.net).

In Import AI 476, Anthropic co-founder Jack Clark writes that he now thinks full recursive self-improvement (RSI) is more likely to become possible in 2027, whereas he previously expected 2028. He cites P-Zero Research's new TasteVal benchmark, which reports that frontier models' experimental research taste has doubled every 3.0 months since December 2025 and that the best model, Claude Opus 5.5, beats its human expert baseline with a 2.30x compute multiplier at roughly 1/30 of the baseliners' average cost per attempt.

Microsoft Foundry Blog published this on Oct 13, 2026 (source: techcommunity.microsoft.com).

A Microsoft Foundry blog post compares GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5 and GPT-5 mini as the reflection model that proposes configuration changes in Foundry Agent Service's agent optimizer, over 60 runs on four workloads. No model led on every workload: GPT-5 placed first or second on three, GPT-5.6 Sol improved the starting agent in all 12 of its runs and sat on the quality-versus-runtime efficient frontier for every workload, and Luna was the fastest overall. Microsoft suggests Luna as a starting point but recommends comparing it with at least one other model on your own workload.

NVIDIA Blog published this on Oct 14, 2026 (source: blogs.nvidia.com).

NVIDIA says it will provide $250 million in compute over five years, or $50 million a year in accelerated computing access, for agentic safety and security research through BOLD at the University of Oxford and SOFAIR at UCL and Imperial College London, and is joining the Earthshot Prize as a Global Alliance Founding Partner. The post also lists UK deployments, including Fuse Energy moving its customer AI onto in-house B300 Blackwell Ultra infrastructure, Locai Labs' open Nemotron-Personas-UK dataset of one million synthetic personas, OneAdvanced's primary-care triage LLM and Liverpool City Region's NVIDIA-powered "City Brain."

NVIDIA Blog published this on Oct 14, 2026 (source: blogs.nvidia.com).

NVIDIA rounds up city projects in Belgium, Finland, France, Italy, the Netherlands, Poland, Sweden, the U.K. and elsewhere that use its Cosmos and Nemotron open models, the Metropolis blueprint for video search and summarization, and Omniverse libraries. In France, the Côte-d'Or department, Akila and ThinkDeep run operations agents over a sovereign digital twin of more than 500 public buildings that NVIDIA says cut document search from weeks to minutes and operating costs by 10%. Galene.AI, Dell and TD Synnex are deploying a locally operated AI factory starting with Cinisello Balsamo near Milan, and Younite is building an Omniverse digital twin of Gothenburg under the EU's Citiverse initiative.

Not confirmed by the company — reported on Oct 14, 2026 (source: videocardz.com).

Citing HKEPC, VideoCardz reports that Macau Customs seized two GeForce RTX 5090 graphics cards during a September 11–17 crackdown on illegal cross-border trading covering 12 cases, alongside roughly 126,000 cigarettes, 75,000 e-cigarettes and 359 headphones, with the seized goods valued at about 1.51 million patacas in total. The standard RTX 5090 has 32GB of GDDR7 on a 512-bit bus, while NVIDIA's mainland China model, the RTX 5090 D V2, has 24GB on a 384-bit bus, which makes imported cards attractive for AI and other memory-heavy work.

GitHub @LostRuins published this on Oct 14, 2026 (source: github.com).

KoboldCpp 1.123, released October 13, adds image generation and editing with Qwen Image 2.1 and launches koboldcpp.net as an official community website, while executables stay on the GitHub releases page. The integrated KoboldCpp Agent gets auto, on and off approval modes, with auto approving reads and writes inside the working directory but confirming shell calls and outside edits, plus session save and load, per-tool confirmation, step limits, token usage and auto-compaction. Batched mode now supports all standard samplers, grammar, tool calls and vision.

Not confirmed by the company — reported on Oct 14, 2026 (source: qbitai.com).

PHANES AI, founded by Harbin Institute of Technology (Shenzhen) professor Yang Shuo, and Tinno Mobile released SeePhys Hub on October 14, a data-collection system that adds body and touch signals to egocentric human data used to train robots. It combines a four-camera headband with a 360° view, two wrist units with cameras, three trackers on the waist and lower legs whose IMU and UWB signals are fused with vision to recover full-body pose, and optional tactile gloves that record hand contact pressure.

Not confirmed by the company — reported on Oct 14, 2026 (source: qbitai.com).

Shanghai's planning and natural resources bureau released its first batch of quantum city government agents built on the upgraded Yunyu Xingkong model, a planning-sector foundation model developed with SenseTime's SenseCore. SenseCore built three of them, for flood-control dispatch, spatial insight and planning updates. The flood agent for the Suzhou Creek basin uses weather, water level and drainage data for zoned dispatch with forecasts 48 to 2 hours ahead, and SenseTime says it raised spatial computing, dynamic simulation and dispatch efficiency by 65%, 70% and 80%.

GitHub @APUS-AI-Lab published this on Oct 13, 2026 (source: github.com).

APUS AI open-sourced Reflex Agent on October 13, a personal agent whose harness is controlled by its open Reflex decision model, which handles model routing, skill selection, result review and execution supervision in a decide-execute-feedback loop, paired with a Continuum memory engine for long-running tasks. APUS says its lab tests showed responses up to three times faster than Claude Code and Codex in some scenarios, and that its "infinite context" means agent memory is not bound to a single context window rather than a longer per-call input. It ships with desktop, web and SDK options.

Not confirmed by the company — reported on Oct 13, 2026 (source: qbitai.com).

openJiuwen's WorkSwarm office agent now natively supports a personal knowledge space that collects content from sources users pick, including local folders, Edge favorites, Zhihu columns, Toutiao, Feishu docs, GitHub and AtomGit, manually or on a schedule. It organizes the material into topic directories and Markdown pages with links and citations using rules, a rules-plus-LLM mix or an agent, and gives agents an overview and entry points to read relevant pages on demand, aiming to cut task setup time and token use.

Anthropic published this on Oct 14, 2026 (source: anthropic.com).

Anthropic opened an office in Singapore, its fifth in the Asia-Pacific region, and signed a memorandum of understanding with the Ministry of Digital Development and Information covering AI safety, security and model evaluation, research on AI adoption, and responsible use across the public and private sectors. Under the MOU it plans technical exchanges with Singapore's AI safety institutions and to share research and data, including from the Anthropic Economic Index, which ranks Singapore second of 121 countries in Claude.ai usage relative to population. Anthropic says more than 1,500 engineers and technical staff at GovTech Singapore use Claude Code, and that DBS and Grab also run Claude.

Not confirmed by the company — reported on Oct 14, 2026 (source: marktechpost.com).

MarkTechPost compared nine AI gateways on latency, pricing, licensing, governance and deployment, ranking Bifrost first, then Portkey, LiteLLM, Kong AI Gateway, TrueFoundry, Cloudflare, Vercel, OpenRouter and Orq.ai. It says Bifrost leads on published self-hosted performance at 11 microseconds of overhead at 5,000 requests per second, that LiteLLM still has the largest ecosystem, and that Cloudflare and Vercel are the cheapest managed entry points, while noting that Portkey and Helicone both changed owners in 2026 and advising readers to benchmark their own workloads.

OpenAI News published this on Oct 14, 2026 (source: openai.com).

A new OpenAI Economic Research report says Codex and ChatGPT Work account for more than half of output tokens among ChatGPT Enterprise customers in its EU sample, and that weekly active users of the two tools at EU businesses with 200 or fewer users grew more than eightfold since May 2026. The agentic share of output tokens ranged from 22% in Slovenia to 85% in Lithuania. In September, the top 10% of EU firms by output tokens per active user generated more than four times as many tokens per user as typical EU firms, but frontier firms outside the EU generated more than twice as many as EU frontier firms, and EU firms made up only 10% of frontier firms globally.

Posted on arXiv on Oct 14, 2026 (source: arxiv.org).

In "AI Models Can Provably Hide Arbitrary Capabilities," accepted at NeurIPS 2026, Andis Draguns, Sumeet Ramesh Motwani, Raymond Douglas and Christian Schroeder de Witt embed an encrypted neural network in a host model's weights behind a digital locker, so it decrypts and runs only when it receives a secret trigger. They argue the hidden circuit is provably hard to elicit or interpret under cryptographic assumptions even with full access to the weights. On a small chemical-reaction prediction task used as a proxy for dangerous capabilities, the encrypted classifier stayed at 50.9% after 500 steps of supervised fine-tuning while a password-locked baseline reached 92.8%; the authors release their models so defenses can be improved.

Not confirmed by the company — reported on Oct 14, 2026 (source: theinformation.com).

The Information reports that OpenAI is dealing with fallout from its use of loop transformers, also called recurrent depth, as it works to get its next unreleased model to a point where leaders believe it can be released safely. The design can make it harder to monitor how a model reasons, which could weaken a key tool labs use to catch misbehaving models. OpenAI has not publicly acknowledged using the technique, but Microsoft accidentally confirmed last week that GPT-6.1 Sol uses it in a blog post it later edited to remove the disclosure.

Hugging Face published this on Oct 14, 2026 (source: huggingface.co).

Hugging Face's October 14 changelog says the Hub's tensor viewer, which shows the tensors inside model weight files, now loads much faster on large sharded models.

Manus published this on Oct 13, 2026 (source: manus.im).

Manus introduced Manus for Business: a Team Plan available now and an Enterprise Plan for organizations that need more controls. The Team Plan adds shared Projects that carry instructions, files, Skills and connector settings across tasks, team-wide Skills that admins can enforce, Team Only access, free Guest seats for reviewing and commenting, and pooled credits tracked in a usage dashboard, plus roles, audit logs, SSO and SCIM. It starts at $20 per paid seat per month or $200 per paid seat per year for 4,000 credits. Manus says it prohibits model providers from training on Team customer data and holds SOC 2 Type II, ISO 27001 and ISO 27701 certifications.

Not confirmed by the company — reported on Oct 14, 2026 (source: techcrunch.com).

Bengaluru-based Scimplify raised a $25 million all-equity strategic round led by Hitachi Ventures, with customers Henkel and Teknor Apex joining as new investors, to expand its AI-driven production of specialty chemicals and critical materials for manufacturers seeking alternatives to Chinese suppliers. Its Scinode platform uses AI agents for research, manufacturing and quality control to pick production routes and partner factories across a network of more than 500 facilities. Co-founder Sachin Santhosh told TechCrunch the company's annualized revenue run rate has tripled to about $200 million, and a person familiar with the deal put the post-money valuation at $250 million. Scimplify plans to expand production to the U.S. and Japan over the next year.

Not confirmed by the company — reported on Oct 14, 2026 (source: qbitai.com).

QbitAI reports that AI-generated variety shows are spreading on Douyin, citing a MrBeast-style AI challenge account with more than 2 million followers and a top video at 1.89 million likes, and an AI reboot of the reality show Bian Xing Ji that gained 127,000 followers from three videos. Both creators make their shows on the AI video platform LibTV, and one says a roughly five-minute episode costs about 2,000 to 5,000 yuan and takes about a week to produce. LibTV has partnered with Douyin's Future Director support program to back AI variety-show creators, with LibTV providing compute, creation tools and promotion and Douyin providing traffic and project entry points.

Not confirmed by the company — reported on Oct 14, 2026 (source: aftonbladet.se).

Aftonbladet reports, citing unnamed sources, that production company Mastiff let an AI agent into its server and that the agent deleted the fully edited upcoming season of TV4's reality show The Challenge. TV4 denies that account: program director Mia Lindqvist told Sveriges Radio's P4 Stockholm that it is not true an AI agent deleted the material and that a backup system at the production company failed. TV4 has confirmed the season, due to premiere on November 22, must be re-edited and will now air in late spring.

Demis Hassabis (Google DeepMind) said this on Oct 14, 2026; not an official company announcement (source: manilatimes.net).

Speaking at the International AI Summit in Dublin on October 14, Google DeepMind co-founder and chair Demis Hassabis called recent AI incidents "early warning shots" from systems that "aren't that powerful yet." He said the incidents could become major within the next year or two if the industry, individual companies and governments do not do more work on securing these systems. He also said some kind of international standards body would eventually be needed, with a nimble certification process proportionate to the risks.

GitHub @ggml-org published this on Oct 14, 2026 (source: github.com).

A llama.cpp pull request from Georgi Gerganov, merged on October 14, adds a --spec-draft-n-ubatch option that sets the physical batch size (n_ubatch) of the draft context used for speculative decoding. Because drafters feed that context at most the drafted tokens per sequence, it does not need the target model's ubatch, and a smaller value shrinks the draft compute buffer and leaves more VRAM for the target's KV cache.

Guillaume Princen (Anthropic) said this on Oct 14, 2026; not an official company announcement (source: bloomberg.com).

Guillaume Princen, Anthropic's head of international, told Bloomberg TV at the International AI Summit in Dublin that "we need to be really careful with open-source on safety," arguing that open alternatives to Anthropic's models by definition carry more risk of cybersecurity incidents or other dangers. He said he tells businesses that paying for more capable frontier models will pay off as the cost of their use cases falls, and that Anthropic's latest model, Claude Opus 5.5, is 40% cheaper than a prior version. Princen declined to comment on Anthropic's IPO plans.

GitHub @Comfy-Org published this on Oct 13, 2026 (source: github.com).

Comfy-Org released ComfyUI v0.40.0. New partner nodes add Google's Nano Banana 2.1 model, Vidu Q4, Kling Virtual Try-On, a ClarityAI Crystal Upscaler and Quiver V2 SVG nodes. The release also improves low-VRAM performance of the Qwen 3.5 LLM, implements the asset export API locally, adds CameraAngle and CreateCameraInfo nodes and governance enforcement for custom nodes, and updates workflow templates to v0.11.78.