1. Google Told Developers API Keys Weren't Secrets. Gemini Made Them Secrets Anyway. For two decades, Google's official guidance was clear: API keys are not secrets. Treat them as public identifiers. Embed them in your website's JavaScript. Commit them to your repo.

2. Burger King's AI Assistant Also Grades Its Workers Burger King is putting an AI named Patty inside employee headsets. The system answers questions about meal prep, walks workers through order assembly, and coaches them in real time.

3. OpenAI Answers Its Moat Problem with Figma and Federal Paperwork A question rattled through the AI industry last week: if frontier models keep converging in capability, what exactly protects OpenAI's position?

In Brief

• Google Launches Nano Banana 2, Opens Pro-Level Image Generation to Free Gemini Users Google released Nano Banana 2 (Gemini 3.1 Flash Image), bringing image generation and editing capabilities previously restricted to paid tiers to all Gemini users. The model runs at Flash-tier speed and adds improved world knowledge, subject consistency, and production-grade output. It is available today across the Gemini app and Google's developer APIs. The Verge

• Andrej Karpathy: Coding Agents "Basically Work" Since December Karpathy stated that AI-assisted programming changed abruptly in December 2025, not gradually. He attributes the shift to higher model quality, long-term coherence, and tenacity on large tasks — calling the effect "extremely disruptive" to default programming workflows. Simon Willison

• Samsung Galaxy S26 Debuts with New Android AI Features at Unpacked 2026 Google showcased its latest Android AI capabilities integrated into Samsung's Galaxy S26 devices at Samsung Unpacked 2026. Google Blog

• Google Translate Adds AI-Powered Context, Alternatives, and Ask Buttons Google Translate now surfaces alternative translations, an "understand" button for deeper context, and an "ask" feature for clarifying ambiguous phrases. The updates use AI to expose nuances that single translations miss. Google Blog

• SkyReels V4 Generates Video and Synchronized Audio in a Single Pass SkyReels released V4, an open multimodal video foundation model that produces video and temporally aligned audio together. The architecture uses a dual-stream Multimodal Diffusion Transformer — one branch for video, one for audio. It accepts text, images, video clips, masks, and audio references as input. Hugging Face

• DREAM Benchmark Exposes "Mirage of Synthesis" in Deep Research Agents Researchers introduced DREAM, a benchmark for evaluating AI-generated research reports. The paper identifies a failure mode where strong surface fluency and citation alignment mask factual and reasoning errors. A four-vertical taxonomy maps the gap between apparent quality and actual accuracy. Hugging Face

• Simon Willison Publishes Agentic Engineering Patterns Guide Willison argues that cataloging your own technical knowledge — what's possible, roughly how — makes you a more effective director of coding agents. The guide frames "hoarding things you know how to do" as the key skill for agent-assisted development. Simon Willison

• PyVision-RL Fixes "Interaction Collapse" in RL-Trained Vision Agents Researchers released PyVision-RL, a reinforcement learning framework for open-weight multimodal models. It addresses interaction collapse, where RL-trained agents learn to minimize tool usage and multi-turn reasoning. The fix combines oversampling-filtering-ranking rollouts with accumulative tool rewards. Hugging Face

• ARLArena Proposes Stable Training Recipe for Agentic Reinforcement Learning Researchers introduced ARLArena, a framework addressing the training collapse that limits agentic RL at scale. The recipe targets longer interaction horizons and larger environments. Hugging Face

• Google and Massachusetts Launch Free Statewide AI Training Program Google partnered with the Massachusetts AI Hub to give all state residents no-cost access to Google's AI training curriculum. Google Blog

Read the full edition →