Thanks for the response Raul.
You have a valid point about complexity. Generically speaking, decentralized architecture are, by definition, more complex that their centralized alternatives. Typically you sacrifice simplicity to gain other benefits such as trustlessness , federated work etc.
Your point about bias and over-fitting is valid but is far from being a generic characteristic of supervised models. There are plenty of scenarios in which data scientists optimized a model to fit the training data even when the training dataset is fairly optimal. Similarly, there are other characteristics of a training dataset that can introduce bias in a model that are not necessarily related with the dataset being biased in-and-out itself.
In terms of your skepticism about blockchain and AI, I will be the first one to tell you that most of the blockchain-AI platforms today look relatively naive from the AI perspective and can’t be used in mainstream AI scenarios. However, that doesn’t negate their value proposition which should improve as tier2 blockchain solutions become more mainstream. Having said that, there is a middleground. The folks at Numerai have been experimenting with some very interesting concepts that achieve certain level of decentralization without being a 100% decentralized platform.