So the questions surrounding artificial intelligence extend to the whole of society, they particularly affect the world of video games, which has been accustomed to using this technology for decades. But what if instead of serving as an opponent, the AI helped players improve? It is from this reflection that this player used this technology to make incredible discoveries in Super Mario 64.
Since the beginning of the 2020s, artificial intelligence has made immense progress to the point of becoming democratized and becoming an everyday tool. If this raises a lot of questions about its use in society, players have been used to this technology for decades because it is the latter which allows them to play alone “against the computer”. But beyond being useful to developers, what if AI became a tool for players to improve? This is the reflection that it was made by an experienced Super Mario 64 player who made incredible discoveries thanks to this technology.
Artificial intelligence as a speedrunning tool
In a twelve-minute video, the user Memario traces on YouTube the research of Krithalith, a player specializing in Super Mario 64 who offers TASi.e. Tool-Assisted Speedruns. For ordinary people, it is a performance whose goal is to complete video games as quickly as possible, except that the “Tool-Assisted” dimension emphasizes that we can use tools to execute series of games. actions often impossible to perform by a human. With TAS, we can obtain the theoretical minimum time needed to complete a game. It is therefore a whole category apart from speedrunning since it is not a live performance, but rather the result of long work over several months to optimize the smallest aspect of a run.
With all that, Krithalith is therefore a Super Mario 64 TASeur known in the gaming community for having found a setup to achieve the Carpetless bug which saves almost a minute in TAS, an incredible advance in this environment. Beyond being an excellent TASeur, Krithalith is also a good programmer and that's why he came one day to develop an AI to teach him how to play Super Mario 64 like speedrunners and TASers do.. By making it ingest all the knowledge surrounding the game over more than 28 years, this artificial intelligence plays like a Super Player, even if it does not manage to do as well as the latest TAS. This is particularly felt in the stars that take longer to recover where it is really in decline. Despite these shortcomings, Krithalith saw the potential in this technology that took routes the speedrunning community had never considered. And clearly, he did well.
When AI does better than players
To illustrate the different progress of this AI, three levels are particularly telling. First of all, Krithalit chose the In The Deep Freeze star which is one of the fastest to obtain, in 5'23 seconds. By itself, without any specific indication, the AI ended up touching the star in 5'33 seconds through many attempts, but without being able to go further than that.. Finally, its creator indicates a slope to take on the left to obtain a slight gain in speed and which is the path taken by the TASeurs. To everyone's surprise, the AI performed less well than its previous tests, arriving at a result of 05'47 seconds. Despite this, what must be remembered is that artificial intelligence has reached the same level as the old TAS, that is to say 5'33 seconds, without indications.
It's the same thing on the level Shining Atop the Pyramid which is a star which is recovered quickly, in 5'87 seconds, but which requires many complicated movements. While the artificial intelligence initially does anything, Krithalith simply tells it the route taken by the current TAS and in just 12 hours, she managed to achieve the same result as the TASeurs who obtained this result in 18 years. Except that the major difference… is that the AI made different movements to get there! Certainly, the path is similar, but she does not press the same keys at the same time, which, in a practice as precise as the TAS, is far from trivial.
With this same logic, the AI ends up doing better than TASers on certain levels. This is for example the case in the level The Manta Ray's Reward where she manages to gain 0.22 seconds compared to the current TAS record, which is quite simply an achievement in an environment where gaining 0.1 seconds can take time. month. To achieve this, artificial intelligence thought of what had never crossed the minds of TASers : press the swim button in the eighth frame instead of the first because this allows Mario to spin less and therefore go faster. It must be said that it does not go without saying to think like this because we would tend to think that pressing at the first frame would be faster. But on its own, the AI came to an unprecedented conclusion, which offers a new field of possibilities for TASers.
With this example, Krithalit proved that artificial intelligence could also be used as a tool to improve in-game performance, even if it only concerns high levels such as speedrunning. We were thus able to discover an unexpected use of AI: it can be used to expand the horizon of knowledge by trying things that no one would have thought of before. It now remains to be seen whether this way of doing things cannot be applied to more general public approaches, in order to further improve the comfort of players.
