Microsoft Research and UC Berkeley released TokenEconomizer, an open-source plugin that learns from past sessions to cut coding agent token costs 18.4% to 30.9% on SWE-Bench without changing the agent.
AWS describes a controlled local setup that routes the ChatGPT desktop app's model inference to Amazon Bedrock in AWS GovCloud (US-West) using API-mode sign-in, device policy and network restrictions.
Crisp launched Halo, a retail AI platform that runs inside a company's own cloud environment, bringing sales and supply-chain analysis closer to existing data and governance systems. It targets CPG manufacturers.
Why it matters: Signals OpenAI moving deeper into regulated enterprise deployments with data-residency guarantees.
OpenAI unveiled the package at DevDay on September 29th, pairing available zero-retention safety processing with a preview of private inference for enterprises handling sensitive data.
doomformer is a project by alp that compiles an integer neural network into a DOOM map, using 1,355 voodoo dolls for arithmetic, with values read back and compared against NumPy.
Kadmos-1 is a small decoder-only transformer built from scratch that holds short conversations and writes small Python functions; architecture, tokenizer and weights are in the GitHub repo.
Sebastian-Ion Nae put two small neural networks on a Raspberry Pi Pico 2 W: a world model predicting next frame and reward, and an actor that plays generated games; code on GitHub.
Uwear.ai is a tool for generating fashion visuals: on-model photos and video from garments, aimed at brands and retailers producing product imagery with AI.
Community quantizer Unsloth released Dynamic V3.1 preview GGUF quantizations of DeepSeek-V4.1-Flash-1017, saying they retain the most accuracy for their size across seven benchmarks; a separate post compared ISTA's GSQ-RCO quants.
Local Code is an open-source AI assistant for Mac that lets users download models and run them locally without separately installing dependencies. The Draw Things team built it as a free, private tool.
Synthflow released Vector, an AI agent for lending workflows that it says delivers 40% higher reach and $1M in commitments; the company makes voice AI agents.
Tria launched Agentic Trading, letting AI agents execute crypto trades while users keep control of their funds; the product targets agents that trade without custody.
Shufti launched a system that links AI agent actions to verified human owners and their permissions, aiming to establish accountability and identity for autonomous agents.
SenseNova-Loop-9B-Preview is a 9B-parameter multimodal language model for long-context and agentic workflows, accepting text and images and combining a looped decoder.
Glog.AI is an AI software-security platform offering vulnerability triage and contextual remediation, pitching AI-assisted fixes for code security issues.
A community fine-tune applies the method behind Qwen3.8-27B-Humanlike-Chat 2.0 (136k GGUF downloads across v1 and 2.0) to Google's official Gemma 4 31B, aiming for more natural, human-sounding chat.
Why it matters: An unusual biological-computing approach to video generation now publicly available.
The Biological Computing Co. released Mischief, described as the first publicly available neuron-derived AI video model, built on Wan 2.2 and aimed at faster video generation; it is pursuing enterprise use with AWS.
Why it matters: Apple is buying AI infrastructure talent as it builds out its own agentic tooling.
Apple acquired Ascend.io, a startup building agentic AI engineering and data pipeline automation tools. AGIBOT separately introduced its Worken framework for measuring robot productivity.
Modellix.ai, Aurora Mobile's unified AI model access service, launched a ComfyUI integration that lets creators use multiple AI media generation models inside ComfyUI creative workflows.
GitHub added new AI agent features to Copilot in VS Code and launched Copilot Fusion, with agents now starting about 40% of pull requests, according to reports.
Tagbin showed BRIXX at TechSparks 2026, a platform where natural-language prompts handle architectural design, construction planning and cost estimation together.
Why it matters: Investor reaction to AI capex shows markets questioning whether spending will pay off.
Tech stocks fell after Amazon reported higher-than-expected capital spending and its slowest core e-commerce growth in three years. Nasdaq faced a $2 trillion October wipeout as AI spending fears deepened; Amazon tapped an $80B supplier loan facility.
TCS launched its Industrial Intelligence Foundry, a platform aimed at building agentic AI value chains for industrial clients while keeping AI use aligned with organizational rules, per diginomica.
Salesforce plans marketing and e-commerce Einstein Copilots, while publishing conversion numbers for agent-native shopping search that cover one specific thing.
Dayforce announced a new agentic experience and multiple agentic capabilities for its HR platform at its annual user conference, raising questions about whether customers are ready.
Percepto launched Axion, a hardware-agnostic physical AI platform that unifies inspection data analysis and gives visibility across mission-critical infrastructure.
Why it matters: Arm selling full server CPUs would reshape the datacenter chip market long dominated by Nvidia and Intel.
Arm is expanding from licensing Neoverse core IP to selling finished server CPUs aimed at AGI workloads. Meta is co-developer and primary customer; OpenAI also plans to adopt the products.
SwitchBot's K11+ Pro is a compact robot vacuum with a small dock, strong suction and Matter support for Apple Home, aimed at homes where full-size vacuums take too much space.
Misumi Americas started the Misumi Imprint Foundation, a $2 million corporate giving organization, its first in 63 years, to fund engineering education, robotics programs and manufacturing skills in the US.
Samsung added an FDA-approved sleep apnea detection feature to Galaxy Watches, arriving first on the Galaxy Watch Ultra 2, Watch 9 and Galaxy Ring, ZDNET reports.
PsiQuantum built an 8x8 photonic switch that loses 1.2 dB, under a quarter of its light, versus the 3 dB data centers allow; larger devices are planned for AI GPU networks.
A new UEFI-level pre-driver GPU unlocker is available for Nvidia RTX 20xx, 30xx and 40xx cards on Windows and Linux, lifting the Tensor core cap on consumer GPUs.
Why it matters: Bringing DGX Spark silicon to laptops makes local AI compute portable for developers.
Nvidia's RTX Spark laptops use the same chip as the DGX Spark desktop but add faster memory and features the DGX Spark lacks; Nvidia also convenes an Open Secure AI Alliance citing a forensics gap in closed models.
Why it matters: One of the first wired AR glasses built on Google's Android XR with Gemini.
XREAL's AURA are its first wired augmented reality glasses running Google's Android XR, with transparent Micro-OLED displays, 70-degree field of view, Gemini AI and a Snapdragon Reality Elite chip.
Why it matters: Sub-$1,500 humanoid pricing would push legged robots toward mass-market volumes.
Noetix is selling a humanoid robot for RMB9,998, an unusually low price that reports frame as just the beginning of a push to put humanoids in homes and businesses at consumer prices.
Why it matters: Foxconn is a key AI server builder, and its partnerships signal how fast datacenter capacity is being assembled.
Foxconn chairman Young Liu pitched vertical integration beyond box moving at Hon Hai Tech Day, with Google aiming to cut AI infrastructure deployment to six months and a new OpenAI tie-up in the works.
Beijing startup SynDexy shipped its first small batch of C1 humanoids for precision assembly at electronics plants, six months after development began; key tasks exceeded 98% success in tests.
ISHO, founded by former Xiaomi executive Wang Teng, unveiled the ISHO Ring smart ring and iFlow bed temperature-control system. Prices start at RMB1,799 and RMB9,999; the ring tracks physiological signals for sleep.
ACM Shanghai shipped its first 310x310mm panel-level plating tool as AI chip packages grow larger and panel-level packaging moves toward customer validation and volume production. Order backlog rose 88%.
A Y Combinator-backed US aerospace startup built and flew an autonomous cargo aircraft in under two months, targeting 1,000-mile freight routes. No payload, timeline or funding details were given.
Why it matters: Debt-financed chip purchases show how AI buildouts are increasingly leveraged.
Indian AI infrastructure firm AM Intelligence is seeking up to $3B in financing for a previously announced purchase of Nvidia's next-generation chips, per Bloomberg.
Hyundai's Grandeur receives a first OTA software update adding Pleos Connect and evolving Gleo AI voice search that reads driver intent, while Hyundai Mobis reported Q3 operating profit and began robot actuator prototype production.
42dot released the first over-the-air update for its Gleo AI service, which suggests places along a driving route and shows gas station prices by fuel type.
Why it matters: Shows how much AI demand is reshaping the global semiconductor supply chain.
Taiwan's TSIA expects member output above NT$9.1 trillion ($279.5B) in 2026, up 40.9% on AI chip demand, while urging stable power. Nvidia holds 69% of AI robot chips, DigiTimes reports.
PokeBot released RoboAPI, letting developers test natural-language control on robot arms. Founder Huazhe Xu says the Poke-Aurora wrist-camera setup is designed to transfer across several arm configurations.
Ramazan Eren Arslan converted a Deutsche Bundespost FeTAp 791 rotary phone into a voice assistant: lift the handset and talk, and Google Gemini answers through the original earpiece, powered by a Raspberry Pi 3B.
ZDNET reports Android 17 adoption has hit just 1%, with only a quarter of Android phones on Android 16, urging users to update for security and features.
Microsoft Research and UC Berkeley released TokenEconomizer, an open-source plugin that learns from past sessions to cut coding agent token costs 18.4% to 30.9% on SWE-Bench without changing the agent.
AWS describes a controlled local setup that routes the ChatGPT desktop app's model inference to Amazon Bedrock in AWS GovCloud (US-West) using API-mode sign-in, device policy and network restrictions.
DevFest 2026 runs October 1 to December 31 across more than 800 global events under the theme "Build, Secure, Scale," with hands-on sessions on Gemini and Google's AI developer stack.
Daniel Catt built a tool that converts markdown to handwriting, driving a Bantam Tools NextDraw 8511 pen plotter via his AI assistant Kitty to write todo lists, addresses and packing lists.
The Guardian reports ghost hunters are using AI chatbots and custom apps to interpret signals, raising concerns among paranormal enthusiasts about the technology's role.
Local Code is an open-source AI assistant for Mac that lets users download models and run them locally without separately installing dependencies. The Draw Things team built it as a free, private tool.
VoidMob is promoting an open-source MCP (Model Context Protocol) server that lets AI agents connect to and use its platform. No new version or release details were provided beyond the integration capabilities.
Mistral shipped Vibe CLI 3.0.4, a minimal command-line coding agent installable via pip. The Deep View separately examines how Mistral tests China's open-model advantage.
GitHub added new AI agent features to Copilot in VS Code and launched Copilot Fusion, with agents now starting about 40% of pull requests, according to reports.
Tagbin showed BRIXX at TechSparks 2026, a platform where natural-language prompts handle architectural design, construction planning and cost estimation together.
Google rolled out a large Google Home update with 42 fixes for the Home app, cameras and speakers, addressing problems with Gemini and Matter devices and deepening Gemini's room awareness.
ETtech looks at how AI lets Indian developers build prototypes in minutes, and how platforms like GoDaddy's Developer Platform handle the code-to-deployment half.
Salesforce plans marketing and e-commerce Einstein Copilots, while publishing conversion numbers for agent-native shopping search that cover one specific thing.
A 77-class banking text-classification model continued with sharpness-aware minimization, seed 13, from a full-train checkpoint; author reports 94.12% frozen test accuracy. Apache-2.0 weights on Hugging Face.
XWiki MCP Server is an open-source, self-hostable extension connecting XWiki to AI agents like Claude Code, Codex and Copilot to query knowledge, create and edit pages and work with structured data.
fastbrowse is an open-source browser agent that reads page buttons, links and fields, picks the next step with a fast model and falls back to a standard model when unsure, citing answers.
Why it matters: A design-native VLM that edits layouts directly could make AI design tools much cheaper and faster.
Canva Brush is a vision-language model from Canva Research that sees a design and makes requested edits, from one line of copy to every slide, staying fully editable; Canva says 69% faster and 46% cheaper per task.
Teable Mobile lets users describe a process by voice and have Teable build the database, app and automations, then publish to a team and get notified when done.
A 77-class banking text-classification model continued with sharpness-aware minimization, seed 13, from a full-train checkpoint; author reports 94.12% frozen test accuracy. Apache-2.0 weights on Hugging Face.
A Gradio demo app for generating images with an uncensored Qwen-Image 2.1 model, plus on-demand all-in-one LoRA adapters, running on Hugging Face ZeroGPU.
Idol turns a single photo, render or generated image into a 3D Gaussian splat, removing the background and writing up to 1,048,576 Gaussians as a standard .ply. It runs on consumer GPUs with 7.2 GB VRAM.
MiniMax H3, a video generation model, is drawing community work: a Pika Labs H3 finetune (Pika-H3-Flex) on Hugging Face, an Nvidia open-source H3-based enhancement tool upscaling to 4K, and a MiniMax M3.1-Flash coding/agent model mention.
A GGUF build of LiquidAI's 1.7B small language model with safety alignment ablated via a baked-in rank-1 LoRA, for running locally and offline in the Atomic Chat app.
Model · text-generation · other · ♥ 103 · 2.3k downloads
A research paper presents an improved distribution matching distillation method for one-step and few-step text-to-image generation, estimating score differences with internal velocity branches to better render prompt details.
A paper presents L3G-World, an interactive world model that uses latent 3D Gaussian memory to keep spatial consistency over long exploration in real time without memory growing with stream length.
A research paper introduces Read, Allocate, Route, a role-coupled tokenization scheme for pixel-space generative Transformers that models low-frequency structure globally and high-frequency detail locally to cut attention cost.
An omni-modal parser that turns documents, images, audio and video into structured records grounded in space and time, giving agents a perception layer to read and cite evidence. Uses one encoder and one schema.
A new paper turns the Platonic Representation Hypothesis into a theorem, proving that Joint-Embedding Predictive Architectures converge toward a shared representation of the world as they scale, and proposes improvements.
Agent Office 3D is a low-poly 3D office that visualises your AI coding agents' live status, with idle agents sneaking off to play Minecraft without using tokens. It drew 102 stars its first week.
Mika is an on-device AI assistant for Android that operates apps and the web to complete tasks for you. The new Kotlin project drew 107 stars in its first week.
DeepSeek Harness v0.2.2-alpha.2 adds an experimental subagent tool that sends self-contained tasks to a locally installed Codex CLI or Claude Code, plus PR views and a stats plugin.
An open-source MIT project that has Claude Code, Codex or other agents threat-model your repository, producing an evidence-traced threat model using 36 architectures and 29 methods.
A self-hosted open-source Kotlin project that has AI agents ring your phone like a real call, even when locked, talks through a voice AI and returns the transcript, with no Twilio needed.
An open-source Rust file launcher that searches inside documents, code and screenshots locally, using OCR, semantic search and MCP agent access. It drew 101 stars in its first week.
A GitHub agent skill that examines a product you are authorised to rebuild, builds a clone, then tests it against the original and fixes gaps until a checklist passes. MIT licensed, about 100 stars its first week.
37studio is a self-hosted open-source alternative to Buffer and Opus Clip: Claude Code makes on-brand videos, posts them to nine platforms and drafts replies, using your own API keys.
ZDNET reports Android 17 adoption has hit just 1%, with only a quarter of Android phones on Android 16, urging users to update for security and features.
The Guardian reports ghost hunters are using AI chatbots and custom apps to interpret signals, raising concerns among paranormal enthusiasts about the technology's role.
Google rolled out a large Google Home update with 42 fixes for the Home app, cameras and speakers, addressing problems with Gemini and Matter devices and deepening Gemini's room awareness.
Salesforce plans marketing and e-commerce Einstein Copilots, while publishing conversion numbers for agent-native shopping search that cover one specific thing.
Canva Brush is a vision-language model from Canva Research that sees a design and makes requested edits, from one line of copy to every slide, staying fully editable; Canva says 69% faster and 46% cheaper per task.
Teable Mobile lets users describe a process by voice and have Teable build the database, app and automations, then publish to a team and get notified when done.
SparkBots AI runs trainable bots on web pages and PDFs, returning key details for leads, contracts or resumes in under a second; the maker says they cannot hallucinate details.
Plaud released version 1.3.12 of its AI note-taking software for macOS, Windows and Linux, which records and summarizes online meetings, phone calls and in-person conversations. A separate review compared Pocket AI and Plaud Note Pro recorders.
Claude now works via CarPlay, letting drivers ask the AI by voice, and Tesla's Robotaxi service extended hours to 2 a.m. and added Grok to the Model Y.
Package 26.1027.3384.0 for ChatGPT on Windows contains partial quality and return-visit report interfaces, but the inspected business-profile route passes enabled=false.
Tencent's AI agent platform WorkBuddy launched remote control, letting users connect to a logged-in computer from the mobile app, view the desktop live and take over tasks. Tencent also opened its Beacon AI data tool to the public.
APK teardowns and reports show Google testing Gemini Live offline mode on Android, a sparkling overlay redesign, playback-speed controls for read-aloud responses, and direct file editing in chat.
Microsoft Research and UC Berkeley released TokenEconomizer, an open-source plugin that learns from past sessions to cut coding agent token costs 18.4% to 30.9% on SWE-Bench without changing the agent.
AWS describes a controlled local setup that routes the ChatGPT desktop app's model inference to Amazon Bedrock in AWS GovCloud (US-West) using API-mode sign-in, device policy and network restrictions.
DevFest 2026 runs October 1 to December 31 across more than 800 global events under the theme "Build, Secure, Scale," with hands-on sessions on Gemini and Google's AI developer stack.
Daniel Catt built a tool that converts markdown to handwriting, driving a Bantam Tools NextDraw 8511 pen plotter via his AI assistant Kitty to write todo lists, addresses and packing lists.
Local Code is an open-source AI assistant for Mac that lets users download models and run them locally without separately installing dependencies. The Draw Things team built it as a free, private tool.
VoidMob is promoting an open-source MCP (Model Context Protocol) server that lets AI agents connect to and use its platform. No new version or release details were provided beyond the integration capabilities.
Mistral shipped Vibe CLI 3.0.4, a minimal command-line coding agent installable via pip. The Deep View separately examines how Mistral tests China's open-model advantage.
GitHub added new AI agent features to Copilot in VS Code and launched Copilot Fusion, with agents now starting about 40% of pull requests, according to reports.
Tagbin showed BRIXX at TechSparks 2026, a platform where natural-language prompts handle architectural design, construction planning and cost estimation together.
ETtech looks at how AI lets Indian developers build prototypes in minutes, and how platforms like GoDaddy's Developer Platform handle the code-to-deployment half.
XWiki MCP Server is an open-source, self-hostable extension connecting XWiki to AI agents like Claude Code, Codex and Copilot to query knowledge, create and edit pages and work with structured data.
fastbrowse is an open-source browser agent that reads page buttons, links and fields, picks the next step with a fast model and falls back to a standard model when unsure, citing answers.
Faylee records screen, console, network requests and clicks behind a bug on one timeline and shares it as a single link. Pasting the link into Claude Code, Codex or Cursor lets its MCP server feed the report to the agent to fix.
Oracle introduced AI agent-ready setup and configuration for its Fusion Applications, aiming to speed implementations by letting agents handle configuration work.
A project uses Claude to put global law into Git, making it semantic, versioned, attributable and queryable by agents, released under CC0. It hit No. 1 on Hacker News.
TechTarget outlines ways teams can roll out AI model updates and version changes while keeping production systems stable, covering testing, monitoring and fallback strategies.
Gemini Code Assist 2.107.0 is the latest release of Google's AI coding assistant extension for VS Code, powered by Gemini models for AI-assisted development.