1.	Introduction
	Artificial intelligence has been around since the start of video games. Features such as NPC (non player character) behaviour and basic pathfinding [1] were the main feats of AI in video games early on. An example of this would be the game Pac-Man (Namco, 1980), where the player must navigate a maze and collect all the “points” while ghosts (NPC’s) would try to chase the player down. At this stage of game development and implementing AI, the NPC behaviour was not very advanced at all, but despite this the game was still widely popular and successful, so a demand grew for more intricate and complex games. With this demand the AI implemented gradually became much more advanced and sophisticated. For example dynamic difficulty adjustment was introduced in the Halo series (MS Game Studios) and Half-Life (Valve, 1998) feature complicated tactics that the NPC’s would employ to try and “kill” the player, where the NPC’s came to those decisions through techniques such as decision trees and reinforcement learning. Recently the NPC behaviour in video games has reached very sophisticated levels as in The Elder Scrolls V: Skyrim (Bethesda Softworks, 2011) where the NPC’s would react to a seemingly limitless amount of choices that the player could make in a completely open world and boundless environment. Despite artificial intelligence making progress in video games, up until recently it was never considered very advanced or even helpful in progressing AGI (Artificial General Intelligence).

2.	Artificial General Intelligence and its ethical issues


AGI can simply be defined as a super intelligence where the machine is considered to be just as smart (or smarter!) than a human [2]. This is different to AI which refers to a machine basically imitating decisions a human would make while having access to a large amount of data.
The AGIRI (Artificial General Intelligence Research Institute) states that its goal is to create human-like and transhuman systems that can operate in undefined domains compared to smart systems that only work in narrowly-defined domains [3]. But what is artificial general intelligence? AGIRI lists features such as: superhuman intelligence, self-reflective and can come up with its own fundamental insights. All of this sounds seemingly impossible but can be done with autonomous agents. An autonomous agent has features such as: can survive and adapt to an environment, has an agenda of its own and is willing to act on this agenda [4]. By definition, all humans and animals are autonomous agents. This of course leads to various ethical issues and implications. A particularly heated issue right now is that what if these autonomous agents implemented by AGI, create an agenda of their own that does not serve the purpose for which we created it? What if it decides to “rebel” against its creators? Although I personally think that a lot of this talk is just fearmongering, one can see their point. However, to address this, people developing AGI have placed an important stake in developing these systems with certain risk reduction principles in mind. A “Golden Rule” approach is often utilized when developing these systems which can be summarised to humans treating AGI’s as we would want to be treated ourselves and vice-versa.
Although the “Golden Rule” is a good approach in human relationships it is a very vague approach to developing “safe” AGI. Due to this, other approaches such as Asimov’s “Laws of Robotics” were introduced [5], which have been proven to be flawed but still a right step in the direction of creating safe AGI. Currently a lot of research is underway to create better laws that AGI must follow to reduce risk, but as that is no easy task and creative new approaches are being thought of and developed.

3.	AGI in traditional board games
Since research on AGI started, traditional board games have been a great way to create projects and further research how AGI could be implemented to a more “real life” scenario.
Everyone knows about computers beating grandmasters in games of chess, famously Garry Kasparov (World Chess Champion and Grandmaster) versus IBM’s Deep Blue. The first match was played in 1996 and Deep Blue was the first computer program to defeat a world champion in a classical game of chess. A rematch was later played in 1997 and was the first time a computer program defeated a world champion in a match under tournament regulations. The win was a significant turning point in the general perception of AI and showed that a computer program could beat humanities great intellectual champions [6]. Despite it being an impressive feat, people criticized Kasparov for bad play [7] which was not usual for him. Then chess was criticized as a game that can be defeated by “brute force” [8] as despite there being a large number of possible moves and board states, there is a common “meta game” of what is a good play to make and what is bad play to make (assuming your goal is to win the game). Although strictly speaking IBM’s Deep Blue was AI and not AGI, it still generated a lot of talk and publicity for AI which in turn allowed people to start trying to make programs that could defeat human world champions in games much more complex than chess. These more sophisticated programs would have to be AGI and not just regular AI. The most famous example of this would be Google Deepmind’s AlphaGo project.
Go is an ancient Chinese game that has simple rules but hundreds if not thousands of times more possible moves than chess. It is considered to require more thinking and intuition than just simple brute force to win. Google Deepmind created a project called “AlphaGo” to try to beat a world champion in a game that was not believed to be able to be cracked by a machine. In 2015 it managed to beat world champion Lee Sedol in several games of Go and shocked the tech world. It was praised as being very effective as it had learned millions of moves from past masters of the games, that it implemented during the match and even managed to find new win conditions when faced with problems. However the real breakthrough is when Google later made AlphaGo Zero which had been instructed with the rules of the game but given no other instructions other than to win. It only took three days for it to surpass all versions of AlphaGo and within just over a month, it had on its own developed extremely complex and advanced strategies that had taken humans thousands of years to discover. It even managed to develop its own strategies and had been described as having “genuine moments of creativity” [9]. This is where we can see just how much more powerful AGI is compared to AI.
The implications of this achievement lead to the idea that AGI could be applied to health problems such as finding cures for diseases in less than one hundredth of the time it would take humans to find. Here is where AGI in games could be the key to maybe finding a cure for things such as Alzheimer’s and DeepMind co-founder and CEO Demis Hassabis believes that if the techniques AlphaGo Zero learned could be applied to real life structured problems then the human race could ultimately be impacted much more positively than would be possible without AGI [9]. Although AlphaGo Zero is undoubtably very impressive, there are harder games that AGI could be applied to, and then researched which could lead to better AGI, in turn leading to new discoveries that could benefit the entire human race.



4.	Open AI

Taken from their website, Open AI is non-profit organization whose purpose is to create AGI that benefits all of humanity [10]. It has notable founders such as Elon Musk (Founder of SpaceX and PayPal). They have created many projects such as the OpenAI Gym which is a toolkit used for developing and comparing reinforcement learning algorithms, teaching agents to do everything from walking to playing games like Pinball [11] and they have also released reports in collaboration with Oxford and Cambridge, preparing for the possible malicious misuse of AI technology, with key points such as: Acknowledge AI’s dual-use nature, learn from cybersecurity and broaden discussion about possible malicious use [12].
What I’m particularly interested in is Open AI’s involvement in the popular online game DOTA 2 (Valve, 2013).
5.	Open AI Five
DOTA 2 is a video game of immense strategic skill, teamwork and individual player mechanics. It is in the top 10 [13] of most played video games in the world and the most well payed game for professional E-sport players to compete in, in terms of tournament prize pool money (the top 39 highest earning E-sport players are DOTA 2 players [14]). It consists of two teams, made up of 5 players. Each team bans certain heros (characters available to play) and then picks one for each player. There are hundreds of items available to buy in the game and hundreds of different ways to level up your character. The game is over when one team destroys the other teams ancient (base). Each year there is an “international” that is held (equivalent to the world cup in football) and is considered very prestigious to win. For the last two years each player on the winning team would win two million dollars each from just the prize pool money alone.
In August 2017 at the DOTA 2 international, there was a show match on the main stage of Dendi (a previous international winner) versus the one Open AI bot in a 1v1 match. The Open AI bot beat him convincingly and the crowd was shocked. They also put the bot to test versus other pro players and it was undefeated versus all of them including, the at the time best 1v1 player in the world and the top overall player in the world [15]. The Open AI bot received a lot of praise for its accomplishments but as one versus one match-ups are more dependent on the mechanical skill of the player instead of theory crafting and strategy, many people said that its not as impressive as winning a real match five versus five.
Open AI then planned to beat the best teams in the world in a fair match. They first started by playing each other in very restricted five versus five matches, and every time it would win from doing something it would learn and adapt. Eventually in April 2018 a team of five bots beat a team of real humans in a restricted game, which gained so much attention that even Bill Gates tweeted about it [16]. They kept allowing the bots to play each other and made the game less and less restricted up until the point where they played the best team in the world in August 2018. The game was livestreamed and hundreds of thousands of people watched five bots compete against the top players who had just won that years International. The bots beat them convincingly and showcased that they played the game better in one year than these humans who have played DOTA 2 (and its predecessor Defense of The Ancients) for over a decade. Not only was their individual mechanical skill immaculate, but their teamwork was considered perfect and they made almost no mistakes. Most outstandingly in such a short amount of time, they had figured out and improved strategies that had taken a decade to be discovered by humans, and the most impressive thing is that they learned it from scratch, trial and error, no previous data about the game or its winning strategies was available to them.
	It was one thing for AI to beat world champions in chess and then another breakthrough for AGI to beat world champions in Go, but to defeat world champions in a game as intricate, thought provoking, challenging and difficult as DOTA 2 is something completely else and is a testament to how AGI implemented in traditional games and video games can help progress AGI at rapid speeds.

6.	Conclusion

AI and AGI is now consistently beating human world champions in extremely difficult games, that a decade ago, no one would have said they would be able to play, let alone win. This is huge progress in the field and could possibly lead to AGI that benefits the whole of humanity. If the AGI applied to these games can be applied to real life scenarios then the possibilities are endless and seemingly impossible tasks such as finding a cure to cancer may be possible.