[nostrademons]

That thought experiment was why I joined Google after 4 years of working in startups. The communication cost overhead the grandparent mentioned is why I left Google after 5 years.

While there, I observed a lot of projects where I was literally working with a room full of world-class people - folks who had given TED talks, folks who had started major open-source projects, folks who had written the "Bible" of their particular subfield of computing - and the final design we came up with was worse than what any one person in the room, working on their own, could've come up with. Good designs tend to be both controversial and coherent: they take a position, not everyone agrees with the position, but they do it anyway because self-consistency has its own benefits that are often intangible but highly valued by users. When you design by compromise, you end up sanding off all the most innovative (= hard to communicate) parts of the design, and end up with only the bare minimum that everyone can agree on.

It's interesting that when you put a bunch of average people in a group, have them independently make a prediction, and then average the predictions, you end up with a result more accurate than any one participant's prediction. When you put a bunch of really smart people in a group, have them cooperate to make a design, and look at the design, you end up with a design that's worse than any one expert's original design.

[ChrisFoster]

Yes, the joy of a great design is in the delicate balance it strikes between simplicity and conflicting needs. There's plenty of anguish in finding the balance even in the head of a single person.

The space of solutions to a design problem is really high dimensional and the quality of the design has many local maxima. So from that point of view it's not too surprising that averaging the features of several designs puts you at a point in the space which is not a local maximum. For a more apples to apples comparison with the averaging of predictions, it might be useful to consider predicting the value of a random variable drawn from a multimodal distribution. In this case it's highly likely that different people will focus their predictions on different modes and averaging several predictions will be worse than taking a single one in isolation.