Anthropic's Code with Claude: Half of devs ship AI-written code unread

At Anthropic's London developer event, Code with Claude, a striking poll showed nearly half the room had shipped production code written entirely by Claude—often without reading it. Developers admitted they trusted the AI enough to skip code review, raising eyebrows about code quality, security, and accountability. Anthropic's CEO Dario Amodei emphasized the company's goal to push automation as far as it goes, claiming these tools boost productivity and let humans focus on higher-level design. However, critics like engineers from OpenClaw warn of an impending 'vibe-coded slop' crisis, where bad or dangerous code floods repositories. The tension mirrors broader debates in AI: how much autonomy should we grant these systems?

Meanwhile, at Google I/O, Demis Hassabis declared we're 'standing in the foothills of the singularity,' introducing Gemini for Science—an agentic AI system designed to execute cutting-edge research with minimal human input. Unlike specialized tools like WeatherNext, Gemini can call on multiple models and tools to autonomously design experiments, run simulations, and interpret results. This shift from specialized to general-purpose AI in science could accelerate discoveries but also raises questions about reliability and reproducibility. Together, these developments signal a new era where AI isn't just assisting but actively taking over core professional tasks, from coding to science.