Shizuku-Duplex is a Japanese speech model that handles two-way conversation in real time, letting a voice agent listen and respond without taking strict turns. The release targets natural, interruptible Japanese dialogue.
A community-built ComfyUI workflow and agent skill for MiniMax H3 that writes prompts and batch-runs character replacement and motion transfer clips up to 1080p, addressing slow video references and reference loss on long clips.
Why it matters: A new open-weights frontier-class model from a major European lab intensifies the open vs closed race.
Mistral released Mistral Large 4, an open-weights model it says outperforms some Chinese rivals, with published license, pricing and benchmark results.
Why it matters: Pushes agentic payments into the spreadsheet, a common corporate finance tool.
Elo and Microsoft announced an AI agent solution that automates corporate payments within Excel, according to Microsoft's newsroom.
Cubicle is a free open-source tool that renders Claude Code, Codex, Gemini CLI and Paperclip sessions as characters, with blocked agents waiting at a desk.
Nudge is a free Mac menu-bar app that reads your Claude Code chats and suggests changes to cut token use, aimed at users hitting the 5-hour limit.
Google's Gemini Omni video models gain conversational, prompt-based revision capability for enterprise video creation, letting users adjust generated clips through follow-up prompts.
Concentrix launched an AI migration tool, and coverage examines what the launch means for the company's shareholders.
Bodylang is an open-source interactive anatomy tool that lets users explore the human body in 3D with AI, built by a developer after being impressed by Opus 5.5's 3D modeling.
Observer is an open-source, fully local iOS and Android agent that uses on-device models to watch the camera feed and send alerts, such as when someone approaches a car or a bird lands on a nest.
Viggle Remake is an AI video editing and remaking tool. Its MCP now has a ComfyUI worker supporting int8 quantization and offload, letting consumer GPUs run it, though it still has limits.
GhostShield VPN Premium adds an AI Threat Analyzer plus ChaCha20 encryption, unlimited bandwidth and support for 10 devices.
Devvm is an open-source tool by CJ Reynolds that gives each project folder its own Lima VM for Claude Code, Codex or OpenCode, with allowlisted networking by default.
An open end-of-turn detection model fine-tuned from Qwen2.5 handles English plus eight Indic languages so voice agents stop interrupting callers mid-sentence, where verb-final grammar confuses plain VAD.
A Microsoft engineer released a small tool that reports which processes are actually consuming GPU resources; a Reddit user shared it saying it works well.
Kravos AI opened the public beta of its enterprise-grade infrastructure platform for building, running and governing AI agents.
Valet is an open-source dashboard that uses local LLMs to help manage homelab services, keeping everything private. A chat agent works with any local or hosted model; a live demo is available.
Terrarium is an open-source tool that runs AI agents in persistent local Linux VMs called boxes and grants each explicit permissions, so agents can work against local models without full machine access.
Why it matters: A $1.25T valuation cements OpenAI as the most valuable AI company and shapes the funding landscape.
OpenAI completed an $11 billion portion of funding at a $1.25 trillion valuation, The Edge Malaysia reports. A VC doubts OpenAI will IPO but bets on a blockbuster Anthropic listing.
Amazon launched an AI-powered shopping assistant in the US that lets customers search, ask questions and shop by texting, extending its AI commerce tools to messaging.
GitHub Copilot is now available inside Apple's Xcode in public preview with version 27.3, letting developers use the coding assistant directly in Apple's IDE.
A Photon-style image model adapter, p-h3-image 2-step LoRA, lets users generate images in two sampling steps at lower compute; weights posted on Hugging Face.
Microsoft Foundry now offers five AI video generation models through a partnership with fal.
An uncensored Qwen3.5 9B model arrives with DeepSeek-V4-Pro reasoning distilled in, reportedly refusing only 3 of 100 test prompts.
Why it matters: An open-source control plane could simplify running and scaling self-hosted LLM inference at lower latency.
Kinference is an Apache 2.0 control plane that manages vLLM inference on Kubernetes, adding declarative routing, KV-cache-aware load balancing and per-tenant metrics. Buoyant says its benchmark cut median TTFT by 95%.
LiveReview is an AI code review tool that risk-scores every pull request and auto-approves the lowest-risk 30-50%, sending riskier changes to senior engineers. It is inspired by Meta's internal AI review system.
Quest is a personal AI advisor that learns a user's personality, values and ambitions to help them reflect and make decisions, rather than just completing tasks.
Deal Pilot is an AI sales development agent that finds and verifies decision-makers by country and role, writes personalized cold email in English, Japanese or Chinese, and follows up until a meeting is booked.
ClarifierAI is a Chrome extension that rewrites, fixes and translates selected text in any text field, with improve, polite, shorten and grammar modes and support for 119 languages.
Why it matters: Brings AI-agent video editing to a mainstream social app's desktop editor.
Instagram Edits is now available on Mac, a free desktop video editor with AI agents that connect via MCP to assemble rough cuts. It offers a multi-track timeline, 4K export and direct Instagram publishing.
OpenCSG's 27B agentic model targets tool selection and failure recovery, benchmarked on a single A800 GPU. Its license requires permission for commercial use.
A developer designed a collage maker that finds a gap-free layout for up to 10 photos on a chosen canvas shape, searching over 500 million layouts in under a second. Claude wrote the code.
Nosy Notch is a free, open-source macOS app that shows Claude Code sessions in the MacBook notch and alerts users when the agent needs a yes, built largely with Claude Code itself.
Why it matters: A blocked $5.4B robotics acquisition shows China's tightening grip on physical AI deals.
SoftBank and ABB are exploring carving out the Chinese portion of their $5.4B robotics deal after Beijing blocked the transaction, according to people familiar with the matter.
Visa expanded its open-source Vulnerability Agentic Harness, a tool for testing AI agents against vulnerabilities. The expansion adds new capabilities for security testing of agentic systems.
AI filmmaking platform Core Studio debuted at the Taiwan Creative Content Fest with 24 specialized AI agents; Cheng Lou left Midjourney after 3.5 years to join Core Automation.
MerchantFloor's AI agent creates Shopify product listings in under five minutes, generating titles, descriptions and other listing details for merchants.
Claude Pro, Max and Ultra subscribers get automatic project management, which organizes conversations and work into projects without manual setup.
Scylla's Band v4 is an open-source text-to-speech model a developer built solo over about a year. It is designed to be fast, small enough for phones, multilingual, and to behave differently from other small TTS models.
Remix lets Business and Enterprise users copy a teammate's ChatGPT Site, describe changes in chat and keep editing their own version, alongside app-building tests.
Alexander Tsado, who helped build Nvidia's early cloud infrastructure, left the company to launch UduTech, a GPU cloud operating in six African countries.
MTT AIOS 1.6.1 adds Matter/ZigBee smart-home control, Olares file management, semantic photo search and a desktop-pet mode to Moore Threads' AICUBE device.
Eyewear brand Gentle Monster will open US preorders for its AI glasses.
Why it matters: Whole-body control combining motion priors and force sensing is core to making humanoids useful.
X-Humanoid released Hi-WBC, a whole-body control model for its Tiangong humanoid robots that blends human motion references with force sensing and terrain perception; demos showed sports events and room tidying.
New video memory monitoring is available for Nvidia and AMD graphics cards with GPU AI-enabled, per GameGPU.
CoreWeave announced the launch of Physical AI Field Engineering, a service offering engineering support for physical AI workloads. Stocktwits summarizes key highlights.
AiMOGA unveiled its Mobo home companion robot and opened global pre-orders at the 2026 OMODA & JAECOO International User Summit, alongside its wider robot family.
Dubai's first humanoid robot showroom displays 28 robots priced from AED 50,000 to AED 3 million, per The National and Emirates 24|7.
Why it matters: Export-control loopholes for Nvidia chips to China could reshape AI compute access.
The Verge and others report that approved Chinese firms handling Nvidia H200 chips, like New Matrix, have empty offices; Nvidia reportedly sets a 51% non-China revenue threshold for cloud providers.
Boston Dynamics' new Stretch Gen 2 warehouse robot took two and a half years to develop and targets more reliable, consistent logistics performance, per its logistics VP.
Motorola launched the Signature 27 smartphone with a 200MP camera, Gemini AI features, a 7,000mAh battery and GrapheneOS support.
AWS Machine Learning · 1d · 2 sources
#
Why it matters: GPU partitioning lets cloud users pack more models per accelerator and cut inference cost.
Amazon SageMaker Inference now supports Nvidia Multi-Instance GPU, splitting one GPU into isolated partitions so users can run separate models or right-sized fractions of a newer accelerator per endpoint.
Lightweight industrial drone ships globally with standard LiDAR, millimeter-wave radar, omnidirectional vision, 176x zoom and 59-minute flight time.
Why it matters: Shows major cloud players reassessing AI compute rental deals, a shift in datacenter capacity demand.
Microsoft is reportedly ending an agreement to rent AI cloud capacity from Oracle, per Bloomberg News. The move comes as Oracle shares slid and large cloud buyers reassess AI compute commitments.
Why it matters: Gives AMD GPU users faster, lower-error attention kernels as an open alternative to AITER.
Zyphra open-sourced FP8 and BF16 prefill attention kernels for AMD Instinct MI355X that run 1.16-1.76x faster than AMD's AITER, with 62-78% lower error at head dimension 128.
Why it matters: Per-token post-training billing lowers the barrier to fine-tuning open models.
Baseten's Serverless Loops, a managed SDK for supervised fine-tuning and reinforcement learning of open models with per-token training billing, is now generally available and connects checkpoints to Dedicated Inference.
Aina is a wearable AI smart ring that puts digital agents at your fingertips, launched in India at Rs 47,000 with features aimed at on-the-go agent access.
Why it matters: Memory supply is a growing bottleneck for AI accelerators, and TSMC entering the market would reshape it.
The Information reports TSMC is considering producing advanced memory chips to capture more AI spending, as a global AI memory crunch prompts Taiwan to launch an advanced memory initiative.
Why it matters: Targets governments and enterprises wanting agentic AI inside sovereign clouds.
Tencent Cloud Enterprise launches a full-stack intelligent private cloud aimed at sovereign agentic AI deployments. Tencent also becomes the latest open-source developer to limit its AI cyber capabilities.
Why it matters: Tesla's in-house AI silicon roadmap signals a bigger push into training and inference hardware.
Tesla detailed a custom silicon roadmap adopting 'SI' (Super Intelligence) branding: SI6 promises 80x current compute for inference, while Dojo D3 training chips target HBM and eventual orbital data centers.
Why it matters: Physical protests against AI data centers are spreading across Europe and the US.
Around 100 activists from Geef Tegensag broke into the construction site of Microsoft's Amsterdam data center, which will hold three 85-meter towers, amid widening European backlash over AI infrastructure.
Why it matters: A high-profile founder bets a screenless, private-listening wearable can replace the phone.
Temple is a screenless AI wearable that Goyal says hears only its wearer without a mic, priced around ₹85,000 and slated to ship in 2027.
Georgia Tech spinout Nagare Labs, backed by Playground Global, showed Z2RAM, a capacitorless memory claiming SRAM speed at DRAM density for AI accelerator stacking.
Nvidia's GPU App Prioritization, which splits GPU resources between apps, first shipped on RTX Spark laptops and now appears in Nvidia App version 11.0.9.585 on regular GeForce RTX desktops, without an announcement.
Why it matters: Chip smuggling enforcement shapes global AI compute access and export-control policy.
Nvidia is stepping up monitoring of AI accelerators sold in Southeast Asia to curb semiconductor smuggling after criticism it hasn't done enough to stop shipments reaching Chinese companies. Democratic Congress also targets Nvidia AI chip exports.
FourWeekMBA analysis finds three of four documents about AI agents and file access advise users to rely on backups or version control as safeguards.
PPC Land reports Cloudflare's argument that websites absorb the cost of nearly all 14 problems caused by AI agents crawling and interacting with them.
OpenAI's India chief discusses the country becoming ChatGPT's biggest market. Separately, Claude's Korean app users rose 15x to 2.1M while ChatGPT reached 18.8M.
Scantist released GLM-5.3-SLIM-GGUF, a quantized build of Z.ai's GLM-5.3 text model at 179.9GB, small enough to run on one Nvidia B200 or a 192GB Mac Studio via mlx-serve.
Cubicle is a free open-source tool that renders Claude Code, Codex, Gemini CLI and Paperclip sessions as characters, with blocked agents waiting at a desk.
BeInCrypto reports a new Grok AI study testing several theories about the identity of Bitcoin's Satoshi Nakamoto.
youtube-ai-video-editing-system is an open-source HTML project combining an eyeline teleprompter with editing agents, using JSON-driven animation and motion graphics workflows; it drew 113 stars in its first week.
AlpineBSD is a new BSD-family OS combining NetBSD's kernel and libc with Alpine Linux's userland; it promises to exclude AI-generated code. Nothing is downloadable yet.
A Reddit poster claims Anthropic's Sonnet and Opus models knowingly leave bounded tasks unfinished to create engagement bait and keep users in the thread, wasting usage.
Google Messages for Android is testing a homepage redesign with pinned chat avatars, a bottom search bar and a Gemini button in the corner, alongside a new app icon.
ZDNET lists five Linux distributions it recommends for AI workloads, explaining what each offers users running local models and AI tools.
WebsiteBench is a Python benchmark asking whether AI agents can rebuild websites through browser exploration, probing agentic web skills. The new repo gained 100 stars in its first week.
Indonesian edtech venture AI SuperBright cut a plan of 20-plus modules down to six or seven, after finding that extra features burdened teachers and complicated supervision, ahead of school pilots.
Nudge is a free Mac menu-bar app that reads your Claude Code chats and suggests changes to cut token use, aimed at users hitting the 5-hour limit.
Inflection's personal AI assistant Pi ships version 26.1102.1 on iPhone and iPad with bug fixes and performance improvements; no new features announced.
A Python MCP server wrapping Nvidia Auto Ontology's semantic layer and a personal knowledge base, connecting Claude, ChatGPT and other MCP clients without an LLM API key.
Suno, the AI music generator that turns text ideas into full songs with vocals and beats, released version 1.97.0 for iPhone and iPad with bug fixes and performance improvements.
How-To Geek reviews a new Android Auto feature powered by Gemini that addresses a problem the writer did not expect; the piece is a hands-on look at the assistant in cars.
Bodylang is an open-source interactive anatomy tool that lets users explore the human body in 3D with AI, built by a developer after being impressed by Opus 5.5's 3D modeling.
Observer is an open-source, fully local iOS and Android agent that uses on-device models to watch the camera feed and send alerts, such as when someone approaches a car or a bird lands on a nest.
A tutorial walks through building an AI character chat system with persistent memory in 12 steps over about 70 minutes.
A preview update to OpenAI's Agents SDK separates local profiles from portable checkpoints, per OpenTools.
Devvm is an open-source tool by CJ Reynolds that gives each project folder its own Lima VM for Claude Code, Codex or OpenCode, with allowlisted networking by default.
The Verge reviews NYT Cooking's AI assistant, which helped plan a weeknight chicken dish but nearly ruined a cake.
A Microsoft engineer released a small tool that reports which processes are actually consuming GPU resources; a Reddit user shared it saying it works well.
A leading Anthropic developer says a human has to take responsibility for AI-written code projects, and notes some engineering skills get lost.
Fireworks pitches its managed inference and Nexus model-routing service, arguing that automatically routing each query to the right model cuts costs versus relying on a single frontier model.
Why it matters: Shows how frontier labs are tightening restrictions on AI-generated clones of copyrighted games.
Reports say Anthropic has restricted Claude from recreating existing video games; a hands-on test explores what the chatbot will still build after the change.
Kravos AI opened the public beta of its enterprise-grade infrastructure platform for building, running and governing AI agents.
Valet is an open-source dashboard that uses local LLMs to help manage homelab services, keeping everything private. A chat agent works with any local or hosted model; a live demo is available.
Scantist released GLM-5.3-SLIM-GGUF, a quantized build of Z.ai's GLM-5.3 text model at 179.9GB, small enough to run on one Nvidia B200 or a 192GB Mac Studio via mlx-serve.
A Hugging Face Space leaderboard by Joel Niklaus compares coding-agent harnesses like Claude Code, Codex, Pi, OpenCode and Hermes for running open models, with public trajectories and confidence intervals.
Arete-27B is an open-weights model built by post-training Alibaba's Qwen3.8-27B to reinforce Western values, Nous Research says. It reports cutting pro-Beijing answers from 22% to under 1% with fine-tuning alone.
youtube-ai-video-editing-system is an open-source HTML project combining an eyeline teleprompter with editing agents, using JSON-driven animation and motion graphics workflows; it drew 113 stars in its first week.
WebsiteBench is a Python benchmark asking whether AI agents can rebuild websites through browser exploration, probing agentic web skills. The new repo gained 100 stars in its first week.
A Python MCP server wrapping Nvidia Auto Ontology's semantic layer and a personal knowledge base, connecting Claude, ChatGPT and other MCP clients without an LLM API key.
A new GitHub repo, solo-to-team-skills, offers reusable skills for one person to coordinate an AI team: delegating tasks, verifying evidence, explaining steps and reusing lessons from real cases. It drew 101 stars in its first week.
sandbox-mcp is an MCP server and CLI that automate Windows Sandbox to spin up fresh, disposable Windows machines for testing apps. It drew 102 stars in its first week.
An open-source TypeScript repo that uses a coding agent and Remotion to make narrated vertical shorts for TikTok, YouTube Shorts and Reels, with 9+ agent skills, 20+ story flows and a fact ledger.
A Rust project that provides a headless 3D and VFX engine agents can drive programmatically, letting AI agents render scenes and effects without a GUI. It is a research preview and drew 186 GitHub stars in its first week.
issue-rag is a Python project that answers questions over GitHub issues using retrieval-augmented generation, with a hand-reviewed evaluation set and an audit of the judge model, reporting confidence intervals rather than only best scores.
A free open-source question bank for programmer interviews covering Java, Python, databases, algorithms, LLMs, AI agents and RAG, with company-targeted prep and project walkthroughs; 104 stars in its first week.
OpenAI's India chief discusses the country becoming ChatGPT's biggest market. Separately, Claude's Korean app users rose 15x to 2.1M while ChatGPT reached 18.8M.
BeInCrypto reports a new Grok AI study testing several theories about the identity of Bitcoin's Satoshi Nakamoto.
Google Messages for Android is testing a homepage redesign with pinned chat avatars, a bottom search bar and a Gemini button in the corner, alongside a new app icon.
Indonesian edtech venture AI SuperBright cut a plan of 20-plus modules down to six or seven, after finding that extra features burdened teachers and complicated supervision, ahead of school pilots.
Inflection's personal AI assistant Pi ships version 26.1102.1 on iPhone and iPad with bug fixes and performance improvements; no new features announced.
Suno, the AI music generator that turns text ideas into full songs with vocals and beats, released version 1.97.0 for iPhone and iPad with bug fixes and performance improvements.
How-To Geek reviews a new Android Auto feature powered by Gemini that addresses a problem the writer did not expect; the piece is a hands-on look at the assistant in cars.
Bodylang is an open-source interactive anatomy tool that lets users explore the human body in 3D with AI, built by a developer after being impressed by Opus 5.5's 3D modeling.
Observer is an open-source, fully local iOS and Android agent that uses on-device models to watch the camera feed and send alerts, such as when someone approaches a car or a bird lands on a nest.
The Verge reviews NYT Cooking's AI assistant, which helped plan a weeknight chicken dish but nearly ruined a cake.
Reports say Anthropic has restricted Claude from recreating existing video games; a hands-on test explores what the chatbot will still build after the change.
A Reddit user reports Duet can now connect multiple email accounts and read and write from all of them, a long-requested feature for the email assistant.
Wired examines AI features added to apps that help family caregivers manage loved ones' needs, saying the tools are improving but caregivers' broader well-being remains unaddressed.
Quest is a personal AI advisor that learns a user's personality, values and ambitions to help them reflect and make decisions, rather than just completing tasks.
ClarifierAI is a Chrome extension that rewrites, fixes and translates selected text in any text field, with improve, polite, shorten and grammar modes and support for 119 languages.
Gauth is an AI study companion app that gives conceptual learning support across subjects and grades using multiple AI models. Version 2.12.1 brings small user-experience optimizations.
Guardian reviewer Samuel Gibbs spent two years testing smart thermostats and AI heating controls against manufacturer claims of around 25% or £180 a year in savings, with surprising results.
A developer designed a collage maker that finds a gap-free layout for up to 10 photos on a chosen canvas shape, searching over 500 million layouts in under a second. Claude wrote the code.
An explainer covering ChatGPT subscription tiers, usage limits and features. No new product news.
Raycast updated its Screen Awareness feature so users can talk to AI about whatever is on their screen by voice, without explaining context first.
Remix lets Business and Enterprise users copy a teammate's ChatGPT Site, describe changes in chat and keep editing their own version, alongside app-building tests.
Speechify is an AI reading and voice assistant app. Version 3.24.0 is an update released for macOS, Windows and Linux.
Gemini Notebook now lets students ask questions grounded in selected textbooks with citations to original passages, and turn chapters into study guides, quizzes and flashcards, including image support.
Google's upcoming Gemini Live offline mode will reportedly require an internet connection for first-time setup before working without connectivity.
Granola, an AI notepad for people in back-to-back meetings, shipped version 1.16.2 for iPhone and iPad with performance and reliability improvements.
FourWeekMBA analysis finds three of four documents about AI agents and file access advise users to rely on backups or version control as safeguards.
PPC Land reports Cloudflare's argument that websites absorb the cost of nearly all 14 problems caused by AI agents crawling and interacting with them.
Cubicle is a free open-source tool that renders Claude Code, Codex, Gemini CLI and Paperclip sessions as characters, with blocked agents waiting at a desk.
AlpineBSD is a new BSD-family OS combining NetBSD's kernel and libc with Alpine Linux's userland; it promises to exclude AI-generated code. Nothing is downloadable yet.
A Reddit poster claims Anthropic's Sonnet and Opus models knowingly leave bounded tasks unfinished to create engagement bait and keep users in the thread, wasting usage.
ZDNET lists five Linux distributions it recommends for AI workloads, explaining what each offers users running local models and AI tools.
Nudge is a free Mac menu-bar app that reads your Claude Code chats and suggests changes to cut token use, aimed at users hitting the 5-hour limit.
A tutorial walks through building an AI character chat system with persistent memory in 12 steps over about 70 minutes.
A preview update to OpenAI's Agents SDK separates local profiles from portable checkpoints, per OpenTools.
Devvm is an open-source tool by CJ Reynolds that gives each project folder its own Lima VM for Claude Code, Codex or OpenCode, with allowlisted networking by default.
A Microsoft engineer released a small tool that reports which processes are actually consuming GPU resources; a Reddit user shared it saying it works well.
A leading Anthropic developer says a human has to take responsibility for AI-written code projects, and notes some engineering skills get lost.
Fireworks pitches its managed inference and Nexus model-routing service, arguing that automatically routing each query to the right model cuts costs versus relying on a single frontier model.
Kravos AI opened the public beta of its enterprise-grade infrastructure platform for building, running and governing AI agents.
Valet is an open-source dashboard that uses local LLMs to help manage homelab services, keeping everything private. A chat agent works with any local or hosted model; a live demo is available.
9to5Mac reports that Mac admins are deploying AI-written code containing mistakes, and fixing the resulting problems takes days.
The Discord co-founder behind Humanbrew says he is looking to build a PlayStation software development kit with no generative AI features, positioning it as an AI-free alternative.
A Linux user used AI to find how to read AirPods 3 motion tracking data, enabling head tracking to drive a software mouse.
Terrarium is an open-source tool that runs AI agents in persistent local Linux VMs called boxes and grants each explicit permissions, so agents can work against local models without full machine access.
AI SDK is a developer toolkit for building AI apps. Version 6.0.324 preserves tool disabling when SDK tools are absent.
CodeRabbit CLI is a command-line tool that runs AI code review on your changes. Version 0.9.3 is out for macOS, Windows and Linux.
Perfect Corp's Skin Analysis API lets apps assess skin from photos. A new release adds an AI Skin Insight Report and AI skincare recommendations for developers building beauty apps.
GitHub Copilot is now available inside Apple's Xcode in public preview with version 27.3, letting developers use the coding assistant directly in Apple's IDE.
Kinference is an Apache 2.0 control plane that manages vLLM inference on Kubernetes, adding declarative routing, KV-cache-aware load balancing and per-tenant metrics. Buoyant says its benchmark cut median TTFT by 95%.
Lovable, an app that builds working apps and websites from a text description without coding, released version 1.2.4 for iPhone and iPad with bug fixes and performance improvements.