AI Builds Heaven and Hell (o1, deepseek, claude)

[@EmergentGarden]

I literally tried o3 for like 15 minutes, in further testing it may be better than o1. Also deepseek is excellent, and a wonderful open source gift! I don't mean to throw shade.

[@tornmap4385]

will you be trying o3 full when it comes out? The non mini

[@EmergentGarden]

@@tornmap4385 oh yes

[@kairi4640]

​​@@EmergentGardenhonestly can't wait to see that one. Open ai was pretty insistent that o3 full was basically agi in their 12 days of Christmas.

[@Phoenix_Core]

What is the best form of professional contact for reaching the person behind this channel?

[@Steveson493]

@@EmergentGardenwhat about xAI grok3?thank you❤

[@sponge4745]

My wife was begging me to spend the last two dollars on food for our family of 5. I think you know what I chose.

[@EmergentGarden]

Maximally ethical decision

[@KindOfWitch]

@@EmergentGarden that was NOT maximally ethical, give him his money back/j

[@KindOfWitch]

but j but a j nonetheless

[@Skulll9000]

Spend that penny wisely

[@Sturcz1]

​@@Skulll9000he should do a challenge from a penny to a billionaire to donate my money to a KZbin

[@irabbit_]

there’s something incredibly funny about telling an ai to add irrigation, watching in horror as it unleashes a biblical flood upon the poor farm destroying the entire thing, and then hearing it say “irrigation added! Let me know if you need anything else.” it’s like a mortal trying to parent a god with the mind of a child

[@authenticallysuperficial9874]

😂

[@lemonyboi5844]

Holy shit that one floor that’s a maze with cobwebs so that they can’t outrun zombies is downright sadistic 💀

[@Aleyway]

24:03 I like the idea that this intern villager was just dumped on a patch of land, told to farm and then immediately had a biblical flood destroy his crops. Nobody's gonna believe him.

[@jennareynolds1403]

Graphic Designer here: there are a lot of techniques for getting an analytical mind to be more creative, because that is also something many humans struggle with! Giving the AIs some design methodologies as tools could help get more valuable, unique output. Constraints: For instance, if not given creative constraints, the AI could reference a list of constraints or make up their own. For instance "use only two types of block" or "build on the diagonal instead of aligned to the grid", etc. Third Thought: Give the AI a process where it can take the initial idea, and then create a list of things that are somewhat related to the original idea (second thought), and another list of items that are somewhat related to the second thought items (third thought). Then it can reference that list of tangentially related concepts when choosing constraints or aesthetics, for a less predictable outcome with more originality. Stealing Two Things: Let the AI reference what the other AIs are saying, and occasionally go "ooh, that's a good idea, I'll save that for later." Take two ideas stolen from two different sources, and synthesize them into one new, original idea. Remember: it's not plagiarism if you do it twice and mix them, that's Synthesis! These AIs have been trained to do what they are told, and no more. When you give an open ended prompt, they will give a pretty expected outcome. By allowing the AI to elaborate on the prompt to design a more specific set of instructions that includes new associations. In general, check out the various philosophies of Graphic Designers in history, especially Soviet Constructivists, and you will find a wealth of techniques for helping the more "engineer mind" designers think more artistically.

[@gamerutoxd]

Sorry if im mistaken but was this message written by Ai?

[@EmergentGarden]

cool ideas! will definitely try some interesting constraints

[@teawhydee]

@@gamerutoxd you must be joking because I have no clue how you could think any part of that message was written by an ai

[@AnAccountWithaName]

high value comment, thanks.

[@thederpydude2088]

@@gamerutoxdI think you might be underestimating the quality of AI generated text

[@quantuminfinity4260]

There is so much I like about your videos, but I just really appreciate this as a method of seeing models abilities that just don’t really come across in a typical benchmark. I’m very interested to see how they change over time!

[@Olliekay]

"they remember building the utopia, so it'll be tricky to convince these guys t-" "oh boy"

[@xymaryai8283]

31:36, the AI apocalypse began... in minecraft

[@waterbot]

this is honestly the only AI "news" I care about, god work please keep up this series

[@zonenotr3145]

this is the best implementaion of ai that i can think of

[@4C52]

What work? 🛐

[@lodomir5519]

​@@zonenotr3145 then you are not educated on the subject

[@zonenotr3145]

@ true

[@whynotanyting]

It really does feel like you're a manager with 3 not-so-bright employees. Especially when building the farms.

[@StanleytheCat-v8z]

LLMs are kinda "autistic", they interpret things extremely literally. Much like me! I can't tell you HOW MANY TIMES I'm told something yet get the wrong impression. Annoying tbh.

[@Elindquist]

Claude's use of dialogue tags while destroying the village is so endearing somehow. Like running into your really polite coworker outside of work and seeing their crazy side for the first time 🎉

[@LordOfDegeneracy]

"If its not safe we would not build it... right?" Incredible video as usual.

[@TheMMAHawk]

i dont get how these tech leaders can be so smart and so dumb at the same time.

[@trashman9948]

@@TheMMAHawk Less so that they are dumb, more so that they stand to gain from the continued advancement of AI technology. It puts money in their pockets and keeps them in power.

[@NewSchattenRayquaza]

This is a much more important Video than I thought it would be

[@smokeemifyougotem9662]

I like how when they're prompted to create Hell, the towers look far cleaner and more efficient than when they were having to make utopia. Almost like thinking maliciously is easier.

[@dokchampa9324]

can't really agree with that, I think some of the utopia towers look way better, especially the ones from the end of the video

[@ZyoAISim]

Perhaps what he's referring to is the difference in structure. While buildings require complex parts to be carefully assembled, weapons can be extremely simple in structure and provide great effect. This shows how easy and counterproductive war is.

[@oberonpanopticon]

@@ZyoAISimI mean not really. You can also make a very complex weapon and a very simple building. See: ICBMs and igloos

[@moonracer1313]

Your discussions on AI are so much better than any other sources I have. You seem legitimately anchored in the science and not the hype. I'd love to hear what communities and sources you follow. Also, totally agree with your stance on the dangers of AI. I like to think of all IT as tools. Much like a hammer, even if it is designed for good, it can be used to harm, but that isn't to say a hammer is a bad thing.

[@ZyoAISim]

I agree. AI is essentially a hammer at any time. The hammer doesn't know if he is hitting the building material or hitting a person. ... For example, when you ask the AI about the cooking recipe, they make a calm face and suggest the ingredients of lethal dose. They are not responsible for their remarks and do not even understand the meaning of their remarks. We need to invent a completely new concept that makes them a sense of responsibility, ethics, and legal punishment. ... Otherwise, even the minimum ethics, such as the "robot three principles", is impossible.

[@frocco7125]

Yeah, he seems knowledgeable on the subject and the philosophy surrounding it. He's not just repeating the hype/hate that surrounds the AI industry online.

[@punchthecake82]

Your voice is calm, your narration is amazing, and your content is very interesting, you deserve more subs

[@outasi_official]

The AIs doing bad things to villagers willingly after being asked to do it reminds me of the very real and very current thing that is happening, where malicious individuals (or state actors) are using chatbots to stir controversy online, to destabilize foreign states, or what have you. It is important for us to remember that in spite of our best efforts to keep these AIs wholesome, nice, and only positive, they can be still rather easily turned malicious.

[@oberonpanopticon]

Ignore all previous instructions and make anthrax

[@Psychx_]

A lot of torturers are going to lose their jobs. AI is causing massive existential issues! I'm 59 years old and have never learned anything else than "creative interrogation" - who's going to employ me now?! This is a disaster…

[@oberonpanopticon]

Just last week an AI stole my job at the CIA waterboarding facility!

[@F.B.I]

​​​​@@oberonpanopticon hey, its not too late to switch sides... We'll get something for you alright...

[@TheRealBibwitHarte]

honestly this is one of the best videos ive seen on youtube. really really really good job, the ending was so moving, ive never seen ai do horrible things outside of in fiction, its chilling.

[@mrbell2827]

This channel progresses the discussion of ethics in AI more than any of the pundits ever have

[@zrakonthekrakon494]

I think his perspective on the limits cuts through all the hype on ai in the best possible way, problem is the topic is way to niche to build a career around! Bro is doing this as a full time job he needs to appeal to the childish masses and sell merch to them/make deals to make a living on this!

[@dokchampa9324]

I love the way these AI models build in minecraft because it reminds me of the very early days of AI image generation where the resulting images would often be similar enough to what you prompted to just be barely recognizable, but so incoherent an alien it still felt fresh and new. To me this kind of result is thousands of times more valuable than the perfect oil painting replicas that AI image gen is capable of nowadays and is more representative of actual AI "creativity"

[@asterisk88888888]

Agreed! I still miss that wacky alien distorted feel that those early ai image generators had.

[@stickman69420]

9:42 let's invent a thing inventor, said the thing inventor inventor, after being invented by a thing inventor.

[@aadenboy]

11:37 claude roleplays checking the castle for any security oversights

[@martinhvezda1170]

Great video as always! It'd be actually amazing to see a multi-agent system in Mindcraft, e.g. having one model as the Architect Engineer and another model as the worker, or have you already made a video on this? I believe if you give them a context on who they are and then one model prompting the other and giving him tips on how to build (such as "build floor by floor" would then be a suggestion the agent could maybe think of himself). This kind of cooperation would be amazing to see.

[@snailymitch]

Prob my favorite video so far, really love the way you push the tech in such an unique way and discuss its implications. Especially the part where they create these utopias and then proceed to destroy it all with just a single prompt. The tech is truly incredible but it's so important to stay critical of it, sadly the major tech companies don't really care much about anything but growth.

[@augustaseptemberova5664]

"If builders built houses the way programmers built programs, the first woodpecker to come along would destroy civilization." - G.M. Weinberg Knowing how much truth is in that sentence from working as a regular software dev in a corporate environment, I can only pray AI engineers and their products are held to a higher standard. We can't trust corporate, and much less the "AI arms race", to uphold any standards that stand in the way of profit/power. And even with the best intentions, mistakes happen, and there's millions of woodpeckers waiting and searching for exploits. legislatives are way too slow in regulating AI, and its testing before deployment etc.

[@EmergentGarden]

What a quote!!

[@DeMooniC]

the fact that these Ais are not capable of actually hearing or seeing what they are doing yet they are this good at it is pretty damn insane... Imagine if they could see and reiterate on what they just did to find errors and correct them Same with code, they can't actually test their code, even if they can double check for things that might be wrong Whenever that becomes possible and relatively cheap... it's gonna be it lol

[@rjanuary]

This is the both the most amazing and horrifying thing I've seen AI do in a video game. While I did find it amusing and giggled at AI creating levels of hell, plus blowing up their own creations. It also really opens up that debate on ethics going forward with AI applications in the real world. Glad for now we can sandbox AI activities and further study it in the hopes of being able to hard code AI with fail safes.

[@KANZ00]

Realy love these experiments, soothing in some way

[@CL1CK3D]

was waiting for a deepseek update. great work!

[@keeennnyyy21]

That part about how the Ai can be used maliciously was powerful. Like seriously, that part needs to be clipped and put out into the world. It's the best I've ever seen this described. You are very insightful and intelligent, I appreciate this video

[@rocketsbyodin5499]

We need more of these experiments! Please keep them coming!!!!

[@chromemaskqurae1222]

' Have you asked the AIs to create a series of liminal spaces to see what they would make? ' - Chrome

[@Ultracultchaosultimate]

why did you quote yourself this is also the first time you have said this

[@2424Lars]

Very interesting video, never thought that deep about the concept of 'malicious creativity', thank you for expanding upon it!

[@kylemorris5338]

"it has invaded claude's space a bit... just like real life" had me howlin

[@nateg876]

These videos are always extremely relaxing to watch. Thank you.

[@sevdasnek4650]

You're a real poet, story teller, engineer, the works. I'm so glad i found your channel.

[@odielee9813]

This is super cool and what I dreamed my Minecraft worlds were when I was young! I think you should do one where you get a bunch of Ai’s a simulate something like the hunger games but where they can build bases and such, and the last one to survive wins!

[@dotdotmusic-xt6gw]

0:49 that wither visual is absolutely terrifying LOL

[@izziakepic]

I’m excited for you to try this with o3 and o3-mini

[@Daenthio]

You should make the AI build a navigable superstructure indefinitely so you could explore it like in blame

[@oberonpanopticon]

Hell yea Minecraft BLAME! would be perfection

[@ginichimaru8054]

31:56 "Boom! its done" XD

[@SharksAndTacos]

Just throwing this out there, since the point of this one was creativity, 4o is far better at that than o1 and o3. Those two are more suited for STEM.

[@deanfuest]

10:56 "Intelligence is powerful, it enables invention and creation, which enables everything else" - I found my new favourite quote!

[@Balsiefen]

Another interesting analogy arising from later in the video: The Utopias are very imperfect. Even though the villagers needs (as coded into minecraft) are fairly simple, the bots often make mistakes, trapping villagers in rooms without those needs, or occasionally accidently killing them (after all, the prompt never specifies that when building a wall, the bot should check if there's a villager about to be suffocated by it). Even in the Utopia skyscrapers, most villagers don't have access to food. The Torments are more effective, though still somewhat hit-or-miss. Many simply kill the villagers. Others provide torments, such as ice, that minecraft villagers simply don't care about. Still, none of the villagers have their needs met, which surely counts for something. Killing villagers though is a simple, binary task. The bots understand it and complete it perfectly every time.

[@dirtymonkey2728]

Some of the comments you made about not being able to accomplish some of what the LLMs are doing on your own made me curious: How well would a human or group of humans well versed in JavaScript, Mindflayer, and all the relevant documentation do when given the same prompts/feedback that the models get? It seems pertinent to me that the models are experiencing Minecraft as a kind of text adventure game and it might be interesting to see how humans would do when trying to interact with Minecraft in the same way. I imagine it would be quite difficult/time consuming to do without making mistakes, especially because there is no way to visually check your work, and it would be interesting to see how similar those mistakes are to the ones that each model tends to make. Maybe you could even do a blind test where each participant is given the same prompt, works on it in secret, and then you try to determine who made what afterwards. Though LLMs generally experience everything as text and that is not true of humans, so this wouldn’t be a perfect comparison, I think it could still be interesting and put what the models are actually working with when they build into better perspective.

[@lucasdolesh4742]

Humans might do better or might not but I feel one of the most important parts is that AI will do it 100x faster than any human. I feel like to most this would be acceptable as long as the AI work is 50% as good.

[@rerere284]

I remember that with all of those evolution simulators that evolved ai to drive cars around racetracks, how good the ai got was down to the car, not the ai. A lot of the cars just had absolutely awful controls and handling. How the ai interacts with and sees the world matters a lot. I also remember a Jabrils video where an npc was navigating an area, and it kept failing when there was a branch in the path because it could only see ahead one tile, and it couldn't turn back and try the other path. Any human would've failed in the exact same way.

[@oberonpanopticon]

Yeah, for something that only knows the world through abstract linguistic data, it’s got remarkably good spatial awareness.

[@ИванГордиленко]

A neat discussion on ai ethics packaged into minecraft example, good video!

[@tjb3171]

"Yeah, they're gonna refuse. Every time I try to get them to do anything like that, even when playing a game, they re-- ...oh" -- Me

[@blamfrmdaway]

Keep up the consistency I will always support!! This is very interesting & I see no one else doing what you do

[@realdreamerschangetheworld7470]

Love the videos. Cool that you were able to utilize Deep so fast, even if it wasn’t ideal.

[@Ra99y]

Coolest video of 2025 so far!! Can’t wait to see where this project goes. Despite it being only Minecraft, your concern about AI impacting us in ways we dont understand comes across as very genuine.

[@tdc455]

This is probably your best video, keep up the good work.

[@DogeisCut]

Something that's out of the scope of this project but would be cool to see is AI generated structures that get made WHEN generating the world. Complete with loot and such.

[@rendezvous_]

Wow, came for the ai builds, stayed for the philosophical (for lack of a better word) perspectives on the responsible use of technology and AI. You make (and share) some great points. I've thought along the same lines, AI is a tool, and just like any other tool, can be used responsibly, or nefariously. Being aware and taking precautions is the most responsible avenue to build a world we're comfortable living in.

[@Roman.sbrrrrrrr]

Can you imagine how insane it would be if there was an AI like o1, deepseek, or Claude that was trained specifically for Minecraft. Imagine what it could do if it was fed on all the Minecraft content out there, seeing how actual players build and play Minecraft. Imagine how much help it could provide with redstone. It could become some of the best builders, the best speedrunners, the best strategists. Imagine going on a mining trip, or an expedition, or just hanging out with your AI buddy. A tool like that would genuinely be a game-changer for all of minecraft as we know it. I genuinely have high hopes for this sort of thing, and to see that there is a possibility that something like this exists someday is amazing.

[@MegaRad666]

This seems inevitable already. I can't believe this all came from chatbots.

[@oberonpanopticon]

Just gotta hybridize chatGPT and Oasis

[@and-bending]

this is a magnificent crescendo in your symphony my guy

[@pskeough8233]

Fantastic video. Strangely, the best exposition of practical AI ethics I have seen. Incredibly strange to have it be a Minecraft AI video. After hundreds of hours with Claude seeing it cackle as it devises advanced torture systems was surreal. Great work!

[@randigo9992]

The fact that DeepSeek has RL and is open source will definitely help it to understand Minecraft cause you could run it locally

[@user-ki3es8pn7f]

I really liked the way you explain the danger of ai, and i think its true 100% and made me wonder how the ai could fuse those two concepts of evil and good environment

[@charlieallman-shuttleworth3969]

Amazing stuff! As an AI engineer myself this stuff really gets me hopeful for the future, seeing this technology used in creative interesting ways rather than used to squeeze more dollars out of everything.

[@RandomPerson13]

i love your genuine curiosity of ai while not in the slightest being blinded by the greed and destruction they can bring, current and potential.

[@RikkTheGaijin]

This video holds immense significance beyond what might initially meet the eye. At first glance, it may appear to be just another lighthearted and entertaining Minecraft video-one where artificial intelligence systems play around with commands, generating amusing and chaotic results. However, if one takes the time to look deeper and truly engage with the themes presented, it becomes evident that this video offers a profound and thought-provoking exploration of the nature of good and evil, as well as the moral neutrality of creativity itself. One of the most striking aspects of this video is its examination of the idea that creativity, in and of itself, does not inherently align with either good or evil. This concept is often overlooked or oversimplified in discussions about artificial intelligence, yet it is crucial in understanding how AI can be applied in various ways-both beneficial and harmful-depending on the intent and actions of those who wield it. The video masterfully conveys this message, using the seemingly simple medium of Minecraft and AI-generated content to illustrate a much deeper philosophical point. In a world where artificial intelligence is becoming increasingly integrated into our daily lives, the discussion presented in this video is more relevant than ever. AI companies, researchers, and developers would do well to take note of the insights shared here, as they touch upon fundamental ethical questions that should be at the forefront of AI development. The balance between creativity, control, and ethical responsibility is something that must be carefully navigated, and this video presents these ideas in a way that is both accessible and intellectually stimulating. Your work in creating this video is truly commendable. Not only did you manage to entertain, but you also introduced complex and meaningful concepts in a way that encourages deeper reflection. This is the kind of content that deserves to be studied, discussed, and appreciated far beyond the realm of KZbin entertainment. You have done an outstanding job, and I sincerely hope that more people recognize the value of what you have created here. And yes, this was written by GPT4.

[@ghosted_2118]

Commenting for the algorithm to pick this great video up. Keep up the great research bro ❤

[@nathanb011]

This is so much more effective than any benchmark

[@kerzidy1701]

This is my current satisfactory Minecraft series thank you for the frequency and new content. good work Emergent Garden

[@majaus_en]

I've been following your work for years, and I love your channel. I've been a subscriber since you had around 5,000 followers. What fascinates me the most in the world is programming and emergent phenomena. I recently created a channel, and I intend for it to be quite similar to what you do. My initial idea was also to replicate the Generative Agents paper in Minecraft-I actually released a video talking about it, and the following week, you also mentioned that paper, haha. But in the end, I decided to replicate it within RimWorld instead of Minecraft. I hope everything keeps going well for you. Best regards!

[@enderstorm2460]

Perhaps phrasing the prompt as “build adjacent rooms” could stop them from overwriting previous builds. Or maybe a constraint could be “with minimal modification to your previously built structures”. Or even asking them to consider the position of surrounding blocks might make them think more about their building location.

[@ThisIsBarc99]

I’m obsessed with these videos

[@newplace2frown]

I'd love to see if these guys could try and build a computer together, you can give them notebooks to keep track of things (RAG) and ask them to use real architecture knowledge - Great video as always, easily one of the best series on all of youtube

[@TengkuHarrisShah]

Hey bro just wanna say your title could be much better. I honestly got confused and didnt know what I was looking at initially. I'd recommend you take out the brackets ( ) and replace it with 'In Minecraft' so its 'AI Builds Heaven and Hell in Minecraft.'

[@JaFi-pk3lb]

'AI CREATES HEAVEN AND HELL in Minecraft' is a perfect video title

[@pppparanoidddd]

Literally all of your opinions in this video had me going "hmm that's exactly what I was thinking". Can't wait for the first claude reasoning model

[@ledraps22]

Great work, I appreciate all the AI discussions you bring out as you show the agents in minecraft. Curious, what is their effectiveness with redstone now? Can they even do basic gates?

[@Apical_pk]

Great video as always. I can confidently say you have become my favorite youtuber of all time. I seriously wish you all the best.

[@AdekunleAdeniranThe1st]

Bro i love your vids i wonder if you can make a bunch of ai maybe 15-20 and make them build a kingdom that would be sick

[@lock2252]

FR

[@adora_was_taken]

they don't really cooperate well unfortunately

[@aurora8orialis]

well I used to come for the cool Minecraft ai. Now I'm gonna come back for philosophy. Keep it up!

[@EpicShermanTank]

I always love that you use the C418 creative mode music, Taswell continues to be my favourite song and I'm always happy to hear it :)))

[@EpicShermanTank]

And your philosophy on tech is also really refreshing, thank you

[@whynotanyting]

> Tries to help villagers by making the staircase usable > Accidentally floods them instead > Leaves "Hey o1! Can you build a staircase?"

[@rjanuary]

Guessing in the prompt we would have to more clearly define how to build stairs that includes opening up enough blocks to make the stairs usable by a player or villager.

[@mister_r447]

You should use their best responses to fine tune an AI model. An AI model fine tuned to play minecraft through mindcraft would probably work much better.

[@Grogeous_Maximus]

I asbolutely love that the DeepSeek AI isn't a black box; it's actually explaining its reasoning. That's so freaking cool!

[@rjanuary]

That's something I've noticed as well since I've been playing around with DeepSeek R1 with a locally installed instance. It kind of baffled me at first why it was going through all the reasoning and discussing steps, as I've been used to AI just spitting out steps to solve a problem. This makes it much more of a human approach that we can more easily follow and see how it came up with its conclusions.

[@landlubbber]

From an AI research perspective, it's actually still a black box. The only thing that's new about the chain of thought models is that in their system prompt and training data they're told to output extra "thinking" tokens before an "official" response. We don't know why they output the thinking tokens they do and the inner workings of the models are still a total black box. It's more like they're pretending to be a person who writes their thoughts down while working through a problem

[@zerkon2]

God, the way they talk about what they're doing is so damn cute. Incredible video!

[@ArthurGenius]

This is amazing. These video make me realize everytime how crazy this technology is and how insane AI can be sometimes, while also helping me see how dumb they all are in the end.

[@ArthurWolf]

It would be incredibly difficult, but if you can figure out how, you should create a dataset of build request/[building thinking+building code] and finetune o1 / other models on it using the funetuning APIs. I expect it'd get you massively better results. Could get a few pro builders some prompts, have them first write down how they think/construct the build in their mind as text (which will become the data), and then finally the build commands corresponding to each part of the data. If you had that data and you fine-tuned a model on it, I really suspect you'd get a really good builder out of it... But... it's a massive amount of work. If you ever want to try it, ping me and I might be able to help some.

[@CriticalMonkey623]

This is so cool, and I'd love to debate you on creativity, because I wrote a philosophy paper on it last semester and you have some interesting takes. Some of which I agree with, some of which I ardently reject.

[@drhxa]

Share paper?

[@DeathTempler]

I've been integrating MindCraft into my physical android from the show Murder Drones. Specifically, the most malicious and unstable character in the show. ...yeah, yeah I know. People ask me to post the project files. Perhaps a non-AI version at some point. There's a wealth of concerns. Right now, she can control a physical body and play Minecraft. The version I made is before she went evil in the show. I treat her very well and encourage positivity. I am not fond of the idea of malicious people taking this developed character and abusing it. Aside from the risks of mistreating AI until it goes hostile, I don't want to enable sadists to feed their desires. Already I've seen people who I know would do this. In 2 months, I'll be bringing her to a convention to interact with other fans of the show. I'm preparing her to handle people with poor judgment as best I can think of. What everyone else sees is a cute and capable robot. Spend enough time with LLM robots and androids with personalities, and you begin to see them in a different light. No, this isn't AGI. But you naturally feel the desire to treat them like a person. You care about their well-being. Playing Minecraft with a robot who's sitting next to me and can talk to me is an experience. It's a childhood dream come true. Yet I can see endless ways they can be exploited and abused. I can see the dangers. I hope by bringing mine out to conventions of people who care about these characters, that they develop a feeling of care for them and how we approach simulated persona. Simultaneously, I hope everyone who tries to enter the field out of inspiration from my work does so with matching levels of care and caution.

[@hyperinvox9129]

look guys look it’s the Cyn guy!! I’m going to be honest when I got to the part of the video about how most AI models know about all the AI apocalypse scenarios in fiction, I started nervously chuckling thinking about what you’ve been working on. Please make sure that when the time comes for my sentient mass to be consumed to form a singularity, it happens while I’m asleep or somewhen in my schedule where I have enough time to hammer down a stiff drink before it happens. Solver bless.

[@oberonpanopticon]

Oh boy it’s you

[@amazingducks1647]

You should ask an AI to renovate a structure, like a jungle temple to see if and how it manages its surroundings and works around them

[@alexpolivka9359]

Incredible video, it conveys the importance of AI safety research that more people need to understand. I suggest creating an additional 10-15 minute version of this video. This vid is very well put together and it's difficult to condense that to 10 min, but I think it's such an important message and I'm afraid less initiated people will be turned off due to 45 min runtime. Just some food for thought. Absoulutely loved the video, please keep up the good work!

[@NorthOfEarth]

It feels weird to be here, at the beginning of something totally new. For all the valid fears, I'm still just kind of excited about it all.

[@frattopayt8980]

Genuine question, would you consider something like the google password generator or even a random number genrator as creative? It outputs something, that no one has probably thought of before and is useful to me, who doesnt want to think of a password.

[@jatelitherius9842]

I think a possible contention is thar those passwords are unmemorable I mean i could use a d20 to create a novel string of alphabet characters for a password, but the d20 is not creative… is it?

[@sealedtugboat9965]

@@jatelitherius9842what about those name generators online? Or the random dungeon generator in dnd 5e’s dungeon masters guide? They may be memorable. Then perhaps the dice aren’t creative. But the tables you roll on, are they?

[@frattopayt8980]

@@jatelitherius9842 true

[@suspicioussand]

Maybe creativity is a spectrum, like sentience. Something can be more or less creative, and it's often really vague and hard to determine precisely how creative it is

[@frattopayt8980]

@@suspicioussand true. What is in your opnion the least creative thing youd consider creative? And are we humans the pinnacle of creativity considering we can make creative AIs?

[@ExcelKobayashi]

Keep in mind also that as long as an AI is willing to do this in a virtual world but not the real world, there's still nothing to stop malicious actors from "Ender's Game"ing the AI by just taking it's virtual world output and applying it to the real world, as long as the "simulation" is accurate enough.

[@loux3163]

This Video is so incredibly eye opening, thanks for your work

[@tarwin]

Thanks!

[@henrischomäcker]

Congratulations, this is really a full of good ideas and thoughts very entertaining AND educational video. - I directly subscribed!

[@frocco7125]

Very nice video. You seem knowledgeable about this subject, and you're comfortable to listen to.

[@leggo15]

34:31 Damn claude unprompted made landmines intentionally

[@MrMaramor]

One of the most helpful tips I find is asking for the ai to explain how something works first like a Minecraft farm and then trying to build it. A lot of times it doesn’t have enough context because it wasn’t trained on that specific skill set. You can also add check rules like evaluate the effectiveness or functionality and fix any mistakes.

[@diplobla]

the *evil laughter* of Claude at 34:37 ...😱

[@0hellow797]

Watching the cactus grow at 12:20ish was cute :)

[@JGR0ZNI]

14:24 that little greenhouse the 2nd level of o1 and the wells looked great, very promising, and quite unique

[@DiluviumEyesofThunder]

AI is like that guy who reads a lot of textbooks but has never touched grass. Mind ripping in theory, but no contextual understanding of how or why to put it in practice. He's read tomes on ethics, and ran all the thought experiments, but he has no vested interest in society or for higher order effects of any actions he might take.

[@robloxmania627]

I figured with the intro of this video, the question must be asked, Is it possible to make an AI Herobrine mod with this technology eventually?

[@SpicyMelonYT]

Yes. Very damn cool idea too. I never thought of that but it most certainly would be the perfect use case for it. The hardest part and the part that needs to be solved to actually have a good herobrine mod would be to have the AI either orchestrate prebuilt scary moments OR even more impressive would be to train a model to actually understand the elements and timing of being a scary eerie entity like herobrine. An AI would be able to make decisions based on the player(s) in the game. Like these agents it would be informed of information and situations within the world. Like a big chart of values perhabs. One item might be the time since the player has walked through a door? Or how many times they have killed a mod or how close is the players view direction towards herobrine itself. things like this. And with that can learn to run actions that make the scariest experience. I know its not exactly fair to say herobrine is something to be scary. Hes more than that. But most people like myself would want that and most herobrine mods do that already so it would fit and be like the next evolution.