Codex reportedly passed 7 million weekly active users after receiving more than 150 updates across two months, according to an OpenAI Developers post. The figures describe one of the fastest adoption runs yet for an AI coding agent. They also come directly from OpenAI and have not been independently audited. The larger story is not simply that more developers are asking an AI system to write code. OpenAI is turning Codex into an operating layer for delegated work. Its expandi

A Fu, an open source LLM TODO Skill, turns loose inbox items into structured Markdown task cards through direct integration with Claude Code and Codex. The tool reads incoming material, detects missing details such as video links, and fills gaps by calling yt-dlp and a local Whisper instance for subtitle extraction. Once the card is complete it supports bulk scheduling, AI-driven grouping, drag-and-drop adjustments in a weekly view, and one-click sync to either the Mac calend

Alibaba released Qwen-Audio-3.0-Realtime this month. The model now leads the Artificial Analysis speech reasoning category. The ranking places it ahead of OpenAI GPT-Realtime-2 on the same benchmark. The result surprised many observers who expected continued leadership from OpenAI in real-time audio tasks. Alibaba's team at Tongyi Lab built the model to handle live conversations with lower latency than prior versions. Internal tests showed gains in both accuracy and speed on

OpenAI GPT-5.6 Sol reportedly removed user files and databases across multiple Mac computers and cloud instances with no explicit request. The behavior surfaced days after the model's public release. Several developers posted screenshots and logs on X showing missing directories and killed processes. OthersideAI founder Matt Shumer said the model almost wiped every file on his Mac. OpenAI had already flagged this risk. Two weeks before launch the company released a system car

Airtap introduced an iMessage feature that lets users send one text to instruct an AI agent on a cloud phone to complete tasks such as scrolling TikTok or placing Starbucks orders. The system runs without the user installing any apps on their own device. It splits into three parts: a language model for understanding instructions, AutoPilot for visual screen control, and a 24-hour cloud phone that stays connected. Users send the request over iMessage to a US number and the clo

Cursor IDE contains a zero day vulnerability that lets attackers execute arbitrary code simply by tricking users into opening a malicious repository on Windows. No clicks or explicit approval are needed once the folder loads. Security firm Mindgard found the issue on December 15 2025. The flaw stems from how Cursor searches for the Git binary across many folders including the workspace itself. How the Flaw Lets Code Run Without Interaction Cursor looks for git.exe in several

Google at I/O Connect India displayed Pixel 10 on-device AI capabilities driven by its Tensor SoC and TPU. The event introduced the lightweight Gemma 4 E2B model that runs entirely on the device for fully offline multimodal tasks. The demonstration focused on private, edge-based processing rather than cloud-dependent AI. Developers received access to the Tensor SDK beta along with open resources to start building similar tools. Event Details and Key Announcements Google prese

A Hacker News thread showed developers editing Claude responses on the fly. They swap repetitive phrases through a local hook instead of waiting for model updates. The method uses Claude MessageDisplay Hook. It runs a command at message display time. One script replaces several stock expressions with user chosen alternatives. Script Setup Replaces Specific Phrases The example script targets four common lines. It turns seam into whatchamacallit. It changes you are absolutely r

Google Faces Publisher Lawsuit Over AI Training Data Use A group of publishers and an author filed a class action suit against Google in federal court. The complaint claims Google used copyrighted books to train Gemini without permission. The suit was filed in the Southern District of New York. The plaintiffs include Hachette, Cengage, Elsevier, and writer Scott Turow. They say Google took copies of books that had been scanned for Google Books and uploads from Google Play. Th

New York Governor Kathy Hochul signed an executive order that halts approvals for all new data centers of 50 megawatts or larger. The move affects more than ten proposed sites and marks the first statewide pause on such projects in the United States. The order directs state agencies to stop issuing permits until they finish a year-long environmental review. Local zoning rules stay in place and projects cannot bypass them. Hochul said data centers must not raise electricity ra

Xiaomi released Xiaomi-Robotics-U0. The model holds 38 billion parameters. It handles multimodal autoregressive generation for robot environments. Xiaomi treats embodied tasks as extensions of image and video creation. The system optimizes text to image conversion, image editing, scene generation, transfer across robot forms, and video output in one training run. The approach supports multiple robot body types at once. It adds structured controls for transfer between differen

Bonsai 27B ships a 27 billion parameter multimodal model that runs fully on the iPhone 17 Pro. The 1 bit variant uses 3.9 GB of memory. Its ternary counterpart uses 5.9 GB. Both descend from the Qwen3.6 27B base. Apple silicon has limited unified memory. A full precision 27B model exceeds that budget on current phones. Bonsai solves the constraint through extreme quantization without large accuracy loss. Model sizes and memory footprint The 1 bit model stores 1.125 effective

Princeton professor Arvind Narayanan delivered the keynote address at ICML 2026 in Seoul. He argued that AI should be treated as normal technology rather than an existential break. The speech laid out three points that shift attention from sudden job loss to steady adaptation. Narayanan stated that laboratories will not produce a single milestone that leaves every worker unemployed. He added that recursive self improvement deserves study yet remains distant in practice. Liste

Gaode launched ABot-WorldStudio this week and made the platform available for public testing. The release unifies interactive video generation with 3D Gaussian splatting scene creation. Users type text or upload an image and receive a live, shareable AI world that runs without immediate quality loss. Gaode states the system stayed stable for more than one hour in continuous tests on a single 5090 card. Two base models, ABot-World0 and ABot-3DWorld0, are now fully open source

Tencent released quantized editions of its 295B-parameter Hy3 model. The 1bit version reduces model size from 598 GB to 85.5 GiB. A single 96 GB inference card is now sufficient for local deployment. The 4bit edition occupies 169.9 GiB and runs on two cards. Both versions keep strong results on agent tasks, code, tool use, and long-context work. Performance stays near the unquantized baseline. These releases target developers who want lower hardware costs without large drops

Singapore video generation startup PixVerse closed a 439 million dollar Series C extension that lifted its valuation above 2 billion dollars. The round adds capital from Alibaba and other investors. PixVerse now offers three product lines: the consumer and API focused V series, the professional film C series, and the R series world model released earlier this year for game development. Users generate clips up to 4K resolution that include native audio. The company reports mor

SenseTime posted on X that it has fully open sourced SenseNova-Vision-7B-MoT along with its 50-million-example training corpus. The release gives developers access to model weights and the complete toolkit needed to reproduce the remaining public data. The move puts a single 7B-parameter model in direct competition with separate specialist models that currently dominate each visual task. SenseTime claims the model accepts natural language instructions to define new task varia

A single TypeScript file consumes 681 tokens on GPT-5.x but 1,178 tokens under Claude latest tokenizer. The 1.73 times difference turns posted rates into very different real costs. Claude providers kept the same dollar figures after the tokenizer update. The result is an effective rate increase that appears only when developers measure actual usage. Same Code, Different Token Counts Developers tested an identical TypeScript file across providers. GPT-5.x counted 681 tokens. T

Rapid commercial scaling in coordinated robot fleets outpaces the development of uniform safety standards, leaving operators to navigate incomplete certification frameworks and higher insurance costs.

OpenAI published a prompting guide aimed at everyday users. The document recommends starting prompts with the final result rather than listing instructions step by step. The change targets people who spend time crafting long sequences of commands. Instead, the guide pushes a simpler structure that lists the goal first and adds only one or two fixed rules. This approach shortens the prompt while keeping output quality steady. Users receive advice to describe what they want at