OpenAI's new model technique threatens AI safety monitoring

openai is exploring a new technique for its models, called 'recurrent depth,' which is used in its astra ai model. this approach can help with model costs and performance but makes it harder to monitor a model's thinking process. researchers, including steven adler and nathan calvin, have expressed concerns that this could compromise ai safety monitoring. the technique may make it difficult or impossible to use chain of thought monitorability, a method that is considered one of the best threads for monitoring large language models. this development has raised concerns among researchers and could impact the industry's ability to prevent ai incidents.