Eliezer doesn't say that it is impossible, only "pretty unlikely". That is, under usual circumstances, when you do the math, the benefits of being a member of a tribe that benefits from group selection, although greater than zero, are much smaller than the individual benefits of defecting against the rest of the group. This is the norm, in nature. This is what happens by default. The rare situations where this is not true, require special explanation. For example, ants or bees can collectively gather resources... but that is only possible because most of them are infertile children of the queen, so they cannot spread their genes better by defecting against the queen. In your example, are the "groups" different species? In other words, is this about how bees would outperform bumblebees? In that case, the answer seems to be that the feature B itself is almost a miracle -- something that turns a profitable behavior into inprofitable behavior, without being itself selected against by evolution... how would you do that? (So how did bees evolve, if for their pre-bee ancestors, a worker being infertile was probably an evolutionary disadvantage? I have no idea. But the fact that there are only about three known examples in nature where this happened -- ants, bees, naked mole-rats -- suggests it was something pretty unlikely.) Then you have humans, which are smart enough to recognize and collectively punish some activities that harm the group. If they keep doing so for generations, they can somewhat breed themselves towards harming the group less. But this is very slow and uncertain process, because the criminals are also smart enough to hide their actions, the enforcement has many loopholes (crimes are punished less if you are high-status, or if you do the thing to enemies), different societies have different norms, social order breaks down e.g. during wars, etc. So we get something like slightly fewer murders after a few centuries of civilization.