1. Cliffs. There are big swathes of theoretical CS where we know things can't be done, and the self-taught people would smash their faces into such problems forever because they won't even realise it's insoluble in principle rather than merely difficult, whereas with a more principled background you needn't expend this wasted effort.
2. Local Maxima. Without a principled understanding it's easier to mistake a local maximum you've stumbled into for a global maximum. After all, all the small tweaks you tried make it worse, so, how are you expected to guess that a violently different solution would be better? Theory could help but you don't have any.