“OpenAI now measures for 'sycophancy' in its models with every release to prevent regressions after a previous update caused a model to be overly agreeable and complimentary.”