The former OpenAI researchers warned that losing the ability to track an AI's step-by-step reasoning process creates major safety risks