# OpenAI’s New AI Learned To Play Minecraft! ⛏ ## Метаданные - **Канал:** Two Minute Papers - **YouTube:** https://www.youtube.com/watch?v=fJn9B64Znrk - **Дата:** 09.08.2022 - **Длительность:** 7:21 - **Просмотры:** 137,263 - **Источник:** https://ekstraktznaniy.ru/video/13489 ## Описание ❤️ Come work for Weights & Biases! Check out open roles at https://wandb.me/jobs ❤️ Check out Weights & Biases and say hi in their community forum here: https://wandb.me/paperforum 📝 The paper "Learning to Play Minecraft with Video PreTraining (VPT)" is available here: https://openai.com/blog/vpt/ ❤️ Watch these videos in early access on our Patreon page or join us here on YouTube: - https://www.patreon.com/TwoMinutePapers - https://www.youtube.com/channel/UCbfYPyITQ-7l4upoX8nvctg/join 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Eric Martel, Geronimo Moralez, Gordon Child, Ivo Galic, Jace O'Brien, Jack Lukic, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Kyle Davis, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owe ## Транскрипт ### An AI playing Minecraft? [] Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today we are going to see what happens if  we unleash an AI to watch Minecraft videos   and then, ask this AI to play  it. And the big question is,   can it play as well as we, humans do?   Well, you’ll get your answer today! ### NVIDIA's AI solution [0:22] This is OpenAI’s project, which is not to be  confused with NVIDIA’s similar project called   MineDojo. So, what was that? That  was an AI project that watched a ton   of Youtube tutorial videos on Minecraft.   Minecraft is an open world sandbox game,   one of the most played games in the world, and  this is a true sandbox, which means that you   can do whatever you wish. You create the story of  this game. Gather raw materials, build something,   explore, or don’t, it’s all up to you. This is  a sandbox game, and your story after all. So,   what could NVIDIA’s AI do after watching  these tutorial videos? Well, a great deal. For instance, we could even use natural  language and ask it to explore an ocean ### Exploring underwater (NVIDIA) [1:22] monument. I really liked this one because  the word “explore” is sufficiently vague   and open ended. A human understands what exploring  means, but does the AI understand it too?    Well, as you see, it does! Amazing. It could  also encircle these llamas with a fence. ### Building a fence (NVIDIA) [1:39] Or, it could also be dropped into  an arena with an Ender Dragon, ### "Fighting" an Ender Dragon (NVIDIA) [1:45] the final boss of the game, if you will. However,   look, this is a very limited setting for this  task as the dragon is not charging at the AI.    But at the very least, NVIDIA’s AI  understood what fighting meant in this game. ### My wish is granted! [2:06] Now, I said the following, quoting: “Starting  this game from scratch and building up everything   to defeat such a dragon takes a long time  horizon, and will be an excellent benchmark   for the next AI one more paper down the line.   I’d love to see it perform this start to end.    Make sure to subscribe, if such a paper  appears, I’ll be here to show it to you. ” And I thought, maybe a year or two later there  will be some activity here. And now, hold on ### The OpenAI project [2:40] to your papers, because OpenAI’s scientists  are here with their own spin of this project,   not one year later, but just one week  later. Wow! And, with this paper,   I also got what I was looking for, which is  an AI that understands longer time horizons. ### Building a wooden shelter (OpenAI) [2:56] They make a huge claim. They say that this is  the first AI that is capable of crafting Diamond   tools. That is very ambitious. Can that really  be? Well, I say that I will believe it when I see   it. Do you see this long sequence of subtasks?   We need to do all this correctly and in order ### What? A Diamond pickaxe? [3:19] to be able to perform that. This takes up to 24000  actions and 20 minutes even for a human player. ### Building the Diamond pickaxe (OpenAI) [3:34] So, let’s see if the AI is also up for the task!   It starts chopping wood right away, then, the   planks are coming along. Later, it starts mining,  finds iron ore, crafts a furnace, starts smelting,   and at this point I am starting to think that it  might make it. That would be absolutely amazing.    However, we still need some diamonds. Can  it find some diamonds? It takes a while,   and…oh my, got ‘em! And from then on, we have  established that this is a smart AI, so from   here on out, this is a piece of cake. There  we go. I can’t believe it. A diamond pickaxe. ### Pillar jumping (OpenAI) [4:24] And here is my other favorite, it can also perform  pillar jumps. This is a cool technique where we   jump, and while mid-air, we quickly put a block  below us, and tadaa. We are now standing a little   higher. If we do this over and over again, we will  be able to reach places that otherwise would be   way out of reach. And the AI  learned this too. How cool is that? ### How does it learn? [4:55] While we are looking at what else it is capable  of, I would like to bring up one more remarkable   thing here. Here goes: it learns from unlabeled  videos. In other words, this means that the AI is   not told what these videos are about, it is just  allowed to watch them and find out for itself.    And to be able to learn this much information  from this kind of unlabeled data just is an   incredible sign for future projects to come.   You see, playing Minecraft well is great, ### Not just for games! [5:24] but this is an AI that can learn new things  from unstructured data. In the future,   these agents will be able to perform more general,  non-gaming related tasks. What a time to be alive! For instance, the amazing DeepMind lab is  already walking this path. First they wrote an AI   that could learn to play Chess, Go, and Starcraft  2, and then, they used a similar system to solve   protein folding. Now, make no mistake, the  protein folding system is not exactly the   same system as the chess AI, but  they share a few building blocks. And here, having two amazing Minecraft AIs  appear within a week of each other is a true   testament to how quickly projects  are coming along in AI research. ### First Law of Papers [6:20] So, let’s apply the first Law of Papers  - what do you think this will be capable   of one more paper down the line?   Let me know in the comments below! Thanks for watching and for your generous  support, and I'll see you next time!