Any efficiently computable function can be represented by a circuit containing XOR gates and AND gates - AND gates alone would not suffice (but NAND gates would). The standard practice in secure computation is to use this {XOR, AND} basis to represent functions, since evaluating a XOR is often very cheap (it only involves cheap local operations, and no communication).
In GMW, a 1-out-of-4 OT is used to evaluate the AND gates (as I said above, XOR gates only require local computation - i.e., XORing the corresponding shares). This is the standard OT, i.e., not Rabin OT: the sender has 4 messages $(m_i)_{i\leq 4}$, and the receiver with input $j \in \{1,2,3,4\}$ learns $m_j$ without learning anything else; $j$ remains hidden to the sender.
(note that such an OT can be constructed from the Rabin OT, and conversely, Rabin OT can be constructed from this standard OT - so although their functionality differ, they are essentially "equivalent" in terms of what they allow to do).