Datacentric AI versus GPT3/few-shot models

Thanks for the great panel! What is the future of AI markets: task specific models trained with manually tailored data, or global models (GPT3) that can easily learn from small datasets?