This paper is really promising, but also confuses me: like most AI evaluation papers in traditional peer-reviewed journals the models are ancient due to the publishing process. But here they do pretty well unlike some of the other papers? https://www.science.org/doi/10.1126/science.adz4433