Swiftly deprecating old models' users depended on mistake, says OpenAI's Altman