There are some of us here to take RNGs VERY seriously. I happen to work on a hardware cryptographic card, which has an embedded hardware RNG.
The basic rule is that PRNGs are all but useless for anything other than toy applications. Even the best ones are subject to predictability, if one had enough data and knows the algorithm being used (and, one has to believe that there are organizations out there that can reverse engineer the hardware/software being used).
There are even some us who are somewhat skeptical of NIST's SP800-90A "Recommendation for Random Number Generation Using Deterministic Random Bit Generators" as being possibly predictable (although unlikely).
As for noise sources, note that not all sources are necessarily random. Some of the posters here have mentioned using Avalanche/Zener Diodes as a source of randomness. There is some empirical evidence that indicates Avalanche/Zener Diodes may exhibit a negative resistance characteristic, under a certain set of conditions, in which case, the typical circuit will produce a relaxation oscillator, which will produce a VERY non-random signal. As evidence of this, consult figure 5, on page 19, of On-Semi's "TVS/Zener Theory and Design Considerations
Especially, note those zigs in the expanded portion of the voltage-current chart, and realize that those zigs represent regions of negative resistance. The theory behind these zigs is something called "Microplasma Discharge Theory", which, as far as I've been able to tell, is not well understood in the physics community.
There are, of course, a LOT more considerations that need to be given to producing a good RNG, but I really don't want to write a book here.
P.S. I'll get my coat. It's the one with the pair of loaded dice in the pocket.