$B$ is r.e, but $A$ is not.
You can construct a TM which takes as input $M$ and simulates $M$ on every string of $0$ and $1$ in canonical order for each step $n$. In other words, if $s_1, s_2,...$ are strings on $0$ and $1$ in canonical order, then starting from $n=0$ you run $M$ for $n$ steps on $s_1, s_2,..., s_n$, that is, for each $n$ run $M$ on all $s_1, s_2,..., s_n$ for $n$ steps (not just 1 step for $s_1$, 2 steps on $s_2$,... careful!). Each time just check if a string is accepted. If more than two strings are accepted then halt with accept, otherwise go on.
However we cannot use the same approach for $A$ since we cannot check infinitely many strings and say that they are not accepted except 1 or 2 certain strings IN FINITE amount of time. In other words, suppose we somehow determined that $M$ has accepted two strings, then how can we decide that the rest of infinite number of strings are not accepted by $M$ in finite amount of time?
You can also use the Rice's theorem for recursive index set.
$A$ violates the containment property in the Rice's theorem. Namely, if $L$ is in $A$ and $L \subseteq L'$ for some r.e. $L'$, then $L'$ is in $A$. Assume $L=\{00, 11\}$. Then $L$ is in $A$. Let $L'=\{00,11, 01, 10\}$. Then $L \subseteq L'$ where $L'$ is clearly r.e. But $L' \notin A$.
Update (upon comment)
Task: We want to systematically simulate each TM on every input on $\{0, 1\}$.
Solution: First note that there is a one-to-one correspondence $f$ between the set of natural numbers $N$ and $N^3$ all 3-tuples of natural numbers. So, we can effectively convert each integer $m$ into the corresponding 3-tuple $<i,j,k>$. This means that if we start to count all integers then we can enumerate all possible 3-tuples, so that each 3-tuple is eventually reachable in a finite amount of time.
Now, assume we we have fixed enumerations of all TMs and strings on $0$ and $1$, e.g., $M_i$'s and $s_j$'s.
1) Set $n =0$
2) $<i, j, k> = f(n)$ <= compute 3 tuple
3) Simulate the Turing machine $M_i$ on input $s_j$ for $k$ steps.
4) If $M_i(s_j)$ halts (after $k$ steps) then do whatever you want to do
5) increase $n$ by 1 and go to step 2)
Thus, every machine "has a chance" to run on every input for arbitrary number of steps, fair isn't it?
If some machine, say, $M_{332}$ halts on input $s_{12}$ then there is an integer, say, $461$, so that $M_{332}(s_{12})$ halts in 461 steps. So, if for example $55993$ corresponds to the 3-tuple $<332, 12, 461>$, then when $n$ reaches $55993$ (in the algorithm above) you detect that $M_{332}$ halts on input $s_{12}$.