x
Evidence of policy-amplified context drift in a long-form ChatGPT session — LessWrong