OpenAI Releases ‘Atlas’ Browser

OpenAI has officially launched ChatGPT Atlas, putting AI front-and-center in your browser, with Windows, Android, and iOS support promised soon. The Mac-compatible browser makes ChatGPT a real-time co-pilot. It reads page context, summarizes content, edits text inline with cursor chat, and might automate tasks via agent mode.

Browser memories let ChatGPT remember past sessions to enhance suggestions, fully controllable and deletable by users. Split-screen keeps ChatGPT visible alongside pages. Simon Willison offered his first impressions, noting Atlas’ technical chops but warning that agent mode can feel clunky and prompt injection stays a risk.

OpenAI can also be quietly training its AI to handle junior banking work through a secretive initiative called Project Mercurybased on Bloomberg. Internal documents reveal that the corporate has hired greater than 100 former investment bankers, many from firms like Goldman Sachs, JPMorgan, and Morgan Stanley, each earning about $150 per hour. Their feedback trains the AI to supply professional-grade financial models autonomously, effectively replicating the grunt work of entry-level analysts.

Sora 2 is tightening up after Bryan Cranston’s likeness and voice appeared without consent. OpenAI rolled out stricter safeguards to limit unauthorized replication of real people, aligned with SAG-AFTRA and the proposed NO FAKES Act.

And in research, GPT-5 uncovered an answer to Erdős problem #1043 (1958). OpenAI’s Sébastien Bubeck clarified that the AI didn’t invent latest math, it actually traced a subtle solution buried in a 1961 German paper by Pommerenke, translating and connecting dispersed knowledge.

Despite this, the ChatGPT mobile app is hitting a growth wall. U.S. engagement and session frequency are down over 20% since July, while global downloads fall 8% this month. Read more.

Google just launched Skills, a unified hub with nearly 3,000 courses and labs from Google Cloud, DeepMind, Grow with Google, and Google for Education. Learners can move seamlessly from AI Essentials to large language models and research foundations, and even construct enterprise-wide AI strategies.

Gemini Code Assist lets users run live coding experiments, while gamified streaks and shareable achievements keep engagement high. On top of that, Skills-based hiring links certified learners to 150+ employers, including fast-track programs for top talent. Free access is accessible for developers, educators, and Google Cloud customers.

Google Fi is leveling up as an AI-powered assistant. Calls get AI noise filtering on each ends, HD and HD+ support improves clarity, and a Gemini-powered billing bot can answer account questions immediately. RCS messaging involves web browsers in December, and Wi-Fi Auto Connect+ mechanically switches devices to secure, high-speed hotspots, showing a W+ icon when lively.

AI Studio’s latest Construct interface makes it easier than ever to go from idea to live app. Modular AI “superpowers,” secure storage for API keys, context-aware Gemini suggestions, UI annotations, and one-click deployment to Cloud Run let each developers and non-coders prototype quickly. With a goal of a million AI apps by the tip of the yr, Studio is positioning itself as Google’s go-to hub for hands-on, agentic AI.

YouTube is cracking down on AI misuse with latest likeness-detection tools for Partner Program creators. Verified through a photograph ID and a fast selfie QR scan, creators can now flag or remove AI-generated videos using their face or voice. The system also supports copyright claims and enforces consent, aligning with the NO FAKES Act. Read more.

Claude Code Web is now in beta, bringing Anthropic’s AI coding assistant straight into the browser and iOS. Developers can run cloud-based, sandboxed sessions with strict network and file system restrictions, manage multiple GitHub repositories, automate pull requests, generate real-time documentation, and oversee asynchronous AI agents, essentially letting engineers act as managers of autonomous coding assistants.

On the research side, Claude for Life Sciences, powered by Sonnet 4.5, now supports full workflows from literature review and hypothesis generation to protocol drafting and regulatory submissions. It integrates with Benchling, PubMed, BioRender, Scholar Gateway, Synapse.org, and 10x Genomics.

Anthropic can also be in talks with Google for a multibillion-dollar cloud deal, reportedly value tens of billions, to expand computing capability for Claude. The agreement, still under negotiation, would make Google Anthropic’s primary infrastructure partner, accelerating model scaling, agentic AI development, and enterprise deployments, and underscoring Alphabet’s strategic role in supporting U.S. AI leadership.

CEO Dario Amodei also addressed accusations of AI fear-mongering from Trump officials, reaffirming Anthropic’s ethical AI mission, bipartisan policy engagement, federal collaborations, California SB 53 compliance, and restrictions on PRC service sales. Read more.

• MCP is redefining how AI thinks by linking models on to live resources and executable tools

• Deepseek-OCR turns image-heavy documents into AI-ready data at 33M pages per day

• Nvidia discusses loan guarantee for OpenAI (Paywalled)

• Reasoning-native computing guarantees AI that truly understands and decides

• Qwen Deep Research can now turn your results into an online page or podcast

• IBM and Groq mix workflow orchestration and high-speed inference for smarter AI agents

• AI-powered eye implant restores reading vision in blind patients

• US scientists create microscopic ‘flower robots’ for drug delivery

• Jules will soon start clarifying all of your requirements before coding

• World’s first living computer goes global with 10 universities on board

• Genspark launches Custom Super Agent and Super Agent Store

• Meta to place killswitch on Instagram AI chatbots “early next yr” as part of latest parental controls

• Jensen Huang says the longer term workforce will likely be a mixture of ‘humans and digital humans,’ who could possibly be licensed out or hired—and want onboarding

• Amazon hopes to exchange 600,000 US employees with robots, based on leaked documents

• Microsoft proclaims open-source benchmark for AI agent cybersecurity investigations

• Humans and LLMs represent sentences similarly, study finds

• Opera brings deep research capabilities to its Neon AI browser

• Apple’s M5 chip might finally give Vision Pro the speed and thermal edge AR/VR has been waiting for

• Salesforce revamps Agentforce to try to drag customers across the gap between AI capabilities and AI adoption

• Adobe Foundry desires to rebuild Firefly to your brand — not only tweak it

• Yelp’s AI can now take reservations over the phone

• Gartner predicts AI will upend job markets, trigger lawsuits and rewire global commerce by 2027

• Netflix goes ‘all in’ on generative AI as entertainment industry stays divided

• Samsung’s latest TVs include Perplexity AI in-built — and a free yr of Perplexity Pro

• Napster returns as an AI companion to your MacBook

• AI assistants make widespread errors in regards to the news, latest research shows

• Goldman Sachs CEO David Solomon says there is a key difference between AI and past job revolutions

• A16z-backed Codi launches AI agent office manager

• KPMG leans into agentic AI in Salesforce, Google deals

• A brand new AI breakthrough makes live voice deepfakes fast, low-cost, and almost flawless

• You may now create a Shopify store by talking to Lovable’s AI

• When LLMs binge on social media sludge, their reasoning literally breaks down