So open AI requires Global illumination now global illumination has some Powerhouse Engineers on its team and this could be seen as a little bit of a aqua higher in acquisition to hired to get some Talent on board and that’s probably a part of it but part of that Global elimination thing the holding Company or whatever Underneath It All you do have a bit of an asset that might be very interesting to open Ai and that is biomes or biomes.gg as they say Gigi so it’s an open source sandbox MMORPG built or the web buildforge play mini games and more all right from your Browser why is this important it does seem like all the sharks are swimming in the same direction there’s a bit of a Feeding Frenzy right now everybody’s digging in the same direction and in this video we’re going to take a look at why that might be happening I think it’s Clear that Ai and games are beginning to intersect more and more at first when chaji PT came out and I started doing some videos every time a research paper mentions some games I was all over it I mean Minecraft AI I was one of my favorite videos to do it’s one of the Most unwatched videos then around the same time maybe even before that the Stanford paper comes out talking about the 25 generative agents similacra of human behavior and that was also kind of a video game I initially did a little video on it when it just came out there’s some more news that are Surrounding this paper we’ll talk about in just a second because I think a lot of people are still missing the full impact of this of of what they did here we’ll we’ll talk about that in just a second this page and I’ll link it down Below but so this page you can see this thing kind of running live so this this has been pre-computed so they ran the simulation over however many weeks or months they did it for and this is sort of sort of like a replay so this already happened there replaying it I mean I Guess that’s kind of obvious by the date that they’re showing here but here’s basically you can scroll down you can see the actual state of these agents so LaToya Williams taking a yoga class and kind of a description of the location kind of like where she’s in current conversations there’s none this person Is talking talking let’s see what their current conversation is because you can actually see I guess they’re not talking let’s see they stop talking this cat is writing so I show one of these guys having a conversation you’re saying not at the moment all right here we go this Is one of the conversations so as you can see here it’s a full conversation hey Carmen good morning how are you how are you Etc but my point here is that a lot of the AI and video games they’re intersecting a lot of the research on AI Seems to be going more and more into video games there’s a number of reasons for that training robots in a simulation I mean if you’ve been following what Nvidia is doing looks like a lot of what they’re trying to do and and deepmind Google deepmind is doing something Similar is their training robots and almost these little simulations where they kind of like like you might have seen something like this where these like little robots play um soccer these robots that are pre-made and they do have like an algorithm there that they come out and they start playing right Out of the box they have some moves and stuff like that but it’s kind of pre-scripted and so sort of the new thing that deepmind is doing and a lot of other people Nvidia Etc is they’re throwing these little things in simulations where they learn how to play And they kind of develop their own moves and stuff like that and basically the point is that the things that come out of this simulation versus the pre-scripted things that like are kind of ready to go out of the box like the simulation robots just destroy these Things that come with their sort of basic knowledge that some human coder wrote down it’s like if this then that you know kick the ball whatever right instead they just throw simulation it runs around and therefore Sears or whatever however time flows in that space and also one thing that they do is You know in a video game in a simulation what you kind of like by default what you expect is everything is like perfect everything’s even like there’s not too much variety right there’s not different surfaces or maybe different wind conditions or whatever that’s kind of different from The Real World where you Can have tons of things go wrong for example simple as let’s say the current in one of the robotic hands is a little bit less than sort of what the testing conditions were right so maybe it’s not gonna be able to grip us hard or if it’s Running maybe one leg is just slightly not as strong as the other or how it was in the simulation so to sort of counteract that by the way two minute paper this guy is great he goes over like a lot of this stuff and kind of Deep Dives he’s the guy that goes what a time to be alive like he says that every time and it’s great it’s becoming kind of like a catchphrase but my point was that in order to kind of create a little bit more Randomness in order for for it To be dealing some random environments like different friction or different power or maybe a little bit of like a wind and stuff like that they actually added a little bit of a Randomness to these games so like the friction on the floor changes a little bit or the wind Conditions or whatever all the little factors that kind of go into this get randomized just a little bit to kind of prepare it for the real world quote unquote but the interesting thing that came out of that is it seems like that Randomness and stuff that they added to It actually when the robots come out of the simulation they tend to be a little bit more robust they tend to be a little bit stronger so that whole randomization is actually like a kind of a big benefit because Dora’s robots not only tend to get better at playing perfect conditions But just overall they they get a lot better at dealing with more situations and more environments and stuff like that and so Nvidia and open Ai and deepmind they all kind of have the same idea the idea being that a lot of the training that robots can do and same Thing with stuff like gptl a lot of stuff these neural Nets you know it requires a lot of data a lot of simulations and running robots and stuff like that in the real world is expensive you know they break down they cost electricity Etc but the more of this Training that you can do in simulations the better saves you time it scales very effectively Etc and so I think this idea is really getting traction and I think we’re going to be seeing more and more of this stuff developing and one of sort of the obvious ways to get there faster Is to use games because that’s basically what a simulation is that a game is a simulation of the world so you can create a game that simulates various environments and then throw you know whether you’re training robot or AI like Chad GPT that is trying to reason answer Questions or stuff like that is so much faster and so much easier to just kind of throw them in there and have them learn from that world autonomous coding agent AKA how to craft code in Minecraft so here’s the Minecraft AI video that I did I apologize if this is getting a Little too met up but I figure this might be the easiest way to show people what’s going on so they have this like automatic curriculum where one gpt4 basically figures out what the next step that they need to be doing so like as you can see here if it’s next to a river And it’s got you know a fishing rod and it’s like oh let’s catch a fish and then the next big thing that happens is the skill library and this is where another sort of instance of gpt4 actually codes itself new skills that it can do and it Uses the API the Mind flare API to communicate with Minecraft but for example here let’s say encounters a zombie so it figures out how to how to deal with the zombie so it crafts and equips a shield it recovers hunger so basically like raises its health and Then it looks for the zombie and engages in combat with the zombie right so it writes an actual script like it codes a little script for itself of how to do that right and it tests it and if that script works as intended well it gets Added to the skill library and the skill library is sort of a list of all the things that this thing can do in the world of Minecraft right so it can mine various logs do you mind logs so it’s probably chop chop wood not minewood log but whatever make a crafting table craft Pickaxe whatever make furnace Etc cook steak combat zombie right and then as it goes through the world it tries to find what it counts certain situations you know what I mean it tries to find relevant skills to use in that situation and because of this it has a very Interesting sort of like learning curve or progression curve whatever you want to call it so as you can see here with a lot of the other sort of state of the art boss that learned to play Minecraft we see if a lot of them is is at some Point it hits a plateau like it’s able to do a few things and it just stops learning it stops improving and in fact this blue line is the same thing the Voyager AI but without out the skill Library so without the ability of make those skills and add it to a sort of Bank of skills that it can use it also plateaus over here but the full Voyager of all the architecture of all the things at it like it just keeps going and going and going it keeps improving keeps getting better now the big thing here to me was this right here so they Call it the iterative prompting mechanism so basically what it is is this Bond tries like I said to write to actually code certain skills for itself once it does that it it tests it so it has these little unit tests that it runs and this paper does describe in detail Exactly how they did that it’s I wouldn’t say that it’s easy or whatever it’s certainly not an easy thing to do but it’s it’s less complicated than I’d expect because they are really leaning heavily on gpt4 as the sort of brain power behind this so here’s kind of so The prom gpt4 to write the code they’re like hey you’re going to write this code here’s how you write this code they give it some things examples of various code that has been pre-written already so they give example when they say a few shot learning or multiple learning or Whatever versus zero shot learning so zero shot learning is where you just ask it do something with no examples right few shot learning means you you give it a few examples here’s a similar piece of code here’s another similar piece of code then gbt4 kind of looks and it’s Like okay so based on this you want me to do something like this I got it and it writes it so they do that here represents you have some like example codes from mind flare that people have already completed Etc one interesting thing here is that a lot of people Sometimes they’ll ask well do these AI agents do they have agency like what happens if no one’s prompting them like are they just gonna stop like how are they going to know what to do I mean here’s an example this is a fully autonomous AI agent kind of prompts Itself because the API every minute or over a few set intervals let’s say it’s like every minute it tells gpt4 where it’s at it’s like you’re in this biome and you have this and this is your health so it’s kind of a script that just like gives it some information About its surroundings right so it goes like hunger position and what kind of equipment it has uh what kind of inventory you have you know there any chests nearby what kind of tasks you have et cetera et cetera here’s specifically what it says so it gives the agent its current state with the Inventory and biome and time whether it’s noon or sunset or night or Etc you know it’s 3D coordinate position Etc and gpt4 it’s prompt you know as it’s prompted it replies to that with the next thing that it’s supposed to do and there’s this other gpt4 right here that Codes the skill so the question is what happens when it’s no longer Prime this this is going to run forever there’s no reason for it to stop because it’s like every minute it gets an updated location but basically the script it just prompts it and then it runs and it keeps going And there’s another thing that’s able to teach you new skills and this is an example of the fully autonomous agent that improves itself but and here’s the thing that I don’t think we’ve talked about on this channel yet but I think this is going to be kind of a big deal For how AI is used for coding in the future for example because a lot of people right now are using it to code and they say hey write me a script for this write me script for that and that works really well because gpt4 can not Only code a lot of the stuff it can explain code very well but this to me really seems like the next step because you’re not asking it for one piece of code with one limited applications you’re not saying hey write me a code to craft a steel sword and that’s it Nowhere in here does it ask for something specific instead there’s one GPT that is given a broad objective it is to do as many pot you know it’s sort of tasked with open end exploration right so it says go out there do as many things as you can be the best and just Constantly learn and expand and keep growing right and so this thing comes up with tasks for itself right if it thinks it needs to fish it says hey let’s go fish let’s learn how to fish let’s figure out how to do that if it thinks it needs a steel armor to protect itself It’s like we got to figure out how to make steel armor and what are the steps required in that well first we gotta mine that over we’re gonna smell out the ore okay use that you know the iron bars that we we got from smelting The Ore to Turn it into armor you know breastplate helmet leggings whatever right this is the important thing to understand we give it this high level task go be the best and then it goes and then it has these little sub tasks like arm myself with you know protect myself with armor And embracing down into a list of armor pieces that it wants that it needs and then this other sort of gpt4 codes the skills it needs to make that happen so my question becomes if we know this is possible and right now this may be like A little bit of a basic version of it but if we understand this is possible is how is this not going to be the way that coding is done in a few years or in five years now I’ve tried to figure out if there’s like this if there’s a like a Name for this thing already so some people refer to that as bootstrapping like have it lift itself up by its bootstraps like basically let it figure everything else out like the other sort of terminology that I’ve heard referring to this is called meta coding so instead of coding yourself your stroke take a Step back and you figure out how to get this thing to code for you but just imagine how powerful that will be instead of planning everything out instead of writing everything out you just describe in detail to gpt4 or whatever next gpt5 whatever next series is gonna be that’s gonna be capable of Doing this but you just describe in detail what it is what you’re looking for you give it some high level ideas and then it splits itself up into multiple instances and it thinks to the task and one by one it starts going through it starts okay we’re gonna need This you know if it’s a financial Apple we need some apis to hook up to the Wells Fargo app the whatever Capital One whatever all their credit cards Etc we’re gonna need to figure out how to present their budgets in a friendly way we use this to this so it generates All that stuff it can do that already pretty well like if you’ve ever tried brain storming for ideas with Chad gbt it works extremely well for that it can produce more ideas faster without stopping so in terms of brainstorming it’ll produce a million of ideas for you And then this other sort of version of of gpt4 this other instance of it right it’s going to start coding these things up and then it’s going to start putting them into an app and doing these little unit tests and figuring out if it works and it can probably do some user Usability testing by simply like going through it and clicking on Stuff Etc like all this this is not science fiction like here are the sort of the Lego pieces for it we just need to like it can do all of it maybe not at the level it needs to yet but it’s improving Constantly so and we we have all the Lego pieces we just need to like put them together in the right thing and it will start doing this so the idea of telling Chad GPT to write your little script that you some one function script and then copying and pasting that into Somewhere that might seem archaic in a year or two I think the approach that most people are going to take is something a little bit more like this where you tell it the Big Goal that needs to happen create the best financial app out there right now for People that enables them to do X Y and Z and then it goes automatically goes and tries to create the stuff and keep in mind if it’s working online for building an app it doesn’t really need to do this in the real world so it doesn’t need to Do it in quote unquote real time like just like those robots learn a simulation where he said something like 42 years past and 24 hours or something like that like it’s the time is compressed a lot of stuff can be done with these video games with compressed Time so that’s one thing to kind of keep in mind why video games present like such a big potential benefit is then when we’re developing these AI autonomous agents that go out there and build stuff for us maybe they don’t need to open up a browser and do it on your Computer in real time maybe what actually happens is million instances of some video game simulation pops up at a million little gpts go to work try and design this app and it gets tested tested in there if you’ve seen some Black Mirror episodes like none of this These are these aren’t like really new ideas a lot of these we’ve thought about it the before you’re science fiction writers like a lot of the stuff that books have been written about before is now becoming closer and closer to reality with that said creating worlds with a sentence Let’s really fast jump back to this giant of Agents interactive simulacra of human behavior so the big news this week or last week I think it was that this became open source so now you can actually download this so June Sun Park he was the one that announced it so this Is the GitHub for it so you can literally like replicate this with your own agents like you don’t have to have 25 you can have whatever number you want you can program you can give them their own sort of backstory and they’ll go and they’ll execute and I’ll do all that Stuff like this is another one that people have mentioned so this is AI town a16z and recent Horowitz the BC company they’re aggressively moving into this direction as well but the point here is a lot of people are commenting when they see something like this right and they’re seeing people talk about how Like oh this will make games so much better and people like why why does it matter that you know various characters in the game are going to have conversations or or they’re gonna have a backstory like who cares like it’s not why why would we be interested in that That’s not why people play games is to like read about their story or whatever and that makes sense but here’s the Counterpoint to that so let me read this part for you so if an end user or developer wanted to wanted the town to host an in-game Valentine’s Day party For example so by the way this is the goal of these 25 agents what happened was is so there’s 25 of them and one of them was sort of given a task you can think of it like a voice of God that like strongly suggests that she should Do something and so that’s the Isabella and she was told that she needs to plan a Valentine’s Day party at Hobbs Cafe on February 14th from five to seven and that’s it no other character was told anything or scripted or influenced in any way it was just her by herself and She went around and she talked to every single one of the other ones and she didn’t even talk to all of them because some of them like for example this is showing that Jennifer and Maria they didn’t hear it from her they heard about it second hand basically right so she Told Sam and Sam in a completely different conversation another day in a different part of the Town said oh speaking of which Isabella has invited us to a Valentine’s Day party at hops Cafe isabelto Klaus Klaus told Abigail and then Abigail actually mentioned this back to Isabella so the important thing To understand is this so so kind of keep this diagram in mind keep in mind how like complicated it is it’s back and forth and it spans multiple days Etc and all the different characters have different relationships with one another so for example some of them like married Till they see each other like every morning and every evening before they go to work some of them don’t really interact as much but if you wanted to create an event like that in traditional game environments you would have to script tens of characters behaviors manually right so you have to say Isabella goes to character number one and says blah blah blah number two says okay Etc and so like line by line by line by line by line you have to code you have to script out every single interaction and then you have to script all those characters going to that Valentine’s Party and then you have to script all the interactions that happen there and then later if you want to reference that event like if something happened at that event that you want them to have a memory of then you have to script that so just the amount of stuff that has to Happen is is limited by how much a human being can sit there and keep scripting this whole thing happened with one prompt so maybe it was like a sentence or two and just your normal English language and then all that just happened all these people are going about their Everyday you know going to work going to school doing whatever they have to do interacting with people Etc this one as she can you know talks to every single one uh some of them not even all of them and then the other thing they did is they actually also seeded something else That was happening and this wasn’t like the point of the game but they also mentioned that one of these guys is running for me or something like that and so they wanted to see how well that sort of information gets around and the point point is that information they Called the information diffusion so that information did get around so like a rumor or something like that kind of diffuses around this uh town because one person will tell this person that person will tell this person so think about how much easier it just got to create these like fully functional autonomous worlds Where you can create stuff just by you know Whispering something in one of the agent’s ears and then they they and and then they they go on your behalf to make that happen so again we went from you know hours and days of somebody sitting there scripting to one or two sentences Told to one person and then that whole thing happens like that whole event that spans days and everything leading up to it just happens learn by watching humans a couple more things really fast so first of all open AI already long before this they had some interactions with Minecraft and video pre-training so take a look at this this was June 23 2020 gym we train a neural network to play Minecraft by video pre-training VP t on a massive unlabeled video data set of human Minecraft plate by the way if you’re not familiar with Minecraft it’s Huge there’s like a billion people playing it online all the time there’s tons of video data on it there’s tons of Text data on it there’s just like a lot of AI train just requires a lot of data so stuff like this is like a gold mine Because of just the vast amounts of data that we have on Minecraft and so the really cool thing here is that they use mice and keyboard they use a mouse and keyboard to control the game the neural net Pilots the Minecraft character with a mouse and keyboard so it watches 70 000 hours of unlabeled Minecraft video and it looks like it has about 2 000 hours where contractors were people that are paid to do this they sit there and they actually play it and you can see which buttons they’re pushing right so this is what I’m pushing to mine the Tree I’m pushing you know the left Mouse button I’m jumping I’m using the spacebar whatever but the point here is they’re using they’re teaching these neural Nets to use a mouse and keyboard and and to do some fairly complex things in the game if you think about it and Minecraft Voyager also has something similar that they’re doing where they’re using these as you can see here so they’re saying that the reason they’re training Minecraft agents is because they get access to Internet scale knowledge base so Minecraft Dojo features a massive database that collects automatically from the internet Minecraft has more Than 100 million active players right there’s like almost a million YouTube videos 2.2 billion words in English transcripts 7 000 Wiki Pages Etc Reddit posts Etc by the way really fast I always want to give credit to the people behind this so Dr Jim fan so he’s uh Lucy gym fan he’s on Twitter and he’s a great follow the other person on this team that I’m familiar with is and this is anima so she’s the senior director of AI research at Nvidia here she is with the CEO of Nvidia I’m not yet familiar With everybody else on this page but I look forward to seeing more coming out of this team and uh yeah I think just really fast it is important to appreciate the researchers because they’re like moving this field forward and uh absolutely Blown Away by all the stuff that they’re doing thank you also Google deepmind is not sitting this one out either the rt2 we covered it on this channel but they created a robot basically this was a pretty I think big breakthrough in robotics because because basically what they did is they took an llm model or an LM model with vision and They said you pilot this robot and that seemed to work great I covered in detail on this channel but I feel like this is going to be how most robots are made in the future instead of being like hard-coded to do certain things we’re Going to get some sort of an AI like an llm model or something like that like they have high capacity Vision language models of blms and they put it together with an action model so now it’s a vision language action model vla and that’s where the big breakthroughs come In because it seems to be really good at sort of like generalizing the task and understanding how to do certain stuff that the llm models can do but now it’s able to like interact with the world and so here’s the paper out of Google deepmind about this rt2 and one thing They kind of bury this at the end and and here it says an exciting direction for future work is to study how new skills could be acquired through new data collection paradigms such as videos of humans and I can’t find that particular passage right now for some Reason but it was somewhere either here or in the paper where it talked about that it was indeed able to learn from watching uh humans pick up objects right there’s like I think they fed like 10 000 hours of humans uh picking stuff up and it was like it learned how to do That I was like oh that that’s how you pick up an apple okay cool data mining here’s my point when you start looking through all this stuff you kind of kind of see where this is going a little bit right so we know that robots can learn Well in simulation so you dump them into a simulation like a video game for robots right and they’re able to acquire skills much much faster and cheaper than they would in the real world right because they don’t break down they have to get serviced and it works incredibly Well right by adding little random factors in it they become even more robust in the real world because they can like they’ve already learned how to like not fall over because there’s a gust of wind or whatever we’re seeing that gpts can code autonomously they can You know you give them a high level task they come here and start coding stuff to sort of get closer and closer to that task and then the Stanford generative agents experiment what it seems to show that it’s easy to create these autonomous villagers and people in Center you give it one prompt and then they go and they execute so which means you can create whole story lines with just a very basic description right now all these big tech companies what are they racing towards the well that they’re all going after is data and it Seems to be that something like Minecraft is just a giant wealth of data not just static stuff that already exists but also all the players that are in there right now playing and interacting and reacting Etc now the only problem is Microsoft well it kind of owns Minecraft right so think of Minecraft as like OPEC they’re like all the oil producing countries of the world United and Microsoft owns it so what’s the solution well maybe build another OPEC right build another Minecraft that’s where biomes comes in this is biomes MMR RPG massive multiplayer online RPG I don’t know if it’s actually An RPG I didn’t verify that I mean it seems like it’s an MMO because there’s tons of different people running around into it I’m sure it’s an RPG I just I don’t know too much about this game but it’s made to be played in the browser Think about that for a second so having an open source version that is like more accessible to everybody that’s hopefully even better in some ways that could be huge the interesting thing is like I don’t think they would have to even have a profitable business model right they Wouldn’t have to sell a bunch of crap that nobody needs like they wouldn’t have to degrade the game to make money if this thing is generating data and this thing is just giving him all that stuff they need they can focus on maybe potentially just creating the best game Possible and not worry as much about squeezing every last dollar out of these little kids that are trying to play it anyways my point here is that everybody’s trying to build their own sort of well of data somewhere where they can train AI agents where they can Interact with real people to see if they can improve their chat Bots and create various story lines I mean I don’t have a exact idea of what they’re going to do with it but it’s obviously a huge well of data that can be used for a variety of different things from robotics to Training LMS to help it improve of coding I mean the idea of simulation and stuff like that is huge gaming is a massive Market in and of itself if you have an MMORPG that is like incredible because it’s powered by AI NPCs I mean if you look at something like World of Warcraft how much money was pulling in back in the day I think World of Warcraft was pulling in I mean it was it was hundreds of millions of dollars monthly so again I’m probably not even thinking through all the applications and I don’t even know exactly what They’re going to be using it for but it’s obvious that this is where everyone’s headache but people behind a16z are some of the like the internet royalty like the people that were there from the very beginning and made tons of money not just building stuff for the Internet but also like investing and all the other people that came before it the the quote of you know software is eating the world if you’ve heard that back in 05 or whatever that’s that came from those people you know if you’ve heard of Andreessen Horowitz that’s like that Whole team and now they’re investing in Ai and data and stuff like this and that’s where openai is heading that’s where Microsoft is heading that’s where Nvidia is heading that’s why like every single one of them this is what they’re going for which really makes you rethink That whole thing that they tell you when you’re growing up that you’re playing too much video games or whatever like this is what we’ve been training for but anyways I just wanted to put this out there I know it’s a little bit half big because like not all the ideas are fully Formed and I think we’re going to be seeing a lot more news on this front but I just wanted to put this out there because it definitely seems like all this stuff is kind of it’s all directionally aligned everybody’s going the same direction and it’s obvious that This is going to be huge I think that overlapping the stuff that we see in the AI plus video game Market I think that’s going to be the Forefront of AI in a lot of ways I think which is just extremely exciting me personally as somebody that Spent a lot of my younger days trying to build little Bots for World of Warcraft so they can run around and do my bidding for me like make a gold and line stuff and grind on our coins or whatever this whole thing just brings a smile to my Face if this is something that excites you as well make sure you’re subscribed to my channel subscribing gives you an immediate and permanent plus 20 buff to intelligence there’s a lot more to come stay tuned be ready we’ve seen a few big technological changes in our lives and Uh this one’s gonna be the biggest one by far my name is Wes Roth and uh thank you for watching Video Information
This video, titled ‘OpenAI acquires Biomes 👀 an open-source MMORPG. ChatGPT plus Minecraft? 🔥’, was uploaded by Wes Roth on 2023-08-19 20:04:28. It has garnered 10703 views and 453 likes. The duration of the video is 00:30:37 or 1837 seconds.
#openai #minecraft #gpt4
🔥 The Hardest A.I. Newsletter Ever. https://natural20.com/
[RELATED LINKS] https://www.convex.dev/ai-town https://www.thomasdimson.com/ https://openai.com/blog/openai-acquires-global-illumination https://reverie.herokuapp.com/arXiv_Demo/#
[RELATED VIDEOS]
Minecraft AI – SELF-IMPROVING 🤯 autonomous agent: https://www.youtube.com/watch?v=7yI4yfYftfM
25 ChatGPTs play a videogame… https://youtu.be/GwsRu9yLXnw
GPT-4 leaked! 🔥 All details exposed 🔥 It is over… https://youtu.be/GwsRu9yLXnw
LLMs as Tool Makers [LATM] – GPT-4 *UPGRADES* lower AI Models. https://www.youtube.com/watch?v=qWI1AJ2nSDY
Sam Altman on UBI and Massive Job Losses from AI Automation https://www.youtube.com/watch?v=5Nsqv3FWXio
[TIMELINE] [00:00] Intro [02:30] Training Robots in a Simulation [06:18] Autonomous Coding Agent [16:47] Creating Worlds with a Sentence [21:34] Learn by Watching Humans [25:26] Data Mining
