I can see what he's saying: to make a great product, basically consider market forces your local gradient, then follow the steepest descent. Surely, though, that suffers from the same problem as the classic version of the gradient descent algorithm: it's easy to get stuck in a local optimum. So I do think your starting point matters.

His example, gmail, suffers from this too. There were many webmail and offline email clients before gmail, the key to its success was its integration with an excellent search algorithm. Without that starting point, it would have most likely been pulled towards some local optimum which had already been discovered. Okay, so Google is full of search experts, so maybe internal market forces would have been different than those of the worldwide market. Most startups don't have that kind of micro-market which takes them to some kind of new optimium though.

Taking the algorithmic analogy further, in practice you would probably not want to do pure gradient descent if it feels like you're heading for a known local optimum, and instead stochastically/heuristically "go against the flow" and experiment with idiosyncratic features.