My guess is one thing to watch out for are local models. Qwen and Gemma are getting pretty good and we're getting to a point where we won't melt a GPU on a lower parameter model. In the next year or so, I could see laptops with 8GB VRAM being able to do quite a bit with 7B to 9B local models. By the end of the decade, things could be very interesting with them if they keep improving at their current rate. =D
Oh good point! I can definitely see local models and efficiency improvements becoming a growing share of the AI ecosystem in the coming years, and local fine tunes also becoming easier. Forgot to mention that the 2034 PTMs in the timeline are also local models trained using local hardware, but trained on selected data!
My guess is one thing to watch out for are local models. Qwen and Gemma are getting pretty good and we're getting to a point where we won't melt a GPU on a lower parameter model. In the next year or so, I could see laptops with 8GB VRAM being able to do quite a bit with 7B to 9B local models. By the end of the decade, things could be very interesting with them if they keep improving at their current rate. =D
Oh good point! I can definitely see local models and efficiency improvements becoming a growing share of the AI ecosystem in the coming years, and local fine tunes also becoming easier. Forgot to mention that the 2034 PTMs in the timeline are also local models trained using local hardware, but trained on selected data!