A greedy algorithm works here. Keep adding elements to $B$ any way you like until you can't continue.
Therefore let $B$ be a subset of $A$ that meets the requirements and is maximal for inclusion. Let $p$ be the number of elements in $B$.
Since $B$ is maximal for inclusion, every one of the $n-p$ elements of $A - B$ belongs to some $A_i$ that already has two elements in $B$. Since two $A_i$'s of this kind can never have the same intersection with $B$, there are at most $\binom{p}{2}$ such $A_i$'s, proving that $\binom{p}{2} \geq n -p$. This inequality is equivalent to $p(p+1) \geq 2n$. Thus $$p + 1 > \sqrt{p(p+1)} \geq \sqrt{2n}.$$ This proves that $p \geq \lfloor \sqrt{2n} \rfloor$.