[praestigiare]

You are imagining this situation from the perspective of already having a more powerful / general model and looking backwards, so it seems obvious. But that is not how things work. Without data to support it, you can imagine all kinds of possible theories - that is not science, and it certainly gets you no closer to understanding.

[jjk166]

Yes, the entire point is that what was obviously correct in hindsight was not clearly advantageous to begin with, the author's thesis is that exploring models which may seem worse than what we currently have can ultimately lead to better overall results. The author is specifically criticizing the narrow definition of science which says every incremental step no matter how small must improve your models' predictive power rather than exploring the possibility space.

[praestigiare]

Neither you nor the OP have provided a reasonable example of a model which is less predictive and also more "true." the OPs analogy has been discussed to death here, and yours is explicitly of a model which is in fact more predictive.

[karmakaze]

I didn't read it so specifically. If optimizing for X you may be in a local maxima so deviation looking in other regions seems fine.