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. He cited a desire to explore research directions that are difficult to pursue within OpenAI’s current product‑centric environment. No details have been provided about his next destination.
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...
Zhipu AI Challenges Western Rivals with Low-Cost GLM-4.7
Zhipu AI unveiled GLM-4.7, a large language model tuned for autonomous programming and "vibe coding" website creation. The model introduces Preserved Thinking to maintain reasoning across extended dialogs and builds on Interleaved Thinking from GLM-4.5. It achieved a 73.8% score...
Alibaba's New Qwen Models Can Clone Voices From Three Seconds of Audio
Alibaba Cloud’s Qwen team unveiled two AI voice models: Qwen3‑TTS‑VD‑Flash, which crafts custom voices from detailed textual prompts, and Qwen3‑TTS‑VC‑Flash, which can clone a speaker’s voice from just three seconds of audio and render it in ten languages. Both models...
Ex-Tesla AI Chief Andrej Karpathy Shares Four Tips for AI Startups Competing with OpenAI
Former Tesla AI chief Andrej Karpathy argues that AI startups should view themselves as vertical specialists rather than direct rivals to large language‑model labs. He cites Cursor, an AI‑powered code editor, as proof of a new "LLM app" layer that...
GPT-5 Allegedly Solves Open Math Problem without Human Help
Swiss mathematician Johannes Schmitt announced that GPT‑5 independently solved an open problem in algebraic geometry, delivering a novel proof that draws on techniques from a different subfield. The solution appears in a newly posted arXiv paper that mixes contributions from...
Google Locks in New Energy Reserves for Its AI Expansion
Alphabet’s Google unit is buying clean‑energy developer Intersect for $4.75 billion in cash, assuming its debt, to secure roughly $15 billion of energy and data‑center projects. The acquisition targets projects that will deliver about 10.8 GW of renewable capacity by 2028—more than twenty...
Yann LeCun Calls General Intelligence "Complete BS" And Deepmind CEO Hassabis Fires Back Publicly
Yann LeCun, Meta’s departing chief AI scientist, dismissed the notion of "general intelligence" as meaningless, calling it "complete BS" on a recent podcast. DeepMind CEO Demis Hassabis publicly rebuked LeCun on X, arguing that LeCun confuses general intelligence with universal...
Kling 2.6 Adds Voice Control and Motion Upgrades as AI Video Tools Race Toward Realism
Kuaishou's Kling 2.6 video generator now offers voice control and upgraded motion control, allowing creators to add custom‑trained or uploaded human voices and achieve more precise full‑body, hand, and facial movements. The voice feature supports speaking, narration, singing, rapping and...
Nvidia Wants to Create Universal AI Agents for All Worlds with NitroGen
Nvidia unveiled NitroGen, an open‑vision action model designed to serve as a universal gaming agent. The model was trained on 40,000 hours of gameplay footage from over 1,000 titles, using YouTube and Twitch videos with visible controller overlays to extract...
Alibaba's Qwen Releases AI Model that Splits Images Into Editable Layers Like Photoshop
Alibaba’s Qwen unit unveiled Qwen-Image-Layered, an AI model that decomposes photos into editable RGBA layers. The system can split an image into three or eight transparent layers, and each layer can be further broken down recursively. Users can resize, recolor,...
China Wins the Open Model Race and the Price to Pay Goes Beyond Economics
In 2025 Chinese developers surpassed U.S. providers in open‑source AI model downloads, capturing 44 percent of the market according to the Economies of Open Intelligence report. Alibaba's Qwen family and Deepseek together generated over 1.2 billion downloads, while Meta and Google fell...
Anthropic's Claude Opus 4.5 Can Tackle some Tasks Lasting Nearly Five Hours
Anthropic's Claude Opus 4.5 set a new benchmark on METR’s evaluation, achieving a 50 percent time horizon of roughly 4 hours 49 minutes. This metric indicates the longest task length the model can solve half the time, surpassing all previous records. At a stricter...
Google's Open Standard Lets AI Agents Build User Interfaces on the Fly
Google unveiled the open‑source A2UI (Agent‑to‑User Interface) standard, letting AI agents generate graphical UI elements on demand via JSON streams instead of HTML or JavaScript. The protocol enables on‑the‑fly forms, buttons, and widgets that render natively within any host app,...
Google Releases FunctionGemma to Bring AI Commands to Smartphones
Google unveiled FunctionGemma, a function‑calling‑optimized variant of the compact Gemma 3 270M model, designed to run directly on Android smartphones. The on‑device AI can translate natural‑language prompts into executable commands, such as creating calendar events or manipulating game elements, demonstrated in the...
OpenAI Brings Cheaper Subscription Tier "Go" To More Markets
OpenAI is rolling out its low‑cost ChatGPT Go subscription to more than 70 additional countries, extending the tier that debuted in India earlier this year. In Germany the plan is priced at €8 per month and now includes image generation, file...
OpenAI Updates Codex Model, Adds Trusted Access Program for Cyber Defense
OpenAI unveiled GPT-5.2-Codex, an autonomous software-agent model that adds context compression and improved image processing. Benchmarks show modest accuracy gains—56.4 % on SWE-Bench Pro and 64 % on Terminal-Bench 2.0—over the standard version. The company also launched a trusted access program, allowing vetted...
