There is plenty of low hanging fruit to harvest with these language models in the next year, let alone ten years. And the models are improving rapidly.

GPT is a sequence to sequence model. Many jobs are effectively also sequence to sequence models. Customer support, low end software development, and accounting all fit this pattern and are good candidates for GPT automation.

Jobs that require highly specialized reasoning may fall to S2S models eventually, but the models would need to be trained in that speciality area. For instance, if exposed to all the sequences involved in the career lives of thousands of actual lawyers, I’m sure you could train up a Lawyer-GPT that would do quite well. But how do you get access to those sequences? They are by definition kept under lock and key.

Similarly, you could train GPT to be a specialist doctor assuming you could somehow capture the working life of thousands of such specialists into a form that the machine could train on. But we don’t have that data yet, so these specialist areas will remain elusive for the time being.