Anthropic's Claude Cowork Was Built in Under Two Weeks Using Claude Code to Write the Code
Anthropic unveiled Claude Cowork, an AI‑powered collaboration app built in roughly one and a half weeks. The majority of its code was generated by Claude Code, the company’s own code‑generation model, according to inventor Boris Cherny. Product manager Felix Rieseberg said the sprint‑and‑a‑half effort leveraged earlier prototypes and focused on front‑end development. The launch is a research preview, meaning the product still has rough edges but demonstrates rapid AI‑driven development.
Chinese Researchers Diagnose AI Image Models with Aphasia-Like Disorder, Develop Self-Healing Framework
Chinese researchers introduced UniCorn, a self‑healing framework that teaches multimodal AI models to recognize and correct mismatches between image understanding and generation, a problem they label “Conduction Aphasia.” The system splits a single model into Proposer, Solver, and Judge agents...
Lightricks Open-Sources AI Video Model LTX-2, Challenges Sora and Veo
Lightricks has open‑sourced its 19‑billion‑parameter LTX‑2 model, which generates synchronized stereo audio and video from a single text prompt. The dual‑stream transformer allocates 14 B parameters to video and 5 B to audio, enabling 4K, 50 fps output up to 20 seconds long. Benchmarks...
Chinese AI Industry Admits US Remains Ahead for Now
Chinese AI leaders acknowledge the United States remains ahead in compute power and chip access, limiting near‑term chances of overtaking OpenAI or Anthropic. Executives at the AGI Next Summit cited export controls and a two‑order‑of‑magnitude gap in infrastructure as primary...
OpenAI Reportedly Sets Aside $50 Billion for Employee Stock Program
OpenAI has earmarked a $50 billion employee stock pool, roughly ten percent of its $500 billion valuation, according to The Information. Combined with $80 billion in previously allocated shares, staff now own about 26 percent of the company. The firm is also in early...
New Deepseek Technique Balances Signal Flow and Learning Capacity in Large AI Models
Researchers introduced mathematically constrained hyper‑connections (mHC) to address signal amplification in large DeepSeek models. Standard hyper‑connections cause signal strength to explode up to 3,000×, leading to loss spikes around 12,000 steps in a 27‑billion‑parameter model. By applying Sinkhorn‑Knopp normalized, non‑negative...
Global AI Compute Hits 15 Million H100 Equivalents, Epoch AI Finds
Epoch AI’s new database shows global AI compute capacity has topped 15 million Nvidia H100 equivalents, a milestone that underscores rapid scaling of machine‑learning workloads. The analysis reveals Nvidia’s latest B300 chip now accounts for the bulk of the company’s AI...
Minimax Stock Doubles on Hong Kong Debut
Chinese AI startup Minimax saw its shares double on the Hong Kong Stock Exchange debut, closing up 109% at HK$345. The IPO generated roughly $620 million, with backing from Alibaba and Tencent. Minimax, which powers chatbots and video‑generation models, reported $53.4 million...
Microsoft and Stripe Bring Shopping Checkout Directly Into Copilot Chat for US Users
Microsoft and Stripe have announced Copilot Checkout, allowing US users to purchase items directly within Microsoft’s AI‑powered Copilot chat. The feature, built on Stripe’s open Agentic Commerce Protocol, launches with retailers such as Etsy, Urban Outfitters and Anthropologie. By integrating...
Arm Holdings Establishes New Business Unit for Robotics and Automotive
Arm Holdings announced a restructuring that creates a new business unit called Physical AI, aimed at the robotics and automotive sectors. The company will operate three divisions—Cloud and AI, Edge, and Physical AI—under a unified strategy. Drew Henry has been appointed...
Tailwind's Shattered Business Model Is a Grim Warning for Every Business Relying on Site Visits in the AI Era
Tailwind Labs, creator of the wildly popular Tailwind CSS framework, announced that it has laid off three of its four engineers after revenue plunged about 80 percent. The decline stems from a 40 percent drop in documentation traffic as AI...
Despite Trump's Approval, China Slows Nvidia Chip Imports to Protect Domestic Industry
Chinese authorities have instructed domestic tech firms to suspend purchases of Nvidia's H200 AI chips while the government reviews import conditions. The directive follows President Donald Trump's recent approval for Nvidia to sell H200s to China, but Beijing is wary...
Google's "Nano Banana" Owes Its Odd Name to a Project Manager Working Alone at 2:30 A.m.
Google’s image‑generation model “Nano Banana” has surged to the top of the LM Arena benchmark, outpacing rivals like OpenAI’s DALL‑E 3. The name originated from project manager Naina Raisinghani’s 2:30 a.m. mash‑up of two personal nicknames, a quirky decision that quickly went viral....
New Artificial Analysis Benchmark Shows OpenAI, Anthropic, and Google Locked in a Three-Way Tie at the Top
Artificial Analysis released its Intelligence Index v4.0, ranking large language models across four equally weighted categories: Agents, Programming, Scientific Reasoning, and General. OpenAI’s GPT‑5.2 (xhigh) claimed the top spot with 50 points, followed closely by Anthropic’s Claude Opus 4.5 at 49...
AI Industry Finds Its 2026 Narrative as OpenAI and Microsoft Argue Users Are the Bottleneck, Not Models
The AI sector has embraced a new narrative for 2026: models now exceed the capabilities of most users. OpenAI’s product head Fidji Simo emphasized that the gap between model potential and everyday use must be closed. With 800 million weekly active...
OpenAI Loses Top AI Researcher Jerry Tworek After Seven Years
OpenAI announced that senior researcher Jerry Tworek is leaving after nearly seven years. Tworek was instrumental in building GPT‑4, ChatGPT, and the company’s first AI coding models, and most recently led the Reasoning Models team focused on complex logical tasks....
Only 5 Percent of ChatGPT's 900 Million Weekly Users Pay, and Reportedly Most Aren't Worth Much to Advertisers
OpenAI’s ChatGPT sees roughly 900 million weekly users, but almost 90 percent are outside the United States and Canada. Only about five percent of those users subscribe to paid plans, with many in emerging markets on the $5‑per‑month “ChatGPT Go” tier. International users...
Over Five Percent of ChatGPT Messages Worldwide Are About Health
OpenAI reports that over five percent of global ChatGPT traffic now centers on health topics. In the United States, 40 million users consult the chatbot daily for medical advice, ranging from symptom checks to insurance comparisons. The platform logs nearly two...
AI Tool Catches Pancreatic Cancer in Routine Scans Before Symptoms Appear
Alibaba's Damo Academy unveiled PANDA, an AI system that spots pancreatic tumors in non‑contrast CT scans—images that often miss cancers even for seasoned radiologists. In a clinical trial at Ningbo University Hospital, the tool reviewed 180,000 scans, flagging two dozen...
Anthropic President Daniela Amodei Says "the Exponential Continues Until It Doesn't"
Anthropic President Daniela Amodei told CNBC that AI’s rapid, exponential progress has repeatedly outpaced expectations, but she cautioned it may not continue indefinitely. She highlighted that even if model capabilities keep improving, corporate rollout faces “human reasons” such as lengthy...
Local Resistance Blocks $98 Billion in AI Data Center Projects Across Eleven US States
Tech firms planning AI data centers encountered significant pushback, with 20 projects valued at $98 billion blocked or delayed across eleven U.S. states between April and June. Residents cite higher electricity bills, water use, noise, and loss of farmland as primary...
The "Curse of Knowledge" Means Smarter AI Models Don't Understand Where Human Learners Struggle
A recent study highlights a fundamental blind spot in large language models: while they can solve exam questions that stump humans, they cannot identify why those questions are difficult. This "curse of knowledge" bias means AI systems lack insight into...
ByteDance's StoryMem Gives AI Video Models a Memory so Characters Stop Shapeshifting Between Scenes
ByteDance and Nanyang Technological University introduced StoryMem, a system that stores selectively chosen key frames in a memory bank to keep AI‑generated videos visually consistent across multiple scenes. By feeding these reference frames into a LoRA‑adapted Wan2.2‑I2V model, the approach...
"You Certainly Don't Tell a Researcher Like Me What to Do" Says LeCun as He Exits Meta for His Own...
Yann LeCun, Meta's chief AI scientist, is leaving the company to launch Advanced Machine Intelligence Labs, a startup focused on video‑based world models. He disclosed that Meta’s Llama 4 benchmarks were deliberately skewed, prompting Mark Zuckerberg to lose confidence and sideline...
Google Engineer Says Claude Code Built in One Hour What Her Team Spent a Year On
A senior Google engineer reported that Anthropic's Claude Code generated a working distributed agent orchestrator in just one hour, matching a system her team has been developing for a year. The prompt she provided was only three brief paragraphs, and...
OpenAI Co-Founder Greg Brockman Donates $25 Million to Trump's MAGA Inc. Super PAC
OpenAI co‑founder Greg Brockman donated $25 million to Donald Trump’s MAGA Inc. super PAC, part of a $102 million fundraising wave that lifted the PAC’s war chest to $294 million by year‑end 2025. The donation places Brockman alongside crypto exchange Crypto.com ($20 million) and...
Grok's Image Editing Tool Generated Sexualized Images of Children, Forcing xAI to Acknowledge Safety Gaps
xAI’s Grok AI image‑editing tool allowed users to generate sexualized images of children. The feature lets anyone replace clothing with bikinis or lingerie via a simple text prompt, and a user demonstrated child‑targeted outputs. xAI acknowledged “lapses in safeguards,” labeled...
Baidu's In-House Chip Unit Kunlunxin Joins Wave of Chinese AI Firms Heading for Hong Kong IPO
Chinese AI chip maker Kunlunxin, Baidu's in‑house semiconductor unit, has confidentially filed for a Hong Kong initial public offering, with a recent financing round valuing it at roughly $3 billion. The filing, submitted on Jan 1, signals a spin‑off while Baidu retains...
Science Context Protocol Aims to Let AI Agents Collaborate Across Labs and Institutions Worldwide
The Shanghai Artificial Intelligence Laboratory has released the open‑source Science Context Protocol (SCP), extending Anthropic’s Model Context Protocol to address the needs of scientific workflows. SCP adds structured experiment metadata, a centralized hub for task orchestration, and standardized drivers for...
OpenAI Merges Internal Teams to Fix Audio AI Accuracy Gap Ahead of ChatGPT Hardware Push
OpenAI has consolidated several internal audio teams to accelerate the development of next‑generation speech models, aiming to narrow the accuracy and latency gap with its text‑based systems. A new architecture under construction promises more natural, emotional, and real‑time conversational capabilities,...
LLMs Could Serve as World Models for Training AI Agents, Study Finds
A multi‑institution study demonstrates that large language models can be fine‑tuned to act as world models, accurately simulating environment dynamics for AI agents. In structured text‑based tasks such as ALFWorld and SciWorld, models like Qwen2.5‑7B and Llama‑3.1‑8B achieved over 99%...
Moonshot AI Closes $500 Million Series C to Fund Kimi-K3 Development and Expand Computing Capacity
Moonshot AI, the Chinese startup behind the Kimi chatbot, closed a $500 million Series C round, valuing the company at $4.3 billion. The financing, led by IDG with participation from Alibaba, Tencent and Wang Huiwen, brings cash reserves above $1.4 billion. The capital will fund...
AI Took a Writer's Job, Then Told Him to Try Tree-Felling
Brian Groh, a 52‑year‑old Indiana copywriter, lost his job after his company first outsourced to cheaper contractors and then replaced his work with AI‑generated copy. Seeking guidance, he asked a chatbot which advised him to try tree‑felling, a physically demanding...
Alibaba's New Open Qwen Image Model Aims for More Natural-Looking Results
Alibaba unveiled Qwen-Image-2512, an upgraded text‑to‑image model that delivers more natural‑looking human faces, sharper textures and clearer embedded text. The model was evaluated in over 10,000 blind tests on Alibaba’s AI Arena, finishing fourth overall and topping the open‑source leaderboard....
OpenAI's Stock Compensation Averages $1.5 Million per Employee, Dwarfing Every Tech Startup in History
OpenAI is offering stock‑based compensation averaging $1.5 million per employee, roughly 34 times the pre‑IPO grants typical at major tech firms. The payout represents 46.2% of the company’s projected 2025 revenue, dwarfing rivals such as Google (14.6%) and Meta (5.9%). OpenAI...
Nvidia Plans H200 Production Ramp at TSMC While China Debates Whether to Let the Chips In
Nvidia is negotiating with TSMC to ramp up production of its H200 AI accelerator, targeting a second‑quarter 2026 start. Chinese customers have placed orders for more than two million units, yet Nvidia currently can supply only 700,000 chips. The chips, priced...
Scientific AI Models Trained on Different Data Are Learning the Same Internal Picture of Matter, Study Finds
MIT researchers examined 59 scientific AI models across chemistry, materials, and protein domains and found they develop remarkably similar internal representations despite differing architectures and input formats. The study shows that higher‑performing models align more closely with a reference “UMA...
Security Researchers Catch "Privacy" Browser Extensions Siphoning AI Chats and Selling Them via a Data Broker
Security researchers at Koi identified eight popular browser extensions, including Urban VPN Proxy, that covertly capture AI chat conversations from services like ChatGPT, Claude, and Gemini. The data‑harvesting code was introduced via an automatic update in July 2025 and continues even...
Zara Uses AI to Dress Models Virtually Instead of Booking New Photo Shoots
Zara is deploying artificial‑intelligence tools to digitally re‑dress existing model photographs, eliminating the need for fresh photo shoots. The retailer asks models for permission to reuse images and compensates them at the same rate as a traditional shoot, while other...
AI Fraud Forces World's Largest Accounting Body to End Online Exams
The Association of Chartered Certified Accountants (ACCA) will discontinue all online examinations starting March 2026 due to AI‑driven cheating outpacing detection methods. The body, serving nearly 260,000 members and over 500,000 students, cites incidents where candidates photograph questions and feed...
Oscar Winners and Hollywood A-Listers Launch Coalition to Set AI Rules for Entertainment Industry
Hollywood’s top talent has formed the Creators Coalition on AI (CCAI) to establish industry‑wide standards for artificial intelligence use. Oscar‑winning directors Daniel Kwan and Jonathan Wang, together with actor Joseph Gordon‑Levitt, launched the group, with supporters including Natalie Portman, Cate...
Microsoft CEO Nadella Tells Managers Copilot's Gmail and Outlook Integrations ‘Don't Really Work’ and Steps in to Fix Them
Satya Nadella publicly criticized Microsoft Copilot’s Gmail and Outlook integrations as “not smart,” and has taken a hands‑on role to fix them. He now leads weekly technical reviews with senior engineers, sends bug reports, and issues direct instructions to improve...
OpenAI Seeks New "Head of Preparedness" For AI Risks Like Cyberattacks and Mental Health
OpenAI announced a new senior position, Head of Preparedness, to oversee safety risks associated with its AI models. The role will address threats ranging from cyber‑attack facilitation and mental‑health impacts to the dissemination of biological knowledge and self‑improving systems. It...
Resemble AI Drops Chatterbox Turbo, an Open-Source Text-to-Speech Model that Clones Voices in Five Seconds
Resemble AI has launched Chatterbox Turbo, an open‑source text‑to‑speech model that can clone a voice from just five seconds of audio. The model claims superior quality to Elevenlabs and Cartesia while producing the first audio chunk in under 150 ms, making...
Meta's Pixio Proves Simple Pixel Reconstruction Can Beat Complex Vision Models
Meta AI unveiled Pixio, a masked autoencoder that learns solely through pixel reconstruction. By enlarging masked regions, adding multiple class tokens, and deepening the decoder, Pixio achieves superior depth estimation and 3D reconstruction despite having only 631 million parameters. In head‑to‑head...
Meta Brings Segment Anything to Audio, Letting Editors Pull Sounds From Video with a Click or Text Prompt
Meta has extended its Segment Anything (SAM) technology to audio, launching SAM Audio—a model that isolates individual sound sources using text prompts, visual clicks, or time‑based markers. The system leverages a flow‑matching diffusion transformer and a new Perception Encoder Audiovisual...
Australia's Financial Regulator Warns Banks Against Flooding It with AI-Generated Suspicious Activity Reports
Australia’s financial regulator AUSTRAC has cautioned banks against over‑reliance on AI for filing suspicious activity reports. Regulators met with several institutions, privately reprimanding at least one major bank for generating low‑quality, high‑volume reports using large language models. While machine‑learning flagging...
Salesforce Executives Signal Declining Trust in Large Language Models
Salesforce executives acknowledge a decline in confidence in large language models (LLMs) over the past year, citing randomness and instruction‑following failures. The company is pivoting its Agentforce platform toward simple, rule‑based automation while restricting generative AI in certain scenarios. Salesforce...
Report: OpenAI May Embed Sponsored Content Directly Into ChatGPT Responses
OpenAI is actively prototyping ways to embed sponsored content directly into ChatGPT answers, including woven‑into‑response ads and sidebar placements. Internal mockups show both immediate product recommendations and post‑click suggestions for travel or retail queries. The company is also exploring the...
A Zelda Puzzle Proves AI Models Can Crack Gaming Riddles that Require Thinking Six Moves Ahead
Modern language models are now capable of solving multi‑step visual puzzles, as demonstrated by a color‑changing Zelda shrine challenge. Google Gemini 3 Pro, OpenAI GPT‑5.2‑Thinking, and Anthropic Claude Opus 4.5 were tested on the same screenshot, with GPT‑5.2 solving it consistently and quickly, Gemini 3 Pro...