Alvin Goldman has proposed various versions of reliabilism, trying to resolve its counter-intuitive consequences. In this paper, I critically analyze some of them, and find all of them unsatisfactory, especially in their treatment of the relationship between external world and reliability. My proposal is to incorporate an internalistic aspect into reliabilism, by revising Goldman's notion of 'normal worlds'. This notion is too relativistic to be used as a basis of epistemic justification, but adding some more conditions may make it more useful. I propose the notion of 'the most successful set of normal worlds' as a replacement, which avoids difficulties with both externalism and individualistic internalism.
