📝 TL;DR Anthropic has launched Claude Design, a new creative tool that lets people build prototypes, presentations, one-pagers, and visual assets by working with Claude through conversation. The big shift is that AI is moving beyond writing and coding, and now trying to become a real design collaborator too. 🧠 Overview Anthropic announced Claude Design as a new Anthropic Labs product powered by Claude Opus 4.7. It is built to help teams create polished visual work faster, whether that means interactive prototypes, pitch decks, product mockups, or marketing assets. This matters because most people have ideas they want to show, but not everyone has the design skills or time to turn those ideas into something polished. 📜 The Announcement Claude Design launched on April 17, 2026 in research preview for Claude Pro, Max, Team, and Enterprise subscribers. Anthropic says users can start from a prompt, uploaded files, or even a codebase, then refine designs through chat, inline comments, direct edits, and AI-generated control sliders. The company is also positioning it as a bridge between design and implementation, with export options to Canva, PDF, PPTX, HTML, and a direct handoff path into Claude Code. ⚙️ How It Works • Prompt-to-design workflow - Users describe what they want, and Claude builds an initial version that can be refined through conversation. • Design system support - During onboarding, Claude can build a team design system by reading code and design files so future outputs stay on brand. • Multiple starting points - Teams can begin from text prompts, uploaded images and documents, website captures, or an existing codebase. • Fine-grained editing - People can leave inline comments, directly edit text, or use sliders Claude creates for spacing, color, and layout changes. • Team collaboration - Designs can be shared across an organization, edited by teammates, and refined together in group conversations.

📝 TL;DR Google briefly published an experimental app called COSMO to the Play Store, then pulled it almost immediately. But the leak may have revealed Google’s real plan for on-device AI assistants: local Gemini Nano, built-in agent skills, and a hybrid system that keeps helping whether you are online or offline. 🧠 Overview According to reports, COSMO is a 1.13GB experimental Android app that appears to combine on-device Gemini Nano with cloud support in one assistant-style experience. The app reportedly includes 14 built-in skills covering things like calendar suggestions, browser actions, deep research, photo lookup, timers, list tracking, and conversation summaries. That matters because it suggests Google is moving beyond a simple chatbot and toward a more proactive assistant that can actually help manage everyday tasks on your phone. 📜 The Announcement This was not an official product launch. COSMO appears to have been accidentally published to the Play Store and then removed within hours, just ahead of Google I/O 2026. The app reportedly looked rough and experimental, which makes it seem more like an internal test bed than a finished consumer product, but the feature set is still a strong signal about where Google may be heading next with Gemini and on-device AI. ⚙️ How It Works • Runs Gemini Nano locally - The app reportedly includes a local model on the device, which points to offline capability and faster private processing. • Hybrid local and cloud mode - COSMO appears to switch between local Nano and a cloud system depending on connectivity and task needs. • Fourteen built-in skills - Reported features include calendar suggestions, browser automation, deep research, photo lookup, summaries, timers, and more. • Browser agent support - One of the most interesting reported features is a browser skill linked to Mariner-style web automation. • Assistant-style interface - COSMO seems designed less like a normal chat app and more like an always-available helper for tasks and context.

📝 TL;DR OpenAI has launched workspace agents in ChatGPT, shared AI agents that can handle long-running workflows across tools, teams, and systems. This is a bigger shift than custom GPTs, because these agents do not just answer questions, they keep working in the background and help move real work forward. 🧠 Overview Workspace agents are OpenAI’s new team-focused agents built directly into ChatGPT. They are designed to automate repeatable workflows like report prep, lead research, software reviews, and internal support tasks, while pulling context from connected tools and continuing to work in the cloud even after you log off. The big picture is simple: ChatGPT is moving from assistant to operator. 📜 The Announcement OpenAI introduced workspace agents on April 22, 2026 in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans. The company says these agents are powered by Codex, can be shared across a workspace, run on schedules, and work inside ChatGPT or Slack. OpenAI also says GPTs will remain available for now, with conversion tools coming later so teams can migrate existing workflows more easily. ⚙️ How It Works • Shared team agents - Teams can build one agent and reuse it across the organization instead of creating separate personal setups. • Cloud-based execution - Agents keep running in the background even when the user closes their laptop or leaves ChatGPT. • Codex-powered actions - These agents can write or run code, work across files and tools, and handle more than simple chat tasks. • Workflow builder - Users start by describing a recurring process, and ChatGPT guides them step by step to turn it into an agent. • Slack integration - Teams can deploy agents into Slack so they can respond where people already work. • Connected context - Agents can pull from workspace tools, memory, files, and systems to complete multi-step work with more context. 💡 Why This Matters

📝 TL;DR A 23-year-old with no advanced math training used GPT-5.4 Pro to help solve an Erdős problem that had been open for 60 years. The wild part is not just that it got solved, it is that the breakthrough came from letting AI brainstorm freely instead of forcing it into the usual expert playbook. 🧠 Overview Liam Price, a young math enthusiast rather than a professional mathematician, helped crack Erdős Problem #1196 using GPT-5.4 Pro. The solution was later confirmed by leading mathematicians, including Terence Tao, and the result is now listed as solved. This matters because it suggests AI may be most powerful not when it imitates expert thinking, but when it helps people escape it. 📜 The Announcement The story gained attention after Scientific American reported that Price used GPT-5.4 Pro to explore the problem in a loose, idea-first way, what people are now calling vibe maths. Instead of starting with rigid formal structures, he let the model suggest unconventional directions and then filtered the useful ones. The problem had been open for decades, and the approach GPT-5.4 Pro surfaced appears to have broken out of a pattern that had kept human researchers stuck. ⚙️ How It Works • Idea-first prompting - Price reportedly let the model generate broad approaches first instead of boxing it into strict formulations too early. • Human filtering - The AI did not replace judgment, it produced possibilities and Price filtered what seemed promising. • Unconventional route - The method used was not the standard framework experts had been reaching for on this class of problem. • Expert verification - Leading mathematicians, including Terence Tao, reviewed the result and confirmed the solution. • Public proof trail - The solution was posted through the Erdős Problems site, where the problem is now marked solved. • Fast exploration loop - The workflow shows how AI can compress the brainstorming stage of difficult research dramatically.

📝 TL;DR Canva has launched Canva AI 2.0, a major shift from design tool to prompt-powered creative platform. And the timing is perfect, because just as Canva goes all-in on AI creation, industry chatter says OpenAI may be preparing its next flagship model to make prompting itself much less important. 🧠 Overview Canva’s new Canva AI 2.0 is one of the clearest signs yet that AI is moving beyond single-purpose tools and becoming the operating layer for everyday creative work. Instead of starting with a blank page, users can start with an idea, brief, sketch, or goal, and let Canva generate layered, editable outputs across documents, presentations, websites, spreadsheets, and more. The bigger picture is that the AI race is no longer just about smarter models, it is about who turns that intelligence into real workflows people actually use. 📜 The Announcement At Canva Create 2026, Canva introduced Canva AI 2.0 as its biggest product launch since the company started. It says the platform is powered by the Canva Design Model, which it describes as the first foundation model built specifically for creativity and real-world design structure. The launch is live in research preview, with new capabilities including conversational design, agentic orchestration, layered object intelligence, memory, web research, scheduling, connectors, brand intelligence, Sheets AI, and Canva Code 2.0. At the same time, unconfirmed industry reporting around OpenAI’s next flagship model is adding even more fuel to the sense that the AI platform race is about to accelerate again. ⚙️ How It Works • Conversational creation - Canva now lets users start with a prompt, idea, rough thought, or sketch and turn that into editable creative work. • Agentic orchestration - The system chooses the right Canva tools behind the scenes to turn a goal into a finished output. • Layered editing - Outputs are built as editable objects, not flat locked images, so users can tweak specific elements without breaking everything else.