It would cost a tenth of that to replicate GPT-4 by training on its outputs. It's quite likely that finetunes of the forthcoming larger StableLMs will rival GP

It would cost a tenth of that to replicate GPT-4 by training on its outputs. It's quite likely that finetunes of the forthcoming larger StableLMs will rival GPT-4. (and by that time GPT 4.3 will be shipping)