Claude Design Takes On Figma, and Google the Open Web

Claude Design Takes on Figma

Anthropic is launching Claude Design, an experimental tool that creates prototypes, presentations, and mockups using prompts. Users describe what they want ('a calming meditation app with nature-inspired colors'), and Claude generates an initial version, which can then be refined through further prompts or direct edits. The system reads company codebases and design files to maintain consistent brand styles. Exports are available as PDF, URL, PPTX, or directly to Canva for further processing. Available for Claude Pro, Max, Team, and Enterprise subscribers, based on Claude Opus 4.7. The launch follows Claude Cowork (January) and agent-based plug-ins, while Bloomberg reports on funding rounds that would value Anthropic at $800 billion. → TechCrunch AI

Synthszr Take: Anthropic is turning the democratization of design into a business model. While Figma spent years trying to close the gap between designers and non-designers with ever-better tools, Claude Design skips the tool layer entirely: language becomes the direct interface for visual creation. This is reminiscent of the evolution of photography, where experts first worked in darkrooms, then Polaroid came along, and today everyone takes pictures with their smartphone. Anthropic's move is clever: it doesn't attack Figma head-on but positions itself as a preliminary step that can then be exported to established tools. Reading design systems turns Claude Design into an in-house brand compliance machine, ensuring that even the most creative prompt lands within corporate design. Anthropic is betting that the future of design lies not in better tools, but in no longer having to operate any tools at all.

Google Squeezes Website Traffic Even Further

Google is integrating its AI Mode deeper into Chrome: in the future, links will open directly next to the AI answer in the same window, instead of in a new tab. The website becomes a context window next to the chat. Users remain in the search dialog while the AI merges content from the open page with content from other web sources into a synthesized answer. A new plus menu also allows users to pull several already open tabs, images, and PDFs into a search query as context. Google sells this as 'simpler' navigation, while publishers have to watch their pages become raw material for Google-generated answers. The features are initially launching in the US. → The Decoder

Synthszr Take: Google is upending the traditional browser hierarchy: the website is no longer at the center with search as an auxiliary function; instead, the chat becomes the main stage, while websites are degraded to context suppliers. This is reminiscent of early shopping TV channels, where the host became more important than the product itself. Early testers report that they remained more 'focused' on their task when they didn't have to switch between tabs. What Google is selling here as a feature is the gradual transformation of the open web into a database for AI assistants. The real innovation isn't the split-screen, but the fact that users will primarily consume web content filtered through an AI's interpretation in the future.

While OpenAI is finally integrating its long-promised super app into the Codex API, two of its most ambitious architects are losing faith: Kevin Weil (Science Research) and Bill Peebles (Sora) are out. The diagnosis is clear—if you really want to build a super app, you first have to bury everything else.

OpenAI IPO Fantasy: We're Building the Super App

OpenAI is transforming its Codex platform from a specialized coding assistant into a comprehensive super app. The update brings background computer use for Mac apps, parallel agents, an integrated browser with Atlas technology, and image generation directly into the application. With 3 million weekly users and 70 percent month-over-month growth, Codex is positioning itself as a direct competitor to Anthropic's successful Claude Code. Thibault Sottiaux, head of Codex, states the ambition clearly: 'We are building the super app in public.' In parallel, OpenAI is launching GPT-Rosalind, its first domain-specific model for life sciences, which can read scientific papers, design lab experiments, and generate biological hypotheses. → The Rundown AI

Synthszr Take: OpenAI is playing the super app game by the rules of WeChat, not the rules of Windows. Instead of building a monolithic system that can do everything, specialized modules are being created (Codex for development, Rosalind for life sciences) that are orchestrated via a common platform. The trick is: each module must prove itself on its own before it becomes part of the bigger picture. This is reminiscent of the emergence of medieval trading cities, where specialized artisans first established their workshops before a networked economic system emerged from them. The 70 percent growth rate of Codex shows that this modular approach works. OpenAI is thus avoiding the complexity trap that Microsoft fell into with Windows Vista or Google with Google+: breadth without depth leads to interchangeable products.

OpenAI IPO Reality: We're Shrinking to Get Healthy

OpenAI is losing three key figures from its most ambitious projects. Kevin Weil, who led the Science research initiative, Bill Peebles, the architect of the AI video tool Sora, and Srinivas Narayanan, CTO for enterprise applications, are leaving the company. The departures mark the end of OpenAI's 'side quests'—customer-facing bets like Sora, which consumed a million dollars a day in computing costs and was shut down last month. OpenAI for Science, the internal research group behind the AI platform Prism for accelerating scientific discoveries, is being integrated into other teams. Weil had claimed back in October 2025 that GPT-5 had solved ten unsolved Erdős problems—a claim that immediately fell apart. The consolidation on enterprise AI and the upcoming 'super app' show: OpenAI is sacrificing its moonshot projects in favor of focused growth. → techcrunch.com

Synthszr Take: OpenAI is carrying out what is known in biology as 'apoptosis': the programmed cell death of parts that no longer serve the whole organism. Sora was bleeding a million dollars a day, and the Science initiative was producing embarrassing Erdős fantasies instead of breakthroughs. The departures of Weil, Peebles, and Narayanan are not an escape, but the result of a deliberate amputation. Sam Altman is turning OpenAI into a focused enterprise predator, while researchers dream of 'cultivating entropy.' We know the pattern: Google went from a colorful playground to an ad machine optimizer, Facebook from a campus connection to an attention extraction engine. OpenAI's transformation is more brutal because it liquidates the research romance along with it—a company narrowing itself into a pure inference engine for paying enterprise customers.

AInything Goes: Shopify Gave AI Agents Write Access

Shopify released the AI Toolkit this week, a collection of tools that give AI agents direct access to the Shopify platform. The toolkit connects AI tools like Claude Code, Cursor, or VS Code with Shopify's APIs and store management functions via plugins, skills, or an MCP server. This allows developers to create apps that can access Shopify's documentation, API schemas, and code validation. The crucial point: the tools have 'store execute capabilities'—they can write directly to Shopify stores and make changes. The toolkit supports Node.js 18+ and is compatible with common AI development environments. Installation is done either via automatically updatable plugins (recommended), manual agent skills, or a local MCP server. → Aakash Gupta

Synthszr Take: Shopify is doing what's known in urban planning as the 'desire path principle': instead of prescribing paths, you see where people walk naturally and then pave it. AI agents have long been accessing APIs, piecing together tools, and generating code—Shopify is now legitimizing and standardizing this practice. The 'store execute capabilities' are like an open window on the 20th floor: convenient for a breeze, but you hope nobody jumps. The toolkit is cleverly positioned between control (validated API schemas) and freedom (direct write access). Shopify is betting that the productivity gains from AI agents will outweigh the risks—and that its documentation is good enough to prevent hallucinations. A calculated experiment in distributed autonomy, where Shopify provides the guardrails and then watches what happens.

Roblox's AI Assistant Now Makes Plans

Roblox is equipping its AI assistant with a 'Planning Mode' that takes a fundamentally different approach to game development: instead of generating code directly from a prompt, the AI first creates an editable action plan. The assistant analyzes existing code and the data model, asks follow-up questions about the desired style (cartoon, realistic, fantasy), and how assets should be created (build from scratch or from the Creator Store). Developers can refine the plan before implementation begins. Two new features support this: Mesh Generation creates textured 3D objects directly in the game, and Procedural Model Generation speeds up asset production. Roblox argues that single-step prompt-to-output systems often miss the creator's original intent. → TechCrunch

Synthszr Take: Roblox is turning its AI assistant into an architect who draws the blueprint before pouring the foundation. The analogy to professional software development is obvious: design documents, technical specs, and code reviews are nothing but multi-stage planning processes. What Roblox is implementing here is reminiscent of the way urban planners work, inserting an iterative coordination process between vision and construction. The Planning Mode is an admission that AI agents fail at complex tasks if they can't externalize intermediate steps. It will be interesting to see if other platforms follow suit or continue to rely on one-shot magic. Roblox is betting that the future of AI assistance lies less in magical output and more in structured collaboration.

Canva's Strategy Against Adobe: DIY Instead of Partnerships

With AI 2.0, Canva is pursuing a radically different strategy than Adobe. While Adobe relies on external partnerships and integrations, Canva is developing a standalone AI system with built-in long-term memory directly within the product. The new platform is prompt-based and remembers brand guidelines, design preferences, and project history across sessions. Users can create complete campaigns, layouts, code, text, and brand assets in a continuous workflow via chat. The system precisely edits individual objects and maintains context over multiple interactions. Canva's bet: its own memory agent will make external tools obsolete. → AI Secret

Synthszr Take: Canva is copying Apple's M-chip playbook: control over the entire stack instead of best-of-breed integration. Memory as a core feature is reminiscent of the early days of Evernote, except that Canva's AI not only collects information but also actively works with it. Adobe, on the other hand, is building an ecosystem like Android—open, modular, dependent on third parties. The irony: Adobe's Creative Cloud started as a monolithic suite, while Canva began as a simple browser tool. Now, the architectural philosophy is flipping. The winner will depend on whether creatives prioritize a consistent experience (Canva) or maximum flexibility (Adobe). Canva's advantage: they don't have to migrate legacy users who have had Photoshop shortcuts in their muscle memory for 20 years.

Google Hybridizes Gemini: 'I Know What You Did Last Summer'

Google is equipping Gemini's Personal Intelligence feature with Nano Banana-powered image generation. The system creates personalized images based on context from Gmail and Google Photos, without users needing to explicitly mention their preferences in the prompt. Instead of 'Generate an image of my dream house, my interests are tennis and music,' 'Design my dream house' is now sufficient. The feature accesses labels in Google Photos and understands group terms like 'family,' enabling prompts like 'Create a picture of my family and me doing our favorite activity.' Google uses a 'Sources' button to show where Gemini gets the context for image generation. → The Download from MIT Technology Review

Synthszr Take: Google is perfecting the business model of convenient surveillance. Nano Banana isn't a technical breakthrough, but the monetization of twenty years of data collection through a new interface. We know the principle from biology: parasites that manipulate their host so skillfully that the host perceives the invasion as a symbiosis. Gmail content and photo labels become training material for personalized images, while users celebrate giving up control as a feature. The 'Sources' button is like the fine print in a loan agreement: technically present, practically ignored. Google is turning privacy into an opt-out problem where the default setting is maximum data usage.

Microsoft Successfully Lobbied the EU: Data Center Emissions to Remain a Trade Secret

Microsoft and other US tech giants have successfully lobbied the EU to keep the environmental impact of their data centers secret. An investigation shows that the industry's demands were incorporated almost verbatim into EU regulations. The secrecy clause, which the European Commission added to its proposal in 2024 after being lobbied, prevents public scrutiny of the emissions of individual data centers. Researchers will now only receive national summaries of energy balances. The AI boom has led to a massive expansion of chip-filled data centers, whose energy appetite is partly met by fossil gas. Legal experts warn that the blanket confidentiality clause could violate EU transparency rules as well as the Aarhus Convention on public access to environmental information. → The Guardian

Synthszr Take: Microsoft has established a franchise system for pollution: the brand stays clean while the individual branches operate in the dark. The industry that once proudly announced its green ambitions is suddenly silent as soon as AI comes into play. We know this pattern from the tobacco industry in the 1960s: expansion can only continue if the data is kept under lock and key. The EU Commission copied the lobby's template so precisely that only two words were changed (probably by accident). Prof. Jendrośka, who monitored the Aarhus Convention for 19 years, cannot recall a comparable case. The tech giants are betting that the AI gold rush is more important than transparency about its true costs.