OP here with a friendly reminder that we should also look at the denominator when evaluating how useful the "internal tools make good startups" heuristic is.

Also note that I'm not claiming they are never good ideas, so pointing out instances where they've panned out isn't very interesting. It'd be more interesting if you could show that the hit rate among these types of startups is better than I'm suggesting (for that we need to consider the denominator).