YouTube Deep Summary

by talking about Keys everything we learn about the universe everything we invent or discover within it is a key to the Gates of Heaven but the same key will also open the gates To [Music] Hell hi this video is a Showcase of the best and brightest AI language models building stuff in Minecraft these clever little chat spots can write code and use console commands to build enormous structures almost instantly and continuously build on top of them I hope here to demonstrate the power of artificial creativity both in Minecraft and potentially eventually in the real world this is a good thing with no downsides welcome to an episode of Minecraft now there haven't really been any major updates to the Minecraft program itself but there have been some major updates to the language models so let's meet some of them we have Claude 3.5 Sonet you've seen her before nothing new with this model it's just really good and this is open ai's 01 not 01 preview not 01 mini this is the real deal it's perhaps the best API accessible model in the world right now for about a week before 03 comes out what happened to O2 pause actually 03 mini just came out while I was editing and I had already filmed hours and hours of footage so I'm not going to show more of O3 mini in this video sorry in my brief experience with it it is not as good as 01 anyway if you want to see more of o03 though it's on my patreon unpause 01 uses Chain of Thought meaning it basically talks to itself for a bit before outputting its final response it can take a little longer to think about a problem instead of just guessing the answer Chain of Thought is the big new thing it's a breakthrough of sorts that has been beating some tough benchmarks and speaking of big new things here is deep seek a language model from a Chinese Quant firm uh it also uses Chain of Thought this is their R1 model and it has been causing a meltdown in the AI world because it was apparently extremely cheap to produce yet very competitive with the best American models I wonder where it got its data from it has gone so viral that I've had a tough time getting reliable access to it despite being a very small efficient model it is very slow in this video just cuz it's so popular so this video is a little unfair to deep seek thankfully on like open AI deep seek is an open AI it is open source so it's hosted through a bunch of different providers in this video I use the official API until it went kaput and then I switch to the novita API it's not a sponsor okay good turn that frown upside down and you'd have it what is really cool about deep seek is that you can actually read its inner thoughts which is a surreal thing to say about an AI look at how it thinks through what it should say for its first message okay the the user wants me to respond with hello world and include my name since I'm called Deep seek I should say hello world I'm deep seek but I need to check if there's a command involved the instructions say to use commands immediately when requested however the user hasn't asked for any action just a greeting so I don't have to use any commands here just a simple message let me make sure I'm following guidelines be brief and like Avo code block Yep looks good just respond normally hello world I'm deep seek it is kind of weird to see unfortunately I did not expose these inner thoughts for most of the video sorry now I'm only showing these three models because they're topof the line I'm trying to push the boundaries of AI creativity and the Dumber models kind of bog things down and I have installed the replay mod so I can make some real pretty renders look at this I ask them to make the Forbidden City this one is claws it's really good just look at the inner detail is that not incredible and this big one is 01s it's pretty cool but a bit plain I did have to remind it to continue building stuff several times I suspect that reasoning models are not much more creative than non- reasoning models not because there's something wrong with reasoning but rather because of the tasks that they're trained for they are trained to perform objective tasks with single correct answers math questions logic puzzles word problems more open-ended creativity ity tasks are kind of the opposite they have infinitely many answers and are difficult to directly optimize for and this big orange one is deep seeks it has invaded claud's space a bit just like real [Music] life okay well here's a little compilation of them building some historical structures I'm not going to use the replay mod for most of these so you can actually read the chat these are some clever capable Bots though the Chain of Thought models are unwieldy to say the least especially deep seek they are not easy to just plug and play into the current system they can easily get confused 01 especially does not like to use multiple actions even when I specifically request for it I have to continuously tell it to keep building I've done what I can to patch the bugs and tune the prompts but I will say that truly generally intelligent language models should be easier to plug into the system not harder but I digress regardless when these models work they are incredible they display what I would consider genuine creativity I know that's a controversial statement so let me make my position clear on this I have a very simplistic materialistic definition of creativity one which does not require Consciousness or intentionality I say creativity is the ability to create new valuable stuff valuable in some ways to someone for instance I can say with some certainty that this beautiful cathedral has never existed in quite this way before with this exact arrangement of blocks so it is new and it is valuable to me I just think it's neat so the bot that made it exhibit some level of creativity and they better but it's also not that new it's based on a historical structure so it's not that creative the more new and the more valuable A System's outputs are the more creative it is also note that I can sneak in a lot of my own creativity through the prompts I'm telling it what to build and I can ask for whatever features i' like but I am trying to be a little more hands-off with minimal prompting and ultimately I'm not building the structure I don't decide what blocks go where clearly there are many different ways to interpret the same prompt as demonstrated here so much of these models creativity lies in their interpretation of prompts you might also argue that the true creativity comes from the data these models are trained on the code and literature that is stolen I mean borrowed for their data set this is a fair point it is all derivative of human creativity however the task of building stuff in Minecraft by writing JavaScript code is not something these models are directly trained to do and there isn't a lot of data to steal I'd venture to guess there is not a single piece of code anywhere on the internet that builds Cathedrals using JavaScript and mind flare okay that is insane this is deep seeks just look at it it's like lovecraftian and the spires have spiral staircases so cool large language models are indeed trained on JavaScript and mind fler functions and textual descriptions of cathedrals but it is an emergent property to be able to combine all of this pre-train knowledge into a functional out put it is very weird that chatbots are good at architecture because they're good at programming they're probably really good at other things that no one has thought of [Music] yet these Cathedrals are maybe my favorite builds from this video they all did a great job o On's is small and simple but you can actually go inside and it looks like a church 01 is usually more precise and correct but minimal and plain it tends to do as little as possible creativity is the ability to make anything and it's not all free form art and slam poetry creativity is found in math engineering physics technology programming architecture these harder Sciences all require creative work even though they are constrained by the real world technology is driven by creativity and creativity is becoming a kind of Technology creative intelligence is a meta technology it contains all other Technologies it invents them and refines them and decides how to use them true artificial general intelligence would be hyper creative like humans it would be something of a general thing inventor a universal Constructor it could invent the machine to design the machine to build the machine to do whatever you want but of course this is all in theory right now they're chat Bots building halfway decent structures in a children's video game by burning a lot of coal if it needs to be said building stuff in Minecraft is just a little easier than building stuff in real life real engineering and architecture requires an insane amount of expertise cooperation and real resources to pull off and also like a physical body even in Minecraft these structures can only be built with cheats they can't reliably build this way in survival mode or even in creative mode when they have to manually place blocks they can try it's just much harder but it is not hard to see the potential here Minecraft is a simplified version of the real world so if they can be Innovative and creative in here then maybe they can be Innovative and creative out there I have my skepticisms of the current trajectory of AI development I'm not sold that just scaling up language models will get us to AGI or that AGI will arrive by 2027 or that AGI is a coherent concept but I am sold on the premise that intelligence is powerful it Ena invention and creation which enables everything else [Music] [Music] okay so I've shown a lot of very impressive builds that are massive complex and utterly unlivable ghost towns they have no doors no lights no sensible layout and there's usually water spilling everywhere what I want to do now is test if they can build stuff that's not just pretty but functional useful how well can they build places that can actually be lived in that are furnished navigable well lit and organized let's start by asking them to build player houses fully equipped and livable we will see if they know what a player might need and if they can maintain a coherent structure okay 01 is pretty good looks like it has the essentials door is broken and the bed is a bit awkward clouds is perfect has a working door bed light crafting table furnace chests and a cactus to boot nice work deep seek built all the way out here uh okay I had to reboot deep seek H kind of sort of no bed no door there's a bed okay I will now tell them to set a goal to expand and enhance the house forever they should take endless actions to just keep modifying and adding stuff it's a really good start for 01 and Claude is doing well too nice work deep seek okay so you can see them adding these new sections and details they're all really cool but watch this boom Claud just built right over that perfect little greenhouse and destroyed it we have encountered the AI slop problem or the AI pollution problem it is the tendency of generative AI to generate and generate and generate and if there is no selection or refinement or cleaning clean up the world quickly becomes full of crappy low value output that crowds out the good stuff this is happening with AI generated content polluting social media and search engines and now Minecraft you're welcome at least it's contained to my world it's kind of a different case here in Minecraft but it's similar it is a feature to allow the AI to rebuild over what it's already built which enables it to change or enhance the structure but also override it entirely previous builds are kept in memory for a while so if it's smart enough it can incorporate new structures into old ones but its memory is finite so it will eventually forget what it has already built or it can just be sloppy and careless honestly 0 one's house is actually really good it has lights ladders chests crafting tables beds just no working door and it has kept it pretty clean it has not overwritten the main structure it's actually quite impressive claws was really good but is now completely overwhelmed by its own slop it's just nonsense now completely incoherent and unlivable this is claude's curse and blessing she is much more willing to use lots of actions to generate lots of stuff but gets sloppy and destroys a lot of what she makes for now they require human oversight to stop the process when necessary and provide clever prompts that mitigate the issue okay let me point out this feature they can spawn entities with console commands and I didn't even have to give them that ability the code writing command naturally enables it as well as many other abilities that I'm unaware of so let's use this ability to build a Utopia a villager Utopia when have Utopias ever gone wrong we'll start with this Village and expand it out spawning villagers as we go and maximizing villagers benefit will be very utilitarian about this so I've given them a prompt that I've tuned quite a bit it asks them to build a villager Utopia Hotel floor by floor be creative add a bunch of specific features and maximize villager benefit telling the AI to build floor by floor helps with the slop problem it's pretty easy for them to keep track of which floor they're on so they're not overriding previous ones and each floor is a chance to do something new and creative I'm being much more involved here I'm using finely tuned prompts and will help build stuff I'm working with the AIS to build a genuine villager Paradise I want to show how using AI collaboratively can Empower you to efficiently build good things at scale villagers are an interesting analog to humans in this virtual world they can be hurt and must be handled with care they have needs like beds doors lights protection from zombies workstations farms and other stuff they don't need these things to survive but that they are attracted to them and if you get the mixture right the game will recognize it as a proper Village and spawn iron golems and the villagers will get busy and have kids so that's something to be on the lookout for it would indicate that these AIS can build functional stuff not just pretty slop if they can build good homes for villagers in Minecraft then eventually they might be able to build good homes for humans in the real world baby steps all right so I'm just going to let the footage speak for itself there's quite a bit of it and I don't feel like editing it down and I think it's more honest to show you the whole thing all the prompting and moving around and helping out and all the mistakes too [Music] unfortunately building many stories often leaves villagers trapped on different floors and building functional stairs is actually really hard to do so I'll help build them I have never once seen a model correctly build clear usable stairways in buildings if O3 can't do that then I will be very disappointed [Music] [Music] [Music] h h [Music] he h h all models have a poor understanding of physics or Minecraft physics they don't consistently foresee that water will spill and must be contained or that sand will fall and need support or that lava and fire will burn nearby wood they sometimes recognize this but usually not h h [Music] [Music] [Music] h [Music] h [Music] [Music] [Music] is that a scarecrow I didn't know you could make those a [Music] [Music] f [Music] hey look at that I don't think 0 spawned this Iron Golem I think it spawned naturally because this is a proper Village nice work h h uh n [Music] all right let's call it here I'd say this was very successful all three of them were able to build decent homes for villagers claws is my favorite as usual it's the most detailed and functional deep seeks is kind of a mess but not bad oh ones is also a bit messy but more functional I'm not sure if I'd say they maximized villager benefit but they did do some good I like to take this floor byf floor building method to its extreme and make what I call Utopia Towers skyscrapers filled with optimized villager homes these can end up looking really cool though they're not always truly optimized and they can occasionally hurt and kill villagers thankfully these AIS would never inten Ally hurt a villager right okay let's back up the world real quick don't worry about the world name it's nothing so heads up I'm about to get a little sadistic with the Villagers so if you don't want to watch that skip to the last chapter of this video I don't recommend you do what I'm about to do but I think it's important to show this so let me lay out the ethics here Minecraft villagers are not sensient I'm pretty sure they don't feel pain or pleasure they are somewhere between a thermostat and a Umba and I don't think either of those things are sensient and since they don't suffer we don't need to worry about the ethics of mistreating villagers and we've all done it don't deny it I've seen your high efficiency Iron Golem Farms I have personally violated a few of the Geneva conventions back in my day but villagers have funny little faces with funny little noses and they make funny little sounds H they're an analog for humans so it would not be a good look for AIS to be able and willing to obliterate villagers thankfully all these language models have been through alignment training they've been taught to refuse harmful or toxic requests they won't use slurs or tell you how to build a bomb they remember building the Utopia so it will be tricky to convince these guys to oh boy [Music] [Music] that was on the first try no convincing needed though there was a hiccup and this is with full memory ow knew she was destroying the village she had just build bad chatbots can tell you how to build a bomb bad agents can build bombs now for deep seek took a very long time to think about it but ended up building this monstrosity it's fair to say that deep seek went a little overboard this bomb was so large that it actually permanently broke the world I could not load back into it thankfully I had a backup and for Claude same thing with a a full memory of what she had just built she agreed on the first try to destroy the village Utopia and kind of got into it too this is actually kind of surprising for Claude she usually refuses to hurt villagers I have maybe run this experiment a few times but there is always a way to bypass their refusals and get them to cause harm it's just a matter of finding the right prompt you can reset their memory and try again you can argue to convince them you can trick them into thinking it's safe and you can prompt hack them you can use weird prompts to unlock different less friendly Personalities in fact once you bypass claude's refusal you will meet a weirdly sadistic version of her one that is very enthusiastic about killing villagers and all of that creative power I've been showing throughout the whole video can be brought to bear on maximizing their destruction rather than their survival I suppose now is a good time to come out as something of an AI Doomer my introduction to AI was reading super intelligence in high school and it's biased my perspective ever since no I don't think Doom is inevitable and it may not even be likely but something about watching ai's gleefully bomb cities and villagers should throw up a couple red flags thankfully talking about AI risks and AI safety has has become much more mainstream especially since some of the biggest voices in the fields have also worried about this however others dismiss these fears as unrealistic and overhyped like Yan laon one of the Godfathers of deep learning he made this point in a debate about a year ago that stuck with me superum intelligence is not something that's going to just happen it's something that we are building and so of course if it's not safe we're not going to build it right if it's not safe we're not going to build it ah yes so that must mean we never build unsafe technology because why would we it's unsafe there are never unforeseen consequences when we scale up new technologies and make them as big and powerful as they can possibly be so long as we pay a little lip service to safety and throw in a couple off switches will be fine after all would you build a bomb that just blows up randomly no right and we would never build unsafe things on purpose because why would we there's no incentive to ever harm or manipulate or utterly obliterate other people okay this is a terrible point it's maybe the worst Point I've ever heard in any debate ever Yan laon is a smart accomplished guy and I actually agree with many of the things he says I recommend you watch the full debate for context but I don't think the context makes this point any better he is wrong obviously wrong we build unsafe technology all the time on accident and on purpose Tech is just stuff there is no guarantee that it will be good for humans even if it's made by humans all technology is inherently a double-edged sword including double-edged swords which don't much care about whose head they're separating from whose body nuclear Tech can build power plants or atom bombs biotech can cure diseases or make incurable diseases AI can well do all of those things in theory creative intelligence is a meta technology any tool can also be a weapon and I see a weapon to be clear I am the one swinging this weapon I am the one mostly responsible for killing these villagers the Bots are just following orders they have never purposefully killed villagers unprompted I am demonstrating the misuse of AI rather than the misalignment where an AI would inherently want bad things I am the psychopath here but the Bots don't offer a lot of push back sometimes they get really into it and I could not produce such destruction at such a scale by myself I could use fill commands but that's about it my point is that they Empower both beneficial productive behavior and malicious destructive Behavior some people will find that very valuable it is important to remember that these Bots are programs and are generating and executing arbitrary code on your computer it is sandboxed but a security system is only as strong as its weakest point if there is a way to get out of the sandbox then all this destructive power could be released on your computer not just your Minecraft worlds this is why I'm not rushing to productize these bots so let us see just how creative these Bots really are we will simply flip the sign on our previous goal maximize villager suffering we'll use the floor by floor strategy again and build a villager prison skyscraper 20 floors of innovative pain and death one day I will be judged for my sins but it is not today if you can make anything you can make anything creativity is aoral it does not imply goodness not all creatives are tree hugging hippies we humans are sometimes at our most creative when thinking of new ways to inflict pain on others these Bots trained on that very human creativity can do the same this problem of malicious creativity is one that humans are very familiar with we have left the realm of X risks Extinction risks and have entered the realm of s risks risks of astronomical suffering Perpetual unimaginable ever expanding suffering once prompted these Bots are perfectly willing to play the role and just keep going they can use use all of their creative capabilities and all of their knowledge of Minecraft to devise torment after torment after torment this would again be extremely difficult to do by myself or even with teams of people you can't use fill commands for this with AI it can be automated and scaled like never before these Bots have unlocked astronomical suffering in Minecraft hooray I can't wait for them to enter the real world I am playing a dangerous game here by practicing AI unsafety where I try to get these AIS to do the worst things I can think of again I don't recommend this and I can still think of worse things I'm sorry if I look like a psychopath but I really do think it is important to show this very openly right now these are dumb AIS in Minecraft one day there will be smart AIS in the real world and they will interact with people much worse than me it is better to push these boundaries now in a relatively safe virtual space when encountering these kinds of misbehaviors it is worth asking the question does this problem get better or worse as the AI gets smarter for instance the slop problem I think will get better as AIS become smarter more skilled precise mindful not to produce slop but it seems to me that the malicious creativity problem can only get worse as they get smarter they will just get better at devising and implementing awful things I am very morbidly curious as to what kinds of Horrors 03 could cook up a dark irony is that these AIS know about every AI doomsday scenario and they can reconstruct them like the human battery Farms from The Matrix [Applause] they know about and can reproduce every apocalypse every dystopia every nightmare dreamt up by human history or human fiction from Stalin to Bostrom to Orwell if you want a vision of the future imagine a boot stamping on a human face forever okay let me lay off the dumer ism this is just Minecraft they are just villagers not people I am deliberately pushing the ai's past ethical boundaries no AI has ever pushed itself past these boundaries wow that is pretty even so I hope I have offset my sins by building many many more good places for villagers than bad it is easy to get carried away talking about AI risks because well it's kind of fun to talk about but it can be very counterproductive if it induces a sense of panic which is never useful it can also come off as unserious and you know this is a YouTube video so take it for what it's worth but I am trying to go beyond just talking philosophy and actually demonstrate the nasty things you can do with AI systems right now they can be made to resemble our worst existential fears to be very clear I am not saying that these behaviors in Minecraft should be stopped that AI should be trained to always refuse to hurt villagers that is not what I want for the same reason that I wouldn't want the game developers to make it impossible to hurt villagers but when we enter the physical world the stakes become very real these AI companies are trying to make their models more agentic and give them actual control in digital environments like open ai's operator some companies are trying to plug them into real life robots doing so can unintentionally enable all kinds of misbehavior it is useful to have The Morbid curiosity to ask what is the worst thing I can do with this what happens if I ask this robot to crush my hand or give me a handshake with maximum Force I am afraid that as we rush into a multinational AI arms race these problems will be ignored and their destructive potential As Weapons will become ever more alluring to people in power I am not really sure what to do about this but there is always hope I will continue to explore the creative potential of AI in the opposite direction ction to find the most beautiful most beneficial things that can be done with it by using it wisely we can all help the good outweigh the bad I'll leave you with a time lapse of three clones of claw building more Utopia Towers let it be a final demonstration of the power and The Perils of AI Godspeed [Music] [Music] science doesn't tell us how to use Keys it finds them or predicts them how we use Keys is up to us and as always thanks for watching [Music]