NVIDIA’s New AI Mastered Minecraft 15X Faster!

Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. After an incredible day of Apple showcasing their  culmination of many-many amazing research papers,   and also light transport algorithms applied to  soundwaves. Today we are going to have a look   at NVIDIA’s new invention that can play Minecraft,  but not in the way you think. It truly is a sight   to behold as it is able to gather resources for  itself, go and mine, catch a fish, these things   I expected. However, what I didn’t expect is  that can even engage in more complex tasks like   building a base, or hunting this adorable  little piggy that kind of breaks my heart. Now, it can do 3 more complex things, including  fighting, which I will show to you in a moment.    But let’s have a look at the paper, and  immediately when looking at the article,   I was like “what? ”. Are you sure? You see, this  paper claims to use large language models to   play Minecraft. How is that even possible?   Large language models are typically adept at   answering text-based questions with text-based  answers. So how does it play a computer game? It doesn’t just play the game, but it plays it  incredibly well. How? Well, hold on to your papers   Fellow Scholars, because it is able to build  a curriculum by itself, for itself. You see,   it gets some text information about the world  it is in, for instance, the items that it has   or what time it is and who is nearby. And based  on this information, it starts reasoning by   itself. That is incredible. Look, for instance, if  we have a wooden pickaxe and some stones, then it   is a good time to upgrade to a stone pickaxe. Or,  if it’s nighttime and there is a zombie nearby,   it is time to grab a sword and shield. Now, these  are simple statements, but the key is that this   little AI makes these deductions by itself. We  don’t need to program these by hand, and thus,   if it were dropped into a completely different  game, it would likely do quite well there too. But wait a minute, this is still just  text. How does it become gameplay? Well,   after creating this curriculum, the AI writes  computer code to control the game and achieve its   goals. If this piece of code works, it observes  what it achieves, and then stores it away in   a skill library. These will be the building  blocks, for new, more complex skills later. So a text-based chat assistant that  can play a video game. That is very   impressive. So, how well can it play the game? With a little help in the form of human  feedback, it can do three amazing things.    One, it can build a house. And I have to  say, that is some proper craftsmanship there,   or proper craftsAIship. Lovely  house. Good job, little AI! Two, it can also build a nether  portal. This allows it to travel   to a different dimension with  unique terrain and resources. And, get this, it can even fight an Enderman.   This creature makes short work of an unsuspecting   player, however, the AI came well prepared with  high-quality equipment, and bam, we are done. Now, these all sound good, but we are Fellow  Scholars here, so we are looking for a detailed   comparison of this technique against previous AIs.   Is this any better? Oh my goodness. Are you seeing   what I am seeing? It can explore so much more than  its predecessors, for instance, much more than the   fan favorite AutoGPT, a version of ChatGPT that  can prompt itself and work on its own. And it   gets better! Wow! Look at that. I can’t believe  it! It is more than 15 times faster than AutoGPT.    AutoGPT has barely come out a few weeks ago, and  it has already been improved 15x already. Wow. Just look at that! What took AutoGPT 75  iterations likely only takes maybe 5 iterations,   likely even less. AutoGPT plateaud at iron tools  where the new technique is barely getting warmed   up. Incredible improvement in just a couple of  months. These text-based large language models

can not only write code to play video games,  but they can even plan their journeys. And just   imagine what they will be capable of just two more  papers down the line. What a time to be alive! Thanks for watching and for your generous  support, and I'll see you next time!