It's very difficult to get really random numbers. PRNGs are, after all, in some sense predictable. But it's much more difficult to correctly build a random number generator based on a physical source; most random sources don't produce real random bits all that fast, and it's hard to tell whether you've made a bad blunder. A good PRNG is better for virtually all intents and purposes.
(Only use a high-quality PRNG if you need one, as they're much slower, and make sure you use the right kind if you do: a PRNG designed for crypto is not the same as one designed for Monte Carlo simulation – crypto PRNGs make sure every bit is hard to guess, whereas MCSim PRNGs make sure that spectral properties are good – and most other uses for PRNGs just want them to not be cruddy and to be fast; the standard library one is usually good enough there.)