The fact that we don't know some details doesn't mean that we can't make good predictions about _probabilities_.

Yes, because there's plenty of datapoints to make prediction by extrapolation. Just look at the progress between GPT-1, GPT-2, GPT-3 and GPT-4 and extrapolate to GPT-5, GPT-6 etc. There could be some roadblocks that would prevent further progress, but the default prediction should be that the trend continues.

Metaculus has had predictions for various AI milestones for years and has been consistently too conservative for resolved questions. It now predicts 50% of the invention of full AGI by 2032 which by extension might also be too conservative.