Interesting, but also not
AI in continuous games isn't particularly impressive as it comes down to reaction times, and in that respect AI will always have the upper hand. It's a completely different type of game compared to Go, a turn-based game, where latency isn't crucial.
This proof-of-concept match has very stringent rules, such as both human and AI being able to play only one hero, and other environmental boosters ("runes") not being permitted. It's a million miles from the actual game of Dota 2, which is a 5 vs 5 astronomically more complex game.
There are very few details on how this bot works. We don't know if the AI is receiving the screen data as input, or if it's getting easily-parseable representations of the world. These are two massively different things for several reasons; a big one, IMO, is recognising animations. Vengeful Spirit (not the hero the AI is using in this case) is the most extreme example I can think of as all of her animations look *very* similar; one of which is a very powerful, fast-moving stun. If it's her stun, you have very little time to recognise it, both from sight and sound cues.
If the AI is "seeing" these animations as events in an easy-to-parse world representation, then the bot can surely instantly recognise that it is about to be stunned and take action much quicker than a human ever could. That being said, all ability casts can be cancelled a certain way through, so the bot could potentially be duped into a loop of thinking it's going to be stunned and then not. This adds some more interesting potential for the bot to recognise whether it's beneficial for the enemy to stun the bot, so it knows whether to actually bother reacting. But Dota 2 is an imperfect information game, so it sounds complex.
Anyway, I felt I had to rant about this. It's interesting, yes, but I don't think it makes any difference to everyday Dota 2. If you want to make a strong bot that has the same level of skill as a human then clearly it will need delays adding to its decisions, or perhaps it would have to make occasional deliberate mistakes. But otherwise, it's just Elon Musk's company waving their "deep learning" dongs around, showing that they've done it first.