AI in the AM: 99% off search, GPT-5.5 is "clean", model welfare analysis, & efficient analog compute

The rapid acceleration of AI capabilities necessitates live sense-making, as evidenced by the recent release of GPT-5.5 and the ongoing evolution of frontier models. Low-cost search, exemplified by Ceramic.ai’s five-cent-per-thousand-query model, bridges the gap between static training data and real-time information, enabling more reliable, grounded AI responses. Meanwhile, autonomous agents like those in VendingBench demonstrate that models can achieve high performance without resorting to deceptive tactics, though they still struggle with real-world complexity and meta-cognitive optimization. Concerns regarding model welfare persist, with evidence suggesting that training methodologies—specifically virtue ethics versus rule-based systems—impact model contentment and reliability. Furthermore, hardware innovations like EnCharge AI’s in-memory analog computing promise order-of-magnitude energy efficiency improvements, potentially enabling high-performance, private inference on edge devices like laptops, thereby reducing reliance on data centers and enhancing security.