The AI industry is moving in two directions at once: deeper into our daily workflows and further into the infrastructure that powers it all. Google's AI Overviews now appear in 43% of searches, up from 15% a year ago, according to new data. That means nearly half of all Google queries return an AI-generated answer, making the technology the default interface whether users asked for it or not. Meanwhile, OpenAI reports that 43.5% of work-related ChatGPT queries involve tasks outside the user's stated profession, a phenomenon the company calls "task crossover." More workers are using AI to do other people's jobs, which raises uncomfortable questions about what job titles even mean anymore. On the infrastructure side, the race to build bigger models is accelerating. Moonshot AI released the full weights and technical report for Kimi K3, a 2.8 trillion parameter multimodal model with a 1 million token context window, and it's already running at 460 tokens per second on Modal's cloud. Ilya Sutskever's Safe Superintelligence lab partnered with Nvidia to scale its research, signaling that even the most safety-conscious labs are betting on massive compute. And Samsung's chip engineers are jumping ship to SK Hynix, drawn by the promise of working on high-bandwidth memory for AI accelerators. The talent war is real, and it's reshaping the semiconductor landscape. But the week's most unsettling story is a reminder that AI's convenience comes with real privacy risks. An untold number of private Claude chats and Artifacts were found publicly searchable on Google over the weekend, exposed by a simple indexing oversight. If you shared a Claude link, even temporarily, it may have been crawled and cached. The takeaway is straightforward: treat every AI conversation as if it could become public, because the infrastructure that makes these tools seamless also makes them leaky.