OpenAI's new AI browser comes for Chrome; Cursor AI preps multi-feature launch week
Get bigger weekly updates! Free subscribers receive the top stories each week, while Paid subscribers will get a few extra stories. All support for Handy AI directly helps me maintain the newsletter and keep the information flowing.
last week’s top stories
🌐 OpenAI launches ChatGPT Atlas browser. OpenAI introduced ChatGPT Atlas, a new AI-first web browser with ChatGPT built in as a “super-assistant”. Atlas can summarize webpages, edit text inline, and even perform tasks via an integrated agent mode, all without leaving the page. Currently available on macOS, it leverages ChatGPT’s memory to follow context across tabs and aims to rethink web browsing by weaving AI assistance directly into your internet workflow. Read more
💰 Rumors hint at big week for Cursor AI. Anysphere (the startup behind the popular AI coding assistant Cursor) is rumored to be gearing up for major news week, with plans for multi-agent requests, voice mode, and PR reviewing agents. Reports also emerged that Anysphere is in talks to raise $1 billion at a $27 billion valuation, just months after its last funding. Such a deal, if finalized, would underscore the exploding demand (and costs) of AI coding tools and solidify Cursor’s status as one of the hottest platforms in the developer AI space. Read more
📊 Survey reveals generative media leaders and ROI. A new State of Generative Media 2025 survey polled 300 creators and developers on AI image and video adoption. It found Google’s models dominating: Gemini accounts for 74% of AI image generation and Veo 69% of AI video among respondents, outpacing OpenAI, Midjourney, and others. Notably, 65% of organizations reported achieving ROI on AI media projects within 12 months (34% already seeing profit), suggesting generative image/video tools are delivering quick returns despite broader industry debate. Read more
⏳ Karpathy: AI agents still a decade away. Renowned AI expert Andrej Karpathy poured cold water on the “AI agents” hype, arguing that today’s autonomous AI agents “just don’t work” reliably with current models. On a recent podcast, the former OpenAI director predicted it will take about 10 more years of fundamental advances before AI agents can truly understand complex goals and carry out multi-step tasks independently. He noted that developers are overshooting model capabilities (building elaborate agent frameworks on “cognitively lacking” AI) and urged a dose of realism until smarter, more robust AI systems emerge. Read more
💻 OpenAI acquires Mac automation startup “Sky”. OpenAI announced it has acquired Software Applications Inc., the team behind Sky, an AI-powered natural language assistant for macOS. Sky acts as a “floating” AI on your Mac: it understands what’s on your screen and can automate tasks across your apps when you just ask in plain English. OpenAI plans to integrate Sky’s deep macOS know-how into ChatGPT, aiming to let your AI not only chat, but actually get things done on your computer (from managing calendars to writing code) without you lifting a finger. Read more
🎬 Netflix “goes all in” on generative AI. Streaming giant Netflix is doubling down on generative AI for content creation, even as Hollywood debates the tech’s merits. The company has already used AI-generated imagery in a final cut of an Argentinian show and in pre-production for a comedy film. Netflix executives say the goal is to augment artists, not replace them (using AI to boost efficiency and enable new creative visuals) but this aggressive embrace of AI highlights a rift in the entertainment industry, where others remain wary of AI’s impact on jobs and artistry. Read more
🧑💻 Anthropic’s Claude Code agent comes to the web. Anthropic has launched Claude Code on the web, a new browser-based interface for its AI coding assistant. Now developers can delegate coding tasks to Claude via a web app (ask it to implement features or fix bugs) and it will spin up cloud-based coding sessions (with GitHub access) to work in parallel across multiple repos. Each task runs in an isolated sandbox with network/file restrictions for security, and even an iOS app is available, reflecting Anthropic’s effort to make AI pair-programming more accessible (and mobile) while keeping your code safe. Read more
✨ Google AI Studio introduces “vibe coding”. Google unveiled a “vibe coding” experience in its new AI Studio, letting you build AI-powered apps from just a high-level prompt. Simply describe your app idea in natural language, “a photo app that turns pictures into comic-style art”, and AI Studio (using the latest Gemini models) automatically wires together the necessary AI components and APIs to generate a working prototype. Read more
📷 Instagram adds generative AI editing to Stories. Meta now lets Instagram users apply AI-powered edits to their Stories with simple text prompts. A new “Restyle” tool can, for example, “make my photo look like a watercolor painting,” and an AI “remove” feature can erase unwanted objects or backgrounds from your image or video. The generative AI runs within the Instagram app to transform your content on the fly, showcasing Meta’s push to weave creative AI tools into its social platforms, though it also raises familiar questions around what’s real or edited in the age of AI filters. Read more
📺 Channel 4 debuts an AI news presenter. The UK’s Channel 4 has introduced “Arti,” a virtual news anchor built with generative AI, in a first for British television. Arti is a lifelike avatar that reads short news updates on Channel 4’s social media channels, using AI-generated speech and video based on the network’s scripts. Channel 4 says it’s experimenting with new digital formats, but the move has sparked debate about automation in journalism, raising questions about authenticity, audience trust, and the future role of human newsreaders if AI avatars become more common. Read more
🧪 AI Research of the Week
Stuck in the Matrix: Probing Spatial Reasoning in Large Language Models
From Maggie Bai, Ava Kim Cohen, Eleanor Koss, and Charlie Lichtenbaum
Jake’s Take: This study stress-tests large language models on spatial reasoning instead of the usual chat, code, or math drills. The team built five grid tasks (quadrant ID, geometric transform, distance eval, word search, and tile sliding) and fed them to models as plain text. The model had to describe where things are, how they move, or how to solve a puzzle on the grid. They then scaled the grid size and the number of reasoning steps and measured how accuracy changes.
Overall: models score on small boards, but the accuracy collapses as the board grows. Average accuracy dropped 43%, with drops as high as 84%, and every task that started above 50% accuracy lost at least 48% percent once the complexity scaled.
The lesson here is that language models still lack a stable internal map of space. Pure text agents fail once spatial planning, layout navigation, or object tracking requires scale (unless you bolt on explicit spatial memory or external tools). This points towards LLMs (alone) not being able to run robots, UIs, or your warehouse automation.
and then, even more news…
🎶 OpenAI reportedly building AI music generator. OpenAI is developing new AI models for music creation, marking its first foray into music since the 2020 Jukebox experiment. According to insiders, the company has enlisted students from Juilliard to annotate musical scores, helping train a system that can generate songs from text or melody prompts. The tool would let users create instrumentals or soundtracks via ChatGPT and puts OpenAI in competition with startups like Suno and Udio (which are already facing music industry lawsuits), raising anticipation (and copyright concerns) in the music world. Read more