I watched the Open AI play against the "team" of pros at the calibration match earlier this month. Couple of observations and takeaways. The first is that the

I watched the Open AI play against the "team" of pros at the calibration match earlier this month. Couple of observations and takeaways. The first is that the bot strategy currently revolves around the special rule of 5 invulnerable couriers. Bots find microing lots of units effortless, so the map constantly showed each bot's courier flying back and forth carrying regen. The bots never had to really go back to base or their shrines to heal. This is important because it changes the meta of the game entirely. The way the game is structured allows only one (very vulnerable) courier per team. Usually this means that after a team fight, teams need to reset since they've expended significant resources for the fight. But that meta was non existent under the rules for matches against the Open AI five. The humans had trouble coping with this as they weren't used to the idea of ferrying regen constantly. Takeaways here - I could go on about the nuances of a single courier. But basically, the bots' gameplay will likely have to change once it comes down to 1 shared courier per team. Not sure how that will affect the architecture of a "no shared mind". Also, humans will likely need to take a page out of this gameplay and realise that couriers are a highly underutilized resource. Every second it's not doing something for no reason is just as bad as a hero not doing anything. The second observation comes from the last game of AI vs pro humans. This was an interesting game where the audience picked a losing set of heroes for the team. Despite a predicted chance of winning being less then 2% (iirc) he AI could have probably won on account of being mechanically better than the humans. But their insistence on sticking to a strategy of "push hard" found them doing really strange things. The strangest of this was Slark running ahead to cut down creep waves in the lane on its own. The human players knew this would happen and they kept forcing the Slark to go hide in the trees and at some point they were always able to corner it and get the kill. Over and over again. The Slark never changed. Similar things happened around the map during this game. What should have happened was that the AI should have adapted to its disadvantage, and poured its efforts into first defending and then snowballing later with its mechanical advantage. But that element of "intelligence" was never there. The takeaway is this. The AI will eventually beat the humans on account of them being always mechanically better. They need very slight changes in their strategy to win 99.9% of the time. They can be aggressive beyond any human possibility because they can calculate everything to perfection. How long it will take them to travel across the map vs how much longer it will take for an opposing hero to have its ultimate ready for example. There are a lot of mechanical components to Dota that the AI will always have an advantage over. But the AI will likely always reveal quirks that can be turned into dumb winning strategies (aka cheese strats). Something like the whole team fighting from the trees for example might just confuse the AI terribly. We don't know but every now and then someone will discover it and the teams working on the AI will have to "patch" the behaviour. Final takeaway from all of that - I'm not sure if training the AI towards "objectives" is really the best metric towards making an intelligent bot. It seems like what's instead happening is that we get software that has no intelligence at adapting in the moment to things its never seen even if they are brain dead. But it'll get better at hiding them through mechanical perfections. Upside - We get AI's capable of doing increasingly complex things in a seemingly perfect manner. Downside - We get a scary future of AI filled with byzantine issues that need to be "patched".