How we built the unbeatable Dota 2 bot

Those who are thinking that AI is being overly hyped, and that the demos like these are not smart enough, think about it, today's AI is smarter enough IMO, as the potential of real-world applications of AI as of now itself could be vast. It just need some practical polished implementations. Did the Dota 2 bot learn the game from scratch by learning the rules all by itself? No, as evident from the post itself - "small amounts of “coaching” with self-play. But think about it. It need not be that smart to replace humans. IMO, many of our jobs do not need to very complicated self-learning. In real world day-day jobs like agriculture, factories, and to a great extend what we call as skilled labour like driving, management etc, the kind of work is often repetitive. AI is already better than humans at recognising speech, vision and to some extend languages. Combining these skills with mechanical robots in a meaningful way can easily replace what we, as humans can do, except in niche fields. Even there computer can offload our job in several ways. Robots/AI can do it much much faster, and much more efficient than us and it will only be getting better and better with overall advancement in research and the amount of data available..

So it's kind of interesting and equally scary that we are looking at a potential future, possibly near, where most of the work can be handled by a trained robot with humans only supervising them in exceptional cases. Robots can then could be only need to be smart enough to alert the human supervisors when they reach a state where they cannot perform the job normally, for e.g. power grid going off in a factory and they couldn't figure out what to do, it could only be a 0.001 % chance. The point is AI need not get as good as humans in learning complex scenarios to replace majority of us, it just need to do 99% of what we repeatedly do, much better than us.

If you watch the vids and read the blog posts and re watch it and reread it. You end up with just more and more questions. The bot learned to raze from the Fog Of War based on playing a pro once or twice? It changed its build based on losing once? (sure you say you white listed it but it still had choices no?). Wand not being a common opening is also not true. It's common AGAINST heavy casters. (I wont rush wand AS a bat rider but I would AGAINST him. I would AGAINST shadowfiend as well.) Was the animation canceling also taught similar to the creep block? independent RL?).

So many things don't fit in.. but that might be because I am neither good at DOTA 2 nor at AI stuff. Good luck with the 5v5's openai. And if the AI bots become public, maybe i'll get to play against them some day.

This is a highly editorialized title that clearly doesn't follow the posting guidelines:

Otherwise please use the original title, unless it is misleading or linkbait.

Please don't do things to make titles stand out ...

(the original title is the much more sedate "More on DOTA 2"; in case it is changed by moderators, at the time of this comment it says "How We Built the Unbeatable DOTA 2 bot", which is not even a quotation from the article)

The submitted title has certainly shaped the commentary (see people reacting to "unbeatable", and whether it's an impressive achievement, vs reacting to the technical content).

So, in a nutshell, I think we should avoid titles like this, and I believe the guidelines transparently say so as well.

This is an unquestionable technical achievement, but its important to keep in mind that the AI view of the world is through the bot API. This is compared to a human that decodes game state from observing the monitor. Not being an avid DOTA2 player I marvel at skilled players that can follow the game based on what looks like a very busy and colorful mess to me.

At the same time most of the deep learning field has been focusing on implementing super-human perceptual abilities (vision, hearing, translation) that come instinctually to humans. The higher level reasoning and memory/attention augmented machine learning is still cutting edge research. I think DeepMind and OpenAI are driving research towards that end.

By "unbeatable" they mean it lasted a few hours in the hands of the public before the hardcoded behaviors were exploited and it was defeated consistently. Turns out hardcoding the mechanism for players to manually block creeps (a more "creative" play style, but a fundamental element of lane control) allowed for the exploit. The bot did not teach itself to control the flow of creeps to gain a laning advantage against the opponent.

Not to diminish the teams' achievement. This was awesome. But again, it's just more over-hyping of AI that should be called out for what it really is.