Even smart people who understand A.I. seem to really be mischaracterizing how this will play out in the near term.
GPT is a model that humans can relate to, because they can use their iown language to interact with it. Don’t ask it to drive you home. It will happily try, but it is not the kind of model you want for that. Don’t ask GPT to do your taxes. Definitely don’t ask it to perform surgery. Again, it would oblige but the results would be about the same as if you averaged the surgical skills of everyone you know.
From here, we build a thousand models. Each optimized for their use case. Each embodying the understanding and values of their creators. In the very near future, like this year, when you ask GPT to do math, it will get help from Wolfram Alpha. When you ask it to drive, it might invoke Tesla. Hopefully by the time you need a surgery, DeepMind will have created AlphaOperation.
All the conjecture about LLMs is hot air if you don’t accept that their architecture and training is not appropriate for lots of things and that we will solve that with a multitude of models that are. Models that might suck at speaking your language but are better than the best humans at their intended task.