Is there "Secret Sauce'' in Large Language Model Development?

A study of 809 AI models shows that at the cutting edge, 80-90% of performance differences are explained by the amount of computing power used for training, not proprietary 'secret sauce'. However, for less advanced models, company-specific techniques and shared algorithmic progress significantly reduce the compute needed to achieve certain capabilities. The research also found huge efficiency variations within the same company, with some models being over 40 times more efficient than others.