The author linked the "blockchain is the worst possible database" idea, but this is a consequence of the design requirements, so making a "better performing" blockchain database won't solve the problem.
The ideas about a decentralized market for AI training data are interesting as well, but we have a hard time securing our local databases. If followed through, this will end up with malicious actors and buying up our personal data from these markets in bulk, and to blackmail us.
Maybe the tech will be developed in the next decades, allowing a remote AI to be trained on my data without exposing it to third parties, but I'm sceptical about the technical feasibility, and this will require significant up-front resources to get there.