"Breakthroughs" are inherently hard to predict. This field is advancing at a very rapid pace. A lot of the improvements are incremental, but even these are co

"Breakthroughs" are inherently hard to predict. This field is advancing at a very rapid pace. A lot of the improvements are incremental, but even these are coming very fast, and moving in different directions at once. I don't think there's any likelihood of replicating GPT-4 on your M1 in the next twelve months, especially not if you're expecting responsive performance. What we could see are a plethora of models dedicated to doing particular tasks. Say, specialists in aspects of programming or accounting or law or medicine. Or general knowledge models with access to a local cache of Wikipedia. Individually, none of these models would have to come close to GPT-4's overall level of power and flexibility. But collectively, they could reach that level of utility.