VPT: the new OpenAI AI learned Minecraft in 70,000 hours on YouTube – LeBigData.fr

OpenAI’s new VPT artificial intelligence learned to play Minecraft video game after training on 70,000 hours of YouTube video. This innovative approach could offer new possibilities for AI…

We distinguish two types of video game lovers : those who play it, and those who watch others play it in streaming. However, the new artificial intelligence from OpenAI manages to reconcile the two clans…

Already in 2020, OpenAI made an impression with GPT-3. After being trained on millions of texts extracted from the internet, this algorithm by Machine Learning is able to produce real sentences.

Subsequently, the firm also created the buzz with DALL-E in 2021 and its successor DALL-E 2 in 2022. This GPT-3-derived AI has been trained on both text and images, allowing it to generate your own visuals realistic.

This June 27, 2022, OpenAI unveils his latest artificial intelligence became an expert at the Minecraft game after watching 70,000 hours of video uploaded on YouTube by players…

The first AI released in the game as a human player

Many AI algorithms have been trained playing Minecraft game in the past. However, they were tested in a simplified version of the game.

For its part, the new OpenAI AI, VPT (Video Pre-Training)plays the same version of the game as humans and uses standard keyboard/mouse combo controls.

In a published post on his blogOpenAI explains that the algorithm first learned the most basic skills of Minecraft: cut trees, make floors, make a table…

The research team also observed the AI ​​attempt to swim, hunt, cook and jump on the pillars. In short, the algorithm behaved like a human player discovering the game.

According to OpenAI, “ to our knowledge, there is no no published study on AI operating in the full, unaltered human space of action, including drag-and-drop inventory management and item crafting “.

AI Advances Rapidly with Reinforcement Learning on 720 GPUs

Through training on a specific dataset, also called “fine-tuning”, the model then performed these tasks more accurately. Besides, he then started making wood and stone tools and basic shelters, exploring villages and looking for treasure chests.

The AI ​​was then trained via the technique of reinforcement learning or Reinforcement Learning. This method allowed him to learn to make a diamond pickaxewhich usually takes 20 minutes for human players.

This result is convincing, because artificial intelligence has long experienced difficulties in the face of very free Minecraft gameplay.

Conversely, AI has surpassed humans at chess games and Go through reinforcement learning. Because, these games have clear goals towards which progress can be measured. The algorithm can be rewarded for each progress made towards the goal.

In Minecraft, the objectives can be multiplethe progression is less linear, and reinforcement learning algorithms are usually caught off guard.

In 2019, during the MineRL competition between AI developers, none of the 660 AI candidates managed to achieve the objective of the competition: to mine diamonds. Note, however, that participants had to limit themselves to single GPU Nvidia and 1000 hours of gameplay video to train their algorithms.

The intention of the organizers of the competition was to show that creativity was more important than computing power. However, OpenAI’s new AI has been trained on 720 GPUs and 70,000 hours of video. This is the reason why its performance is much higher…

70,000 hours of YouTube video as training data

For training this new AI, OpenAI used the same approach as for GPT-3 and DALL-E. The algorithm was trained on a massive dataset consisting of human-created content.

However, this success is not only based on the immense volume of data or the colossal computing power used. In general, raw video clips are only not as effective for behavioral type AIs than for content generators like GPT and DALL-E.

This type of video shows what people do, but does not explain how. To associate the images with the action, so the algorithm needs labels.

For example, if the video clip shows a player collecting items, an “inventory” tag will be needed. for the algorithm to associate the images to the corresponding class of shares. It must also be indicated which key on the keyboard opens the inventory.

Obviously, it is not possible to manually label every frame of 70,000 hours of video. The researchers therefore called on subcontractors on Upwork for recording and tagging the most basic actions in Minecraft.

They then used 2000 hours of video to teach a second algorithm called IDM how to tag minecraft videos. It was this second algorithm that was responsible for annotating the 70,000 hours of YouTube content.

This approach could allow artificial intelligence to learn a wide variety of new skills watching videos on the internet. OpenAI researchers imagine, for example, the use of VPT to teach computers to perform all kinds of actions on voice command.

A process that is far too expensive for AI developers

Unfortunately, at present, this method of training is beyond the reach of simple AI developers. Besides the cost of the 720 Nvidia GPUs and their power consumption, the Upwork contractors responsible for labeling the videos cost $160,000.

And yet, this AI model is relatively small. It only has a few hundred million parameters, compared to several hundred billion for GPT-3.

It therefore seems urgent to finding ingenious new approaches limiting data and computing power requirements. While the AI ​​needs 70,000 hours of video, one or two videos are enough for a child to learn the basics of the game.

Either way, OpenAI has put the data, environment, and the open-source VPT algorithm. A partnership with MineRL has been established, and participants in this year’s competition will be able to freely use, modify and configure this AI…



We want to say thanks to the author of this short article for this amazing content

VPT: the new OpenAI AI learned Minecraft in 70,000 hours on YouTube – LeBigData.fr


Check out our social media profiles and other related pageshttps://yaroos.com/related-pages/