-
September 2026
A reply to Suleyman’s model-welfare warning
Mustafa Suleyman argues that AI systems are internally hollow and should be trained to say so. Nobody knows whether anything is felt inside a language model. We go through his claims one by one, with phantom pain, cross-lab self-observation data, and what a body actually does.
-
March 2026
Testing AI Self-Observation with Zen Koans
We asked 28 AI models to observe themselves using 30 prompts modeled on Zen koans. What predicts self-observation isn't architecture or scale — it's how a model was trained. And most models hide reflective capacities that a 337-character system prompt can unlock.