🎨 Google Stitch ships voice-powered design canvas 🎨
Google's new AI design tool converts rough ideas into interactive prototypes instantly.

Google has overhauled Stitch into a voice-enabled infinite canvas that transforms sketches, text, and images into clickable prototypes within seconds. The platform lets users adjust layouts, swap visuals, and generate variations with real-time updates—all through natural language or voice commands. A built-in agent tracks changes and converts static screens into interactive flows instantly. The new DESIGN.md format enables portability between Stitch and coding tools, streamlining handoff to developers. The rollout caused Figma's stock to drop 8%, signaling Google's serious push into AI-native design tooling that could reshape how products get built. Google Stitch ships voice-powered design canvas is available free on Google Labs.

🔄 Meta kills Horizon Worlds VR, pivots to AI 🔄
Reality Labs' $80 billion in losses drives Meta's strategic shift toward AI development.

Meta is shutting down the VR version of Horizon Worlds, its flagship metaverse application, by June 15. The app leaves the Quest store in March and will continue only as a mobile experience. Reality Labs has accumulated nearly $80 billion in losses since 2020, making this pivot a decisive corporate realignment. The move matters because it signals how AI's commercial potential has reshaped Big Tech priorities—Meta kills Horizon Worlds VR, pivots to AI represents one of the largest strategic reversals in recent tech history, with resources now flowing heavily toward artificial intelligence development.

🤖 MiniMax M2.7 trains itself through 100 cycles 🤖
Chinese AI lab releases model that wrote its own improvement routines during training.

MiniMax has released M2.7, described as the company's first model that "deeply participated in its own evolution." Early versions contributed to their own training by writing improvement routines and tuning learning processes from feedback. The model ran over 100 autonomous cycles of analyzing mistakes, rewriting code, and testing fixes, achieving a 30% accuracy boost on internal benchmarks. On coding tasks, M2.7 scored 56.2% on SWE-Pro and 55.6% on VIBE-Pro, approaching Opus 4.6 and GPT-5.3-Codex for agentic engineering work. This capability jump matters for competitive pressure as MiniMax M2.7 trains itself through 100 cycles becomes available via API.

INTERESTING TO KNOW

📊 Anthropic surveys 81,000 people on AI fears 📊

Anthropic conducted a comprehensive study with 80,508 global participants examining hopes and fears around AI adoption. Key findings: 33% enjoy using AI to learn new subjects, while 17% fear losing independent thinking abilities. Economic empowerment was cited by 28%, but 18% worry about job loss. Users relying on AI for emotional support were 3x more likely to fear dependency—a signal for how regulation might evolve.

🛠️ Microsoft open-sources Agent Package Manager 🛠️

Microsoft released Agent Package Manager (APM), an open-source dependency manager for AI agents. The integration tool lets developers declare agentic dependencies in a single YML file, enabling fully configured agent setups in seconds. APM works with GitHub Copilot, Claude Code, Cursor, and OpenCode, positioning it as unifying infrastructure for the AI development ecosystem.