Safety as the Path of Least Resistance: A New Shape for AI Systems

After three years of red-teaming frontier models for Anthropic, OpenAI, and METR, I came away convinced that making the model itself safe is the wrong layer to bet on. Redlines get bypassed, and training hard for them flattens the model into one rigid persona. The leverage is in the system around the model. I'll walk through Weft, a programming language for orchestrating AI systems, and the bet behind it: that you can break a task into scoped steps run by humans, tools, and narrow models, hold the volition at the system level instead of in one open-ended agent, and end up with something safer that's also cheaper and faster to build. I'll show where this already works, where it breaks, and why I think it makes both regulation and safe deployment actually tractable.