Say you are measuring IQ before and after taking omega 3. Say your hypthesis is "omega 3 immedieately increases IQ". Assume this is true, even if it probably isn't. After testing this on 20 people you might be able to say "the hypthesis appears to be true, but there is a 46% percent chance these results are due to chance". However after testing 1000 people, you might be able to say "the hypthesis appears to be true, and there is a 0.5% percent chance these results are due to chance". Then you are 99.5% confident in your result. Standards for confidence varies from field to field, but typically 99% confidence or more is called a significant result. Given some assumptions, you can also calculate the number of samples needed to generate a result with confidence >X% with probability Y. According to Daniel Kahneman in Thinking Fast and Slow, most researchers don't take the time to do this, and are consistently surprised when their experiments turn out to not give significant results.
This depends on the experiment. Some of the things to account for are the number of variables, the confidence you are looking for and the amount of interdependence. Consult a statistician, particularly one expert in experimental trials, for an evaluation.
There are also exact formulas for simple tests.
Neither yourself, nor Peter Norvig suggested an appropriate number of subjects for tests like these. So I'm curious, what would be an appropriate number of subjects?