[roenxi]

> I think that's a little unfair. Most folks have been pointing out that OpenAI's current momentum-based "deathball" strategy seems to fall apart without infinite regen and a limited hero pool, both facilitated by the current set of restrictions.

People said similar things about Go AIs and ko fights. And in the end it turned out that neural networks handled kos fine but ladders were a challenge.

On the deathball strategy in particular, consider that we expect a superhuman DotA AI to change the DotA metagame, so playing off-meta doesn't tell us anything. AlphaGo would invade 3-3 point a lot more enthusiastically than a human player. This was considered a classic beginner mistake for many years; now the theory has been readjusted to cope with the fact that AlphaGo stuck with it and just considered it a good move.

We can safely say that the courier change has made a deathball strategy more powerful and it seems quite likely it is not an optimum strategy. But we can't be sure until OpenAI tests it, and we absolutely can't be sure that OpenAI won't just learn a new style when the conditions change.

The criticisms have merit, but nobody has enough data predict anything about the future. Particularly a professional DotA player.