[cbm-vic-20]

While it's outside my area of expertise, it seems that the "secret sauce" of GPT is not the algorithm, but rather the huge corpus of data to train from, and a massive amount of [CG]PU time (cost and kWh) to process it and run it through the training process.

This probably hasn't happened before because nobody wanted to pay the millions of dollars necessary to do it. Maybe nobody had any confidence that it would produce the type of results that OpenAI is now seeing.