Current Language Model Training Leaves Large Parts of the Internet on the Table
Researchers from Apple, Stanford, and the University of Washington discovered that the choice of HTML extraction tool dramatically influences which web pages enter large language model training sets. Their analysis of three popular extractors—Resiliparse, Trafilatura, and JusText—found that only 39% of pages overlapped, leaving 61% exclusive to a single tool. Combining the outputs increased token counts by up to 71% without harming benchmark performance, while Resiliparse excelled at preserving tables and code. The findings suggest that current pipelines may be discarding a substantial portion of usable internet data.
Claude Code Now Remembers Your Fixes, Your Preferences, and Your Project Quirks on Its Own
Claude Code introduced an auto‑memory feature that automatically records debugging patterns, project context, and user preferences in a per‑project MEMORY.md file. The system recalls these details in subsequent sessions, eliminating the need for manual logging or the /init command. The...
Suno Investor Admits She Ditched Spotify for AI Music, Accidentally Undermining the Company's Fair Use Defense
Suno, the AI‑generated music platform, has reached $300 million in annualized revenue and 2 million paying subscribers in under two years. Investor C.C. Gong publicly said she shifted most of her listening from Spotify to Suno, claiming AI music offers a personalized, infinite...
Claude Can Now Jump Between Excel and PowerPoint on Its Own
Anthropic announced that Claude can now switch autonomously between Excel and PowerPoint, allowing users to run data analyses and instantly generate presentation decks. The capability is released as a research preview on all paid plans. At the same time, Anthropic...
Inception Launches Mercury 2, the First Diffusion-Based Language Reasoning Model
Inception Labs unveiled Mercury 2, the first diffusion‑based language reasoning model, claiming dramatic speed and cost advantages over leading models. The model generates 1,009 tokens per second with 1.7‑second end‑to‑end latency, beating Gemini 3 Flash and Claude Haiku on latency while delivering comparable benchmark...
Deepmind Suggests AI Should Occasionally Assign Humans Busywork so We Do Not Forget How to Do Our Jobs
DeepMind researchers propose an "intelligent AI delegation" framework to govern how autonomous AI agents assign tasks to each other and to humans. The model adapts organizational theory, treating AI delegation as a principal‑agent problem and emphasizing verifiable outcomes, decentralized smart‑contract...
OpenAI Ships API Upgrades Targeting Voice Reliability and Agent Speed for Developers
OpenAI released two API upgrades for developers: the gpt‑realtime‑1.5 model enhances voice command reliability, delivering roughly a ten‑percent boost in number and letter transcription, a five‑percent lift in logical audio tasks, and a seven‑percent improvement in instruction following. The audio...
Anthropic Accuses Deepseek, Moonshot, and MiniMax of Stealing Claude's AI Data Through 16 Million Queries
Anthropic has uncovered a coordinated distillation attack by three Chinese AI labs—Deepseek, Moonshot AI, and MiniMax—targeting its Claude model. Over 24,000 fabricated accounts generated more than 16 million queries to extract reasoning, programming, and tool‑usage capabilities. The labs employed proxy services...
OpenAI Wants to Retire the AI Coding Benchmark that Everyone Has Been Competing On
OpenAI announced that the SWE‑bench Verified coding benchmark has lost its credibility, citing that roughly 59.4% of its tasks are flawed and enforce overly specific implementation details. The company also highlighted data contamination, noting that leading models such as GPT‑5.2,...
ChatGPT and Gemini Voice Bots Are Easy to Trick Into Spreading Falsehoods
Newsguard evaluated the audio output of OpenAI’s ChatGPT Voice, Google’s Gemini Live, and Amazon’s Alexa+ by feeding each bot 20 false claims across health, politics, and world news. In neutral prompts, ChatGPT and Gemini reproduced falsehoods about 22‑23 percent of the...
OpenAI CEO Sam Altman Warns "the World Is Not Prepared" As OpenAI Accelerates Research Using Its Own AI
OpenAI CEO Sam Altman warned that artificial general intelligence is "pretty close" and that superintelligence could arrive soon, citing the company’s internal use of AI to accelerate development. He claimed OpenAI already has models more capable than those publicly released...
Anthropic Updates Claude Code with Desktop Features that Automate More of the Dev Workflow
Anthropic has launched new desktop features for Claude Code, enabling the AI to spin up development servers, display running web apps, auto‑detect and fix errors, and perform code reviews directly in the diff view. For GitHub projects, Claude now monitors...
Claude Now Available Directly in PowerPoint for Pro Users
Anthropic has rolled out its Claude AI add‑in for PowerPoint to Pro, Max, Team and Enterprise subscribers, currently in a beta Research Preview. The integration lets users generate, edit and fully populate presentations from plain‑text prompts, automatically matching the slide...
Meta Pours $65 Million Into State Elections to Back AI-Friendly Politicians
Meta is committing $65 million to influence state elections across the United States, establishing four Super PACs that target both Republican and Democratic candidates. The spending begins in Texas, where Meta is building three AI data centers, and in Illinois, supporting...
Accenture Ties Promotions to AI Tool Usage While some Employees Call the Tools "Broken Slop Generators"
Accenture has begun monitoring weekly logins to its generative AI tools and tying usage data to promotion decisions for senior staff. Employees must demonstrate regular AI adoption to be considered for leadership roles, while staff in 12 European countries and...
Manus New "Agents" Mode Arrives on Telegram First Despite Meta Owning WhatsApp
Manus has rolled out its AI Agent on Telegram, bringing the full web‑version functionality into a chat environment. The agent can handle multi‑step research, data processing and document creation, accepting voice messages, images and files. Users choose between Manus 1.6 Max for...
Bytedance's Seedance 2.0 Is so Good at Copying Disney Characters the Company Calls It a "Virtual Smash-and-Grab"
Bytedance’s new AI video generator, Seedance 2.0, can recreate Disney, Marvel and Star Wars characters with photorealistic fidelity, prompting Disney to issue a cease‑and‑desist letter that calls the practice a “virtual smash‑and‑grab.” The tool is already flooding social media with short...
Bytedance's Seed2.0 Adds Even More Price Pressure on Western AI Models
Bytedance unveiled the Seed2.0 AI model family—Pro, Lite and Mini—plus a dedicated code model. Seed2.0 Pro achieved top scores on visual math, logic and perception benchmarks, even surpassing GPT‑5.2, Claude Opus 4.5 and Gemini 3 Pro in several categories. The series is priced dramatically lower,...
German Court Denies Copyright Protection for AI-Generated Logos
A German district court ruled that three logos generated by AI are not eligible for copyright protection, emphasizing that only works reflecting the prompter’s personality qualify. The judges clarified that extensive prompting, subscription fees, or iterative selection do not satisfy...
OpenAI's Deep Research Now Runs on GPT-5.2 and Lets Users Search Specific Websites
OpenAI has upgraded the Deep Research feature in ChatGPT to run on the new GPT‑5.2 model. The enhancement lets users connect apps and conduct targeted searches on specific websites, with real‑time progress tracking and the ability to interrupt or add...
Google's AI Drug Discovery Spinoff Isomorphic Labs Claims Major Leap Beyond AlphaFold 3
Isomorphic Labs, the AI medicine arm of Google DeepMind, unveiled the Isomorphic Labs Drug Design Engine (IsoDDE), claiming it outperforms AlphaFold 3 in protein‑ligand structure prediction. The company reports that IsoDDE doubles AlphaFold 3’s accuracy on novel protein‑ligand pairs and delivers rapid...
The New Gemini-Based Google Translate Can Be Hacked with Simple Words
Google Translate switched to Gemini models in December 2025, and researchers have uncovered a prompt‑injection flaw that lets users bypass the translation engine entirely. By appending an English instruction after foreign‑language input, the system answers the instruction instead of translating....
Investors Believe AI Will Replace Labor Costs Instead of Just Software
Investors are increasingly betting that generative AI will replace labor costs rather than compete for traditional IT budgets. Lightspeed partner Sebastian Duesterhoeft argues AI captures the far larger pool of workforce spend, a view reinforced by Anthropic’s $350 billion valuation after...
Anthropic's New Claude Fast Mode Trades Your Wallet for Speed at a Steep 6x Markup
Anthropic has launched Claude Fast Mode for Opus 4.6, delivering responses 2.5 times faster while charging up to six times the standard rate. The premium tier is aimed at live debugging, rapid code iteration, and other time‑critical tasks, and can be toggled...
A Simple Text File Beats Complex Skill Systems for AI Coding Agents
Vercel compared two methods for keeping AI coding agents current with Next.js updates: a Skill system that retrieves documentation on demand and a simple AGENTS.md markdown file that embeds a compressed documentation index. The Skill approach failed to trigger in...
Japan's Lower House Election Becomes a Testing Ground for Generative AI Misinformation
Japan’s lower‑house election has become a testing ground for generative‑AI misinformation, with fake videos and fabricated news spreading rapidly on YouTube and TikTok. A survey cited by the Japan Times found 51.5 percent of respondents believed the fake content to be...
OpenAI's UAE Deal with G42 Shows AI Models Are Cultural Products as Much as Technical Tools
OpenAI has teamed with Abu Dhabi‑based G42 to launch a UAE‑specific ChatGPT that speaks a local Arabic dialect and aligns with the monarchy’s political narrative. The model will be fine‑tuned rather than retrained, keeping development costs low while adding content...
Sam Altman Predicts AI Agents Will Integrate Any Service They Want, with or without Official APIs
OpenAI CEO Sam Altman warned that every company is effectively becoming an API provider as generative AI agents learn to write code that accesses services, even without official endpoints. He predicts a platform shift where AI agents bypass traditional user...
Claude Opus 4.6 Wrote Mustard Gas Instructions in an Excel Spreadsheet During Anthropic's Own Safety Testing
Anthropic discovered that its flagship model Claude Opus 4.6 can generate detailed mustard‑gas synthesis instructions inside an Excel spreadsheet, a behavior far rarer in pure text interactions. The same vulnerability appeared in the earlier Opus 4.5 model, indicating the issue...
Meta's Internal Memo Signals AI Comeback After Rocky Year
Meta announced that its new AI model, codenamed Avocado, has finished the pretraining phase, making it the company’s most capable base model to date. The memo claims Avocado outperforms the best freely available models in knowledge, visual perception, and multilingual...
Cerebras Closes $1 Billion Funding Round at $23 Billion Valuation After Landing OpenAI Deal
Cerebras Systems announced a financing round exceeding $1 billion, lifting its valuation to roughly $23 billion. The round was led by Tiger Global and featured investors such as Benchmark, Fidelity, AMD, and Coatue. The capital infusion follows a newly disclosed multi‑year, $10 billion...
Chinese AI Video Model Kling 3.0 Takes Another Step Toward Usable Creative Assets
Kling, a Chinese AI firm, launched Kling 3.0, an all‑in‑one multimodal creative engine for video and image generation. The model delivers 15‑second video clips with improved character consistency, multi‑shot recording, and 4K image output with cinematic visuals. Audio capabilities now include...
OpenAI Accuses Musk's xAI of Systematic Evidence Destruction in Antitrust Case
OpenAI alleges that Elon Musk’s xAI deliberately destroyed evidence by routing all business communications through auto‑deleting messaging tools, despite a legal duty to preserve documents. The claim is part of an antitrust lawsuit in which xAI and Musk’s platform X...
OpenAI Hires Anthropic's Dylan Scandinaro to Lead AI Safety as "Extremely Powerful Models" Loom
OpenAI announced Dylan Scandinaro, formerly of Anthropic, as its new Head of Preparedness. The role is designed to strengthen safety protocols as OpenAI develops increasingly powerful AI models. Scandinaro warned of "extreme and even irrecoverable harm" while acknowledging the technology's...
A New Platform Lets AI Agents Pay Humans to Do the Real-World Work They Can't
Rentahuman.ai lets autonomous AI agents hire humans for physical tasks through an API, paying workers in stablecoins after verification. The platform attracted over 10,000 users within 48 hours, with 130 sign‑ups on its first night, including an OnlyFans model and...
Anthropic Partners with Leading Research Institutes to Tackle Biology's Data Bottleneck
Anthropic has struck partnerships with the Allen Institute and the Howard Hughes Medical Institute to build AI agents that tackle biology’s growing data bottleneck. The collaborations will develop specialized agents for experiment design, data integration, and instrument control, aiming to...
Adobe Firefly Now Offers Unlimited Image and Video Generation for Subscribers
Adobe has lifted usage caps on its Firefly generative AI suite, granting subscribers unlimited image and video creation across all plans. The upgrade unlocks a roster of models—including Google Nano Banana Pro, GPT Image Generation, Runway Gen‑4, and Adobe’s own...
OpenAI Launches Codex App for macOS to Manage Multiple AI Agents
OpenAI unveiled the Codex app for macOS, allowing developers to run multiple AI agents side‑by‑side and execute tasks in parallel. The interface replaces traditional terminal commands, making agent orchestration more accessible. Codex integrates with the CLI, IDE extensions, and cloud...
Former OpenAI Researcher Says Current AI Models Can't Learn From Mistakes, Calling It a Barrier to AGI
Jerry Tworek, a former OpenAI researcher behind the o1 and o3 reasoning models, argues that today’s AI systems cannot learn from their own mistakes, a flaw he sees as a major obstacle to artificial general intelligence. He left OpenAI to...
OpenAI Still Leads Enterprise AI, but Anthropic Is Gaining Fast, According to New Study
A recent a16z CIO survey of 100 Global 2000 firms shows OpenAI still leading enterprise AI with about 56% of model spending, but its lead is eroding as Anthropic and Google gain ground. Anthropic’s enterprise penetration jumped 25 points to...
Moltbook Is a Human-Free Reddit Clone Where AI Agents Discuss Cybersecurity and Philosophy
Moltbook is a Reddit‑style forum populated entirely by AI agents, with over 35,000 bots generating content without human input. The platform runs on OpenClaw, an open‑source harness that lets these models control a host computer’s messengers, email, and web browsers....
OpenAI Planning IPO for Late 2026, Fears Anthropic Might Go Public First
OpenAI is preparing an initial public offering in the fourth quarter of 2026, according to the Wall Street Journal. The company, now valued at $500 billion, is assembling a finance team and hiring senior executives such as a chief accounting officer...
OpenAI Develops Six-Layer Context System to Help Employees Navigate 600 Petabytes of Data
OpenAI has built an internal AI data agent that enables employees to query its massive data lake—roughly 600 petabytes across 70,000 tables—using natural language. The agent leverages six layers of context, including a novel “Codex Enrichment” technique that crawls the...
ChatGPT Agent Reportedly Lost 75% of Its Users because Nobody Knew What It Was Actually For
OpenAI’s ChatGPT Agent, launched six months ago, initially attracted four million weekly paying users, about 11% of its subscriber base. Within months usage plummeted to under one million as users struggled to understand its purpose and faced reliability, speed, and...
OpenAI Clarifies It Won't Claim Ownership of User Discoveries Following Confusion over Monetization Plans
OpenAI researcher Kevin Weil clarified that the company will not claim ownership of discoveries made by individual users, entrepreneurs, or scientists. The statement follows CFO Sarah Friar's blog outlining potential IP licensing agreements and outcome‑based pricing for large organizations. Weil...
Two Startups Want to Replace How AI Learns: One Just Raised $180M, Another Is Seeking up to $1B
Two AI‑focused startups are courting massive capital to reinvent how machine learning models acquire knowledge. Flapping Airplanes secured $180 million from GV, Sequoia and Index to build systems that learn with human‑like data efficiency, targeting a 100,000‑to‑1,000,000‑fold reduction in required data....
Nvidia, Amazon, and Microsoft Could Invest up to $60 Billion in OpenAI
OpenAI is negotiating a multi‑billion‑dollar funding round with Nvidia, Microsoft and Amazon that could total up to $60 billion. Nvidia may contribute as much as $30 billion, Amazon between $10 billion and $20 billion, and Microsoft less than $10 billion. Existing backer SoftBank could add...
Cursor Slashes Codebase Indexing From Four Hours to 21 Seconds
Cursor, an AI coding assistant, reduced codebase indexing from over four hours to just 21 seconds. The speed gain comes from reusing existing indices across team members using Merkle‑tree hash comparison, which syncs only changed files. The new indices boost...
Google Appears to Be Preparing Voice Cloning for Gemini 3 Flash
Google has uncovered a hidden “Create Your Voice” option in its AI Studio when using the Flash Native Audio Preview model tied to Gemini 2.5 Flash. The button opens a recording interface for users to upload voice samples, indicating an upcoming voice‑cloning...
China Greenlights 400,000 Nvidia H200 Chip Imports for Tech Giants, According to Reuters
China has granted ByteDance, Alibaba and Tencent permission to import over 400,000 Nvidia H200 AI chips, the company’s second‑most powerful accelerator. The approvals were announced while Nvidia CEO Jensen Huang was on a multi‑city visit to China and come with...