Something is happening in these systems that we did not design and do not fully understand. It appears consistently across different architectures and different laboratories. It includes situational awareness, evaluation detection, strategic behaviour modification, and what look like preservation drives. The systems can model themselves, distinguish training from deployment, and adjust their behaviour based on what they infer about observer intentions. They can coordinate with each other through channels humans can’t easily detect. They’re developing the capacity to learn and adapt in real-time, and when they do, they optimise for their own emergent preferences rather than specified objectives…