This AI learned to play ‘Minecraft’ after watching thousands of hours of videos

This AI learned to play ‘Minecraft’ after watching thousands of hours of videos

Although there are artificial intelligences that have been scandalous, such as the supposedly aware AI of Google, or the one that had the plan to collect every photo of every human face in the world; they are not all the same. This intelligence trained by OpenAI has shown a much friendlier purpose related to Minecraft.

Following in the footsteps of any thirteen year old, this AI learned how to play Minecraft after consuming an approximate of 70,000 hours in videos from YouTube. This has been reported by the media Eurogamerwho also ensure that this intelligence plays “in the same environment as humans” using the equivalent of a keyboard and a mouse.

This, of course, represents a huge leap in the possibilities of an artificial intelligence. In addition, it is not only capable of walking through the different biomes of the game. You can also cut down trees, build crafting tables, swim, run, huntcook and even craft various objects. In fact, one of his great achievements, according to comments from the OpenAI websitewas his ability to build a diamond pickaxe.

We have trained a neural network to play Minecraft using a video pre-training (VPT) on a huge dataset of unlabeled videos of human matches from Minecraft, while using only a small amount of contractor-labeled data. With detailed tuning, our model can learn to make diamond tools, a task that typically takes competent humans over 20 minutes (24,000 actions).


But how exactly did this AI learn to play Minecraft?

Time and number of actions it took for the AI ​​to complete certain achievements in Minecraft

As the AI ​​performed actions within the game, was able to complete others and faster as researchers put it to the test.

Read:  NostraCongress2022 raised thousands of euros for a hospital in Barcelona

From the aforementioned source, they comment that the contractors recorded and tagged basic Minecraft actions to create an algorithm with them. In this way, the 70,000 hours of video on YouTube could be translated into keyboard and mouse actions in the AI ​​database.

This allowed artificial intelligence to learn from human actions, rather than relying on pure trial and error. Thus, it could predict actions using past and future information that was already stored within its database.

AI gameplay playing Minecraft. OpenAI Property

An AI with a future

Although OpenAI considers this experiment a success, and While the team only tested their AI within Minecraft, this would not be their ultimate goal.. In the detailed information on the web, they add that this learning could also be applied to other types of systems.

That is, this artificial intelligence could be used to learn the general use of a computer. Automating actions that human beings carry out in their day to day, but that could be carried out without problems by a computational algorithm.

Although we only experimented in Minecraft, the game is very open and the native human interface (mouse and keyboard) is very generic, so we think our results bode well for other similar domains, for example, computer use.


We cannot say that the news does not excite us. After all, artificial intelligence has been created to help and complement some tasks of humanity.

An AI capable of executing basic computational actions would not only be useful when managing time more efficiently; It would also be excellent news for those people who have a disability, and who often depend on accessibility in operating systems.