My layperson impression is that biological brains do online retraining in real time, which is not done with the current crop of models. Given that even this much required months of GPU time I'm not optimistic we'll match the functionality (let alone the end result) anytime soon.
>>foobar+(OP)
I'm actually playing with this idea: I've created a model from scratch and have it running occasionally on my Discord. https://ftp.bytebreeze.dev is where I throw up models and code. I'll be releasing more soon.