Constructing one-way functions from average-case hardness is a long-standing open problem. A positive result would exclude Pessiland (Impagliazzo ’95) and establish a highly desirable win-win situation: either (symmetric) cryptography exists unconditionally, or all NP problems can be solved efficiently on the average. Motivated by the lack of progress on this seemingly very hard question, we initiate the investigation of weaker yet meaningful candidate win-win results of the following type: either there are fine-grained one-way functions (FGOWF), or nontrivial speedups can be obtained for all NP problems on the average. FGOWFs only require a fixed polynomial gap (as opposed to superpolynomial) between the running time of the function and the running time of an inverter. We obtain three main results: Construction. We show that if there is an NP language having a very strong form of average-case hardness, which we call block finding hardness, then FGOWF exist. We provide heuristic support for this very strong average-case hardness notion by showing that it holds for a random language. Then, we study whether weaker (and more natural) forms of average-case hardness could already suffice to obtain FGOWF, and obtain two negative results: Separation I. We provide a strong oracle separation for the implication (∃ exponentially average-case hard language ⇒ ∃ FGOWF). Separation II. We provide a second strong negative result for an even weaker candidate win-win result. Namely, we rule out a black-box proof for the implication (∃ exponentially average-case hard language whose hardness amplifies optimally through parallel repetitions ⇒ ∃ FGOWF). This separation forms the core technical contribution of our work.