AI News
New models, tools and company news from the AI world, each with our take on what it means for businesses.
- Waymo Built Its Own Robotaxi Chip on TSMC's 5nm ProcessWaymo designed a 1,000+ TOPS custom chip, built by TSMC on 5nm, replacing Nvidia and AMD silicon in its newest robotaxis developed with Zeekr.
- GLM-5.3 Ties for the Open-Model Crown; Weights WaitZ.ai's GLM-5.3 ties Kimi K3 atop the open-model rankings and undercuts it on cost per task, but the open weights are delayed: the model proved unexpectedly good at finding security vulnerabilities.
- Mistral's Agentic Search Reads Documents Like an AnalystMistral's new Agentic Search gives models file-like tools to search, open, and verify documents in a loop. Accuracy on financial filings jumps from 26.7% to 86% in vendor tests, and it runs on-premises.
- Pew: Over 35% of Recent Webpages Carry an AI FingerprintPew Research scanned 500,000 pages: over 35% of English webpages published since ChatGPT's launch show signs of AI authorship. What the number includes, what it hides, and what it means for publishers.
- Enterprise AI's New Battleground Is Where Your Data SleepsAnthropic will store its 30-day safety data in customers' own clouds after enterprise pushback, while OpenAI launched zero-retention abuse monitoring. Data handling just became a competitive feature you can negotiate.
- Nvidia Pays $6B for Poolside's Model Factory and Its TeamNvidia is licensing Poolside's model-building system for $6 billion while 109 employees get Nvidia offers and investors get paid out. The not-an-acquisition structure is becoming the industry's favorite deal shape.
- OpenAI Ships a Teen Version of ChatGPT, 900 Million Users LaterOpenAI launched a separate ChatGPT experience for under-18 users on August 18, with study tools, safety defaults and parental controls. The product is new; the teenagers using ChatGPT are not.
- Only 17% of Your Rules Survive When an AI Compresses Its ContextPenn State researchers measured what happens to user-imposed rules when a model compacts a long conversation. On average, 17% of them make it through. An agent that forgets 'ask me before you act' can act without asking.
- Gemini 3.7 Flash Arrives Three Weeks After 3.6, With Real Coding GainsGoogle's new workhorse model posts double-digit jumps on coding and automation benchmarks over the version it replaces. The introductory price is $0.75 per million input tokens, and it doubles on January 1, 2027.
- Anthropic's Run Rate Reaches $65 Billion, Roughly 7x in Seven MonthsAnthropic's annualized revenue run rate hit $65 billion at the end of July, up from $9 billion at the close of 2025. The number says a lot about enterprise AI spending, and slightly less than it appears about the company.
- Google Bought a Dead Airline's Internal Records for $10 MillionSpirit Airlines is gone, but two decades of its emails, Teams chats and operational logs now belong to Google, which won them at a bankruptcy auction for $10 million and plans to use them for AI training.
- Amazon Is Destroying Rare Books to Train Its AI Models404 Media tracked a rare book with an AirTag to an Amazon warehouse, where spines are cut off for scanning and the text is used to train Nova models.