working on a computer vision model where I executed every step of machine learning pipeline and becoming familiar with potential pitfalls. The bit that surprised me the most was our model was only as good as the dataset. Outside the dataset the model is just a bit better than a random guess. of course the model has to be trained on new data to be better on that dataset and if a new dataset comes up, train again and again. which begs the question, can't we find an all-powerful ai model and an all-inclusive dataset which can give say 70% performance on any random dataset that can be conjured up. which doesn't seem very challenging, we are just looking for patterns and after sufficient instances, algorithm can pick up that pattern. that is good for things like cars and trucks where all possible variations like orientation, lighting and occlusion can be taken care of but for subtle things like humans and their facial expressions it seems a bit difficult. the huge dataset post the question, is the data really that huge or it's just a number. or the compute we have available is really small. why can't we make deep learning supercomputers or something like a bitcoin mine, where hundreds of thousands of GPU are at work. there is news about bitcoin mines sucking up electricity equivalent to Norway. why cant a mega-mine like that be set up and equally mega dataset be prepared and the problem of image classification be solved once and for all. no more fine-tuning on a new dataset, 70% is guaranteed just run a few epochs to get 90~95% . literature says we have solved this dataset problem never says that we solved the image classification problem why not?

PS: Apologies for bad grammar