How stupid this idea is?
What purpose or use case are you imagining? Several ML architectures have one model work on the outcome of another model. Like GANs have a discriminator model that works on the output of a generator model. Boosting models connect sequentially with one model working on the residuals of the output from previous model. But if you're just simply connecting model x and model y sequentially, i.e. not even doing a teacher-student arrangement, you basically just have a hybrid model x+y with different loss functions and backprops for part x and part y.
You’re ahead of your time
If ml 1 is expensive external service, makes sense to use its output to train an in-house ml 2, which you can call for free
that's what Transfer Learning is all about
Transfer learning is training on model on a task, and then applying it to another task. This isn't transfer learning.
you can choose a pretrained model to generate raw intermediate outputs to feed in a downstream model, you can also fine tune the pretrained model on your domain before using the output in another model ... in both cases it will be transfer learning .. and will be using a model's ground truth in some form to feed into another model ..go improve your understanding of TL first
Are there other options for generating the ground truth labels?
House of mirror.
why do you want to do that ? why not use the other model?
This is called meta learning.
Not stupid if the first model is well trained. It’s actually a good approach. But you’ll have to monitor for data drift of the first model. You can also fine tune the first model by backpropagating from the loss of the second model. works pretty well in my experience