The idea is simple. To accelerate computation that is not necessarily in parallel. I thought why can’t you use a parallel guesses instead. These guesses are parallel models that can predict some value in the future. Also the models take advantage or inspiration from the already true computed values. One of the models takes over when it guesses enough right values in a row. It is accelerated because the model is a simplification and only accurate withing a range.
To make it more robust. I guess that you can backtrack and verify a randomly selected small number of samples with the slow model. If there is an unacceptable error in the computation the it will start over from that place and skip the fast model for that range.