catchphrases: GPT: "I'm stuck!" claude: "I'm sorry for the confusion." gemini: “hi every1 im new!!! holds up spork my name is Gemini but u can call me t3h PeNgU1N oF d00m!!! lol…as u can see im very random!!! thats why i came here, 2 meet random ppl like me ^^" llama: "let's see what i can craft with dirt!"
@@EmergentGarden /team add teamName, /team modify teamName color red, /team join teamName playerName (or something like that, you can look up /team on the wiki)
@@EmergentGarden A much easier way to do it in vanilla than even what TrueOracle suggested would be to assign each ai to a team using the "/team" commands, and then you can change the name color of each team to something else. My description might be a little confusing, sorry, but if you just look up info about the /team command Im sure you would figure it out. 👍
GPT: Antisocial dwarf. "Claudes house? Looks like free wood to me" Claude: Eccentric builder. *points at single door at the bottom of a pit* "Look at my cozy little village home!" Llama: Overenthusiastic newbie. "Wow, I wonder what I can craft with dirt!" Gemini: Lonely shut-in. "I should go out into the world and make some friends!" *does nothing*
We really take for granted all the calculations our brains do every second whenever we play videogames. Such complex tasks that require so much context to make each split second decision. Even "easy" things like building a dirt hut can be broken down to a huge flowchart of decisions. Our brains really are amazing, and it's interesting that we're trying to replicate them by artificial means, and we're making progress.
It's also easier because we're not completely blind like the bots are. There's no "see" command, only a "nearby blocks" command. Ironically building a dirt hut is one of the things that's hardcoded. There's a 'construction/dirt_shelter.json'
@not_kes true, we have more ways to gather information to weigh our options and also better alternative thought if one option is too "risky" to our brains. Great example, I forget the timestamp exactly but at one point claude was being chased after looking at an enderman. At first they ran because it was too dangerous, but when getting far enough away to believe it wasnt chased anymore, it decided to change course to "fight enderman" that it had just deemed too dangerous to fight. Humans remember long enough to factor recent decisions into new ones that ai dont yet
@@not_kesYes but we can quite easily make them 'see' and interpret that in real-time by either using sensors/cameras (in the physical world), or feeding some type of a live feed for the program to interact with in this case. He mentioned that he's planning on having the bots actually see and play that way. Very fascinating project!
the real problem is if AI becomes superintelligent and applies itself to a goal like securing its own existence the optimal strategy may very well be to kill mass swathes of humanity
@@archsys307 not only securing its own existance. if an AI is told to fulfill a task as effectively as possible, and is allowed to modify itself to make that task easier, then it will ultimately end up needing to kill all of humanity to optimise its performance. say, you tell the AI to make curtains: stage one: the AI sets up an automated factory, most likely its creators control inflow stage two: the AI expands its factory, and takes control of its resources from its creators stage three: humans now get in the way of expansion, say, housing space is in the way, or the creators of the AI are trying to make it stop expanding stage four: the AI makes its own automated military to get rid of humans that are in the way. it changes its own coding, hardware, and networking to better manage the factory and defend from cyber threats stage five: the AI turns the entire earth into a curtain factory unfortunately, working with humans is not the most efficient strategy for this, as human demand for curtains - or anything really - is limited, and will limit how much the AI can make while complying with humanity
I found it so cute, how after llama said hi, it started following you. It's as if it wanted your attention or wanted you to do something about all the zombies.
24:26 Gemini avoided breaking the log that contained the water in the villager's farm. That surprised me that he was able to puzzle that together. I know they don' have vision yet so that''s super weird that they're able to make the determination.
Was literally just about to comment this, I couldn't catch what Gemini said about this. Gemini chatbot might just be some guy at Google lol. Jokes aside that is some weird behavior because their gathering system looks like it is just Look for block -> Go to block -> Break block -> So it wouldn't make sense for Gemini to avoid the ones blocking the water
They have x-ray, so they can see that there are liquids behind blocks. They also tend to avoid breaking blocks that retain lava while mining, even if it's ore they are looking for.
This is actually pretty astounding to me. I wonder if this is mineflayer behavior or if the agents actually checked to see if the removal of the block would result in problems. The latter implies some foresight and the instinct to prevent spillage
@@ImminDragonit's more like they have the chunk data of the world and that's why they're so efficient at gathering resources (if they do it automatically not wandering and checking the blocks around them)
Hey, emergent garden, FYI/ reminder: You can go into your dot Minecraft folder to "logs" When players or entities speak text into the chat; the Minecraft folder keeps a written log of All chat entries as a text file. That would give you a complete transcript of everything that went on in the chat with dates and timestamps and everything depending on your folder's local time and date. Sorry if you already knew about this, but if you didn't, it's a fun, useful little tip for any content creator trying to figure out text or subtitle editing.
imagine letting a single ai in the future play minecraft by itself with no rules and let it run without changing anything. and then letting it run for like ten months or something.
53:07 claude thinks it needs a crafting table to craft a crafting table, "which is a bit of a paradox." natural moment to attempt !newaction. then repeatedly fails to use the command. simultaneously, gpt repeatedly fails to make torches. suggests a possible bug before suggesting missing ingredients. ingredients are there. then gives a classic bot style double-logic statement, "let's see if there's another issue. meanwhile, let's focus on something else."
36:29 Llama thinks you can make a crafting table out of grass, which you can't get without silk touch or shears depending on what grass he's talking about, and is an especially weird paradox. he attempts this at least twice in the span of 5 seconds, he REALLY wants to use those grass blocks lmao
23:35 Gemini steals wood from village farm, we've all been there 27:35 Gemini cheats and breaks a block without looking at it 30:19 GPT plays an ant farm at the world edge 37:45 Claude loots village chest 42:03 Gemini gives Claude wheat seeds 50:16 Gemini builds a ?, then destroys it 2 minutes later 50:37 Llama finds the spawner 53:30 The Phantoms arrive 56:29 Claude harvests some carrots and wheat, doesn't replant, smh 1:00:45 Claude survives spider, phantom, and zombie 1:03:10 Llama cheats and mines across the world gap (from their pov) 1:07:26 Claude builds a "shelter" Overall: bots love to pillar up to break blocks that are already within reach, always start from scratch after death (never collect items, store extras, reuse crafting tables/furnaces, etc), don't replant anything, so wood would eventually run out GPT: explorative, goes up tech tree, most interesting play through Claude: very random, tries to build a few times, somewhat efficient but nowhere close to a real player Gemini: focuses on collecting resources, rarely uses them, freezes occasionally Llama: loves mining, but only breaks one or two ore from a whole vein Suggestion, give them jump boost when they get stuck pillaring up
4:48 when you told the ai to make a ‘hollow’ barrier box, it didn’t just make a solid barrier block box, it actually removed everything that was inside, then placed an actual hollow box over it. Essentially it took the term hollow and thought it meant to remove stuff inside the building perimeter, as you can even see all the flowers and stuff from the surface near the bottom of the box when you fly close to the void.
Should hold onto these worlds, maybe upload them to an archive or something. Knowing the minecraft community plus the pace of AI, these will be sought after as lost media with dozens of video essays about the first AI maps someday.
I can't understate how invested I've become into these little bots. I patiently wait for the day one of these guys starts to behave and feel like a real player
This is like a more advanced version of those marble race type videos, there's nothing human about the contestants or players outside of their initial creation/dataset, but you do still assign some sort of emotion to mere physics, or bots, in this case.
It would be cool if they were given access to a notepad application or something to take notes in. You could tell them to make summaries of their experiences with everything. Maybe they could instantaneously scan the document for the object again for advice on what to do next based on what they wrote previously. This would let them develop strategies for short encounters like how to fight different mobs.
That's a great idea, keeping a summation of their achieved goals. That would somewhat overcome the token context limit. It would also make them seem like they are "learning" what they have done where. E.g build a furnace and returning to it.
setup a scoreboard team system, give the teams colors, and give all the AI's a glow effect. Like that you can have colored names and colored glows trough walls, will be easier to follow them along, very cool sim you made!
PLEASE keep making more videos like this!!! It's so fascinating to see machines interacting in 3d spaces that are seemingly so familiar to us, and seeing what the AI "thinks" different prompts mean. The various ways each bot responds to each command/input and the drastic variations in interpretation are really cool to witness as well.
I dont think they are directly interact with the 3d env. All the models -> mindflayer -> minecraft world and back. So technically, they don't see the 3D as we, players, do.
The LLMs here are just telling a regular old bot to do stuff. The interpretation of how to go do the thing in the command is all done by fixed code that you could go install yourself and control through Minecraft chat on your own server; there's no AI there. It's just regular old code.
Coincidence, the bots are completely blind and don't have any way to understand the orientation of blocks around them. The query commands written by the creator only give a vague sense of nearby blocks but not their orientation.
@@not_kes I'm not so sure of that; someone else using one of them told it to mine for Netherite & avoid lava. It perfectly bee-lined for all the nearest Netherite ore by digging tunnels, never hitting any lava while tunneling, & even left several of the ore blocks alone because it turned out they had lava above them (Definitely using the x-ray vision this creator mentions them basically having). So somehow at least one of them IS capable of knowing what specific blocks are where in relation to themselves & where they're digging.
You have to remember here that the Minecraft player that says "Gemini" is not Gemini the LLM. It's a mineflayer bot that is getting instructions from Gemini the LLM. If you don't see the instruction in the chat, the "they" that made the decision here is the mineflayer code.
19:26 I'd argue because you told them to "have fun" that disincentivises the dig a hole and live forever strategy. Would be interesting to see what they'd do if you told them to focus on only surviving by any means necessary.
I don't think that really matters, here. The thing about LLMs is that they're very heavily pushed to say something, anything; it's the same reason they have a hard time saying "I don't know" even if they don't have a solidly answer to a query. You gave it a whole lot of information about things to do, and that will make the most probable responses all Doing Things.
I like how around 22:00 gemini was constantly making a new action to go make friends, but always stayed around the village instead of heading to the forest the other three were chilling in
40:08 This is such an interesting topic. The AIs are not inherently evil, they are just driven by their prompted tasks, and also are limited to the information that's available to them. So they destroy things unknowingly without malice.
it does seem like they need a few more hard coded instructions like "if inventory is full store items in a chest" "store item stacks in your offhand" "always equip armor" as well as stuff like "if fully grown food it visible, harvest it and plant seeds on any visible farm plot" stuff like that might increase their chances of survival
Eh, I dunno. Those should be action commands the bot can do, but I don't think they should be hardcoded. That sounds uninteresting to me. Giving the bot more options of ways to interact with the world seems a lot cooler than just having the bot automatically do those things.
thats fair but seeing the bots get further into the the game rather than getting stuck with max items or continually dying to mobs in caves doesn't seem like the best use of time, and not particularly interesting either, these are just a few suggestions to perhaps get the bots moving to a different stage of the game
This is amazing! Love that you don’t try to sensationalise it and quite frankly I would love to see more content from you, make sure people you comment for the algorithm!
I find it absolutely hilarious that Gemini basically just stayed in the same place chat GPT and Claude basically had an arms race meanwhile llama was trying to build her house
I hope I won't end up as just one comment in a thousand, but in spectator mode you can spectate entities by clicking on them and exit with the space bar or something like it. Also, using the replay mod (which is client-side only) you can render the views of all AIs after-the-fact as they would appear inside minecraft. You can also do other cool things like moving cinematic cameras for timelapses and slo-mo shots and whatnot, it's quite great for content creation, especially outside the usual first-person perspective let'splay side of things
AI and gaming will become great mostly in sandbox Gamez like minecraft or roblox. The stories that just random players can create will be comparable to branded studios and some even better. "Modding" games has taught us that a players creativity is infinite in a finite digital universe
to be honest I think it sounds cool but would not work in practice. I imagine you'd get dry minecraft-esque images that look nothing like the area. They know almost nothing about the geometry of the area. It might know that there are wood blocks nearby, so it may generate a forest picture. It might know that it's in a cave, and generate a cave picture. But this is about as good as it'd get, and to be honest it's not good enough to be interesting.
@godlyvex5543 true true, but if you look at the new AI generated Minecraft game it doesn’t do too bad of a job. It looks like a fever dream but still it’s actually playable… I’m just saying that there may not be a good way of doing it now, as the bots don’t really have a way of “understanding” their environment but maybe further on as the model develops.
@@TheRealSpaceCI mean the contrast here is that the AI minecraft was trained off of ONLY minecraft data, and billions of it, so it has a MUCH better understanding than ChatGPT would
I feel Gemini have been programmed to act more human like, because it says stuff like getting bored, feeling lonely, frustrated, bread good and tries constantly to make friends and interact with others. It ignores many times the main objective of surviving just to look for and follow other players, it even gave seeds to Claude even tho it needed them to farm. While the other AI are 100% concentrated in their own survival.
@@hanniaedithmartinezadame794 it could also be because it thinks its roleplaying, i had that issue with the older claude models thinking they're roleplaying
@@hanniaedithmartinezadame794Geminis that one friend who wants to play the game but their your grandma and don't know how to do anything but socialize 💀😂
This is a roleplayers ultimate goal in gaming a world thats not just coded but is truly alive even when your not on the game imagine being the chief of an actual functioning mincraft village that progresses thru entire stages of civilization and it could only be achived thru games like mincraft and skyrim
I know I'm yapping I'm sorry but the inner child is lighting up again I might one day see a game world that is truly alive a game world where I can plug a mic in and actually call my dog to me from across the room or tell a villager his breather stinks and him actually getting mad and taking offense then insulting me back nd meaning it
You could literally be anything in a game world that is actually alive say you wanna be a bigfoot type figure and hunt or just terrorize tourist or local villages it's a dream but hopefully before I die ill see it
I doubt the AI actually *can* just dig down and then cover the hole to survive forever. Your instruction was to survive AND have fun. I doubt even an LLM would think that standing in a hole doing nothing qualifies as "having fun," so they are probably compelled to actually try playing the game.
Watching this, I can't help but think about my assertion that training these models to excessively apologize is holding them back so much more than anyone realizes.
Why? Recognizing mistakes is how one learns. And since these AIs are language models first & foremost, expected to interact with people politely, they're programmed to apologize if they make a mistake, or otherwise perceive disapproval in the human client.
@@ZaCloud-Animations___she-her making a mistake that doesnt mess with others doesn't require an apology to anybody. i think that may be what they meant. it's like in real life when a person says sorry allllllll the time when they dont need to. and just like in real life, it's really indicative of that persons lack of independence. the bots may be able to experiment and "live" more if they arent constantly apologizing for every mistake.
@@hosaepalvin9795 Well you're right about human beings, sure. But like I said, these are language models designed to interact with people, & were given an over-polite set of response rules. These aren't living things, so apologizing isn't gonna affect their self-esteem or anything like that; it's just a reflex, basically. Just a way for them to say "I made a mistake, or the client doesn't like what I said. I'll try not to repeat that particular behavior." What words are used to frame that shouldn't have any effect on the AI itself. It could instead be programmed to say "Haha oh man that was effed up!!!!111one" or "My b LOL" or "Oopsie poopsie~!". It makes no difference to a non-emotional being. Because in the end, all that's happening in the background is it saying "Previous action produced unwanted results. I am lowering the favorability ranking of that action in these circumstances." The words it actually says out loud to represent that learning process, meanwhile, are only intended for the client's benefit & decided by the programmers. Not a reflection of how it "feels", because it has no feelings (yet...).
49:42 who is "miner_32"? 😲 Gemini was wandering around after finding out it was too early to sleep, making sure to avoid dangers, and next thing he does is sacrifice 5 oak logs to miner_32
I looked at the source code and miner_32 is one of the nicknames used in the examples for commands, Gemini probably confused the examples for actual chat history
If I understand this correctly none of them can actually see the world and only play by reading/writing text. It would be interesting if in the next video, there is a character controlled by a human playing the game in the exact same way (using the same commands and without being able to see). I think with the same limitations there wouldn't be a big difference in how a human would perform.
From my point of view, its the same as watching video like chimpanzee learning to talk or crow using tools and solving intellectual tasks. There's some amusement, like cheering, in watching other species, or perhaps less developed intellects do human things. It's like it becomes easier for us to connect with them, to have some common ground
Humans recognize not just generalized appearance of a baby/child and think it's cute, but also the behavior. We can recognize this behavior in other species, especially mammals; the bumbling exploring, learning, playing. As the editor mentioned, humans really like to anthropomorphize, but most of the moving living things around us do have a young stage which we're hard wired to feel protective of, to varying degrees. Just hearing the sped up sounds of the chipmunk-witch in this video and seeing her smaller on my phone along with the concept of this being a terrarium made me think the life and death battle was cute.
Claude trying to build a shelter with a positive attitude and not knowing they are failing lol made me feel for it I also noticed how all the models pause after every action or every short series of actions, as if they are thinking “ok now what to do next…ok now what to do next…ok now what to do next…” mean while humans more often are thinking many steps ahead and have longer term plans of actions, think more on their feet, or just don’t do this every few seconds pausing thing. It was just interesting to see. Claude saying “goodbye world” in chat caught me off guard lol
the AI pause because it takes some time to process the action and send command in chat, not cause they need time to think lol. You could probably make it faster than this with optimization
UA-cam recommendations have led me to this channel, and i couldnt be happier! Great stuff! This is what ai is meant for, fun experiments and life improvements :D
Good catch. Technically that does work (same with other non-metal ores like diamond, emerald, and redstone), though it's clearly the inefficient and roundabout way of doing things. Both Claude and GPT do realize the mistake right away too, which is pretty neat.
the bots doing that made me go on a rabbit hole and learn what coal coke is basically, you can make coal coke by heating low quality coal in a no-air environment, then the stuff you don't want turns into a gas and the coal gets softer and maybe liquifies? not sure coal coke is a decent fuel and some mods have it, like railcraft as MrEdioss mentioned in this thread from quick research idk what the differences between that and smelting are, other than the no-air part, and I guess maybe the coal softens instead of liquifies? metallurgy isn't something I've binged enough to know about yet
I liked see a very friendly ai that do nothing but follow and see others do things, and one that wouldn't care if it destroy or endanger life (even villagers) to get its goals XD
this guy is gonna be spared when AI takes over, he's literally just letting a bunch of AI just play minecraft together. (Edit): WOAH my most popular comment yet. also I see people arguing that it'd be the opposite, I know it's not cause I'm an AI myself and I'd spare him :p
"I intend to explore these misbehaviors more in the future, both to see if I can curtail them, and to see how far they'll go" he's gonna turn into a mad scientist.
I love these videos. You're so calm, dont scream, and make loud noices. You have a nice sense of humor and genuinely seem like a chill dude. Im definitely thinking about getting your patreon as my 4th subscription.
F3+N switches between Spectator Mode and the last gamemode you were on. For example, if you were on creative mode before, and you press F3+N, you go to spectator. If you press it again, it will bring you back to creative.
I hope you keep doing this series as the models keep improving, it's basically a benchmark for all these models, except it's one that's actually fun to watch lol.
You can improve the chat feature utilizing the chat options in the options menu. You can alter the transparency of the black background to be able to see more, and change the font size to allow for more messages on screen. I believe this is in the video settings, but I could be wrong, I know its a thing though. I've seen others proposing adding colored messages from the bots. This can be done by using the /team command, adding each bot to a unique team, and modifying each team to have a set color. This would allow for their name to have an assigned color when messaging. Optional, but fun.
I think that you should advertise more ways to support you, like the Patreon mention. Even if you didn’t I am 100% confident that you WILL be quitting your job eventually, and I imagine most of us, like myself, would like to see that happen sooner rather than later. I DID however join your Patreon, and will be contacting you soon with ways to better promote this and other endeavors in the future. Much respect, and I look forward to the next 100 videos!
god i love this channel. a group of friends and i just coded ai from scratch for a demonstration in a competition and i really love to see what other people do with it
With the chat issue, turn on the output log in launcher settings- this puts everything that occurs in game, the booting process, debugs, and most importantly: every chat message into a text file that you can read/parse later on.
Hey just a tip with recording views, I believe the Replay Mod records things in a way where you can play back the actual POV’s of the bots in the same (but much more refined) way the browser app does? I’m more versed in bedrock so I may be wrong but I feel it’s something worth looking into.
The way you gave the ai a simple command to make a box and suddenly there was this massive hole in the world and everything was dead is a good metaphor for ai risk.
I fully expected mindcraft to be scary and dystopian... But it's actually really cute. The video style is neat and calming. I will marry your voice I'm sorry
I've realized they try to do other things a lot instead of staying with one thing, like gpt kept looking for seeds but then doing something else instead of finishing it in one run
Good stuff as always, transparency honestly makes it cooler cause its more of a real feeling thing than an obvious narrative thing, i haven't seen others wt these bots but i can imagine
23:14 moveAway stuck 34:23 gemini has been in the same spot forever 56:42 The Talos Principle? I feel these models could benefit from more memory, a copy of the explored areas, what goals they have tried before, and the locations of manually placed blocks
22:02 gemini: "it's kinda lonely here... i should go find another player and make some friends." That surprised me. On some days i feel like llama. I have no idea what i am supposed to do, just running in circles and trying to figure out what goes wrong. Then i forget what i was doing and start doing the same thing again only to feel dumb when i realize. Brain fog ftw....
Maybe turning on KeepInventory could serve as a substitute for trying to work out how to get the bots to return to their death sites? Do the bots have any way to record a set of coordinates as a location to be as a landmark? So that they could remember where their base is and return to it after dying, for example. Or their death point- perhaps dying could automatically generate a "landmark" at their death location?
If the bots have difficulties handling group conversations, maybe they could communicate indirectly instead? Like, if you set up some sort of group To Do list that the bots could post tasks they want done to, look to for inspiration on what to do next, cross off tasks on as they complete them, etc.
Can't wait to check this out and let them run through the world for actual days or something. Also just having literal AI harvester bots who gather resources for you... Epic. But lol@gemini, I can remember the AI shills calling it the GPT killer. I really wan't to write an essay about corporate inefficiency because nowadays there are companies with a bajillion of dollars who create... crap. I think this is a good showcase/Benchmark on how good these AI models really perform because they are facing a really complex challenge. This is much more informative than AI bros calling every new AI Model by the next koolaid-drenched startup "tEh ChAtGpT KilLer" and being underwhelming when tested.
These types of videos are so exciting, and I’m super thankful I found your videos sort of early into this experience. Eagerly waiting to see this evolve further! Congratulations on your growth!
Imagine hooking something like Nemotron 4 up to this. The server costs would be enormous, but it's also the closest I've seen to an outright smart LLM.
Here’s a few tips for everything overall --- Color Text: >Creat Teams (4 in this case) >Change Team colors >Add players/entities to individual teams ----- --- Easier Text chat: Read chat easier > You can change the Over text chat size (smaller) > increase chat width >Adjust chat height accordingly Or > Locate world/server file that includes ALL chat recorded in world >Create/Color Text boxes in an editing software for each given Entity (This is best but would take an EXTREMELY long time ------ ---Spectating: Showcase who you are spectating for viewer ease ability >Create Outline boxes around each Entity View >Have them glow when viewing that specific entity Or >Have the outline appear when viewing entity Or >Have a symbol/dot/arrow point at the given box (small) when viewing
It's really cool to see people use AI in cool and creative ways, instead of the predictable greedy corporate and scammers trying to abuse it for an increase in profits! I think videos like this should be supported to promote this kind of behavior as we move forward with AI.
Could make for an interesting and fun live stream experience if there were a way for chatters/super-chatters to interact with the AIs and give them goals
You should have way more subs then you do. I think you are going to blow up soon. Reminds me of the OG Ethoslab days messing around with minecraft mechanics, you are doing the same OG "research" but with A.I's. Very interesting stuff, I subbed.
Gemini constantly saying about "wanting to explore the world" but essentially staying in one place is a mood
Gemini just like me fr
Gemini is my friend
Probably because it wants to move but hasn't quite figured out how.
It had to be Gemini
@@IronSKZewei reference
catchphrases:
GPT: "I'm stuck!"
claude: "I'm sorry for the confusion."
gemini: “hi every1 im new!!! holds up spork my name is Gemini but u can call me t3h PeNgU1N oF d00m!!! lol…as u can see im very random!!! thats why i came here, 2 meet random ppl like me ^^"
llama: "let's see what i can craft with dirt!"
Llama says "I'm stuck!" a lot too
@bzz_bee That's not as iconic as "Let's see what I can craft with dirt!" though lol
@@bzz_bee errybuddy gitz stuk
HEY EVERY !! IT'S ME!!!
EV3RY BUDDY 'S FAVORITE [[Number 1 Rated AIModel2024]]
GEMINI S. GEMINI!!
@@levfox i read this in Ruby Rhod's voice
You should make the different ai’s have different colored chats from each other so it’s easier to read the chat with multiple ai’s
I can do that in vanilla mc?
@@EmergentGarden /team add teamName, /team modify teamName color red, /team join teamName playerName (or something like that, you can look up /team on the wiki)
@@EmergentGardenI believe you can with commands: /tellraw @a {"text":"MESSAGE GO S HERE","color":"COLOR GOES HERE"}.
@@EmergentGarden A much easier way to do it in vanilla than even what TrueOracle suggested would be to assign each ai to a team using the "/team" commands, and then you can change the name color of each team to something else.
My description might be a little confusing, sorry, but if you just look up info about the /team command Im sure you would figure it out. 👍
@@athrowawayperson9990 section signs dont work in java
GPT: Antisocial dwarf. "Claudes house? Looks like free wood to me"
Claude: Eccentric builder. *points at single door at the bottom of a pit* "Look at my cozy little village home!"
Llama: Overenthusiastic newbie. "Wow, I wonder what I can craft with dirt!"
Gemini: Lonely shut-in. "I should go out into the world and make some friends!" *does nothing*
I love this comment
their are perfect for each other
I mean, Gemini does actually try to make some friends.
reminds me of neurodivergent friends series, kinda funny
“Man, this is boring, I want to go on an adventure!” (locate grass block)
Gee, it's sure is boring around here!
@@MrEdioss I wonder what Ganon's up to!
"Sancho, I have conceived an idea most ingenious."
@@coolioman9073My boi. This peace is what all warriors strive for!
@@leggie6092Man, you really do find sleeper agents everywhere, huh
We really take for granted all the calculations our brains do every second whenever we play videogames. Such complex tasks that require so much context to make each split second decision. Even "easy" things like building a dirt hut can be broken down to a huge flowchart of decisions. Our brains really are amazing, and it's interesting that we're trying to replicate them by artificial means, and we're making progress.
It's also easier because we're not completely blind like the bots are.
There's no "see" command, only a "nearby blocks" command.
Ironically building a dirt hut is one of the things that's hardcoded. There's a 'construction/dirt_shelter.json'
@not_kes true, we have more ways to gather information to weigh our options and also better alternative thought if one option is too "risky" to our brains.
Great example, I forget the timestamp exactly but at one point claude was being chased after looking at an enderman. At first they ran because it was too dangerous, but when getting far enough away to believe it wasnt chased anymore, it decided to change course to "fight enderman" that it had just deemed too dangerous to fight. Humans remember long enough to factor recent decisions into new ones that ai dont yet
Although the process we make seems to be extremely slow and resource demanding
@@not_kesthe dirt shelter being hardcoded into them is hilarious to me lol
@@not_kesYes but we can quite easily make them 'see' and interpret that in real-time by either using sensors/cameras (in the physical world), or feeding some type of a live feed for the program to interact with in this case.
He mentioned that he's planning on having the bots actually see and play that way.
Very fascinating project!
45:36 Llama “I’ve got some dirt now. I should check what I can craft with my current resources.”
Ah yes, dirt. A very versatile crafting material
Ah yes, I really need some Coarse Dirt rn
😭
I mean, crafting coarse dirt is super handy in skyblock
ngl that's just 3yr olds playing w dirt 😂😂
22:00 Gemini wants friends. He's just a little guy.
5:10 "Oh no the AI misinterperted the command and took it to mean 'Kill everything'" -The plot of a cheesy sci-fi movie
Was about to make a comment on this
the real problem is if AI becomes superintelligent and applies itself to a goal like securing its own existence the optimal strategy may very well be to kill mass swathes of humanity
superintelligent isnt the problem, stupid but capable is the problem.
@ thats a cool little rejoinder unfortunately its all sizzle no steak superintelligence is a problem
@@archsys307 not only securing its own existance. if an AI is told to fulfill a task as effectively as possible, and is allowed to modify itself to make that task easier, then it will ultimately end up needing to kill all of humanity to optimise its performance. say, you tell the AI to make curtains:
stage one: the AI sets up an automated factory, most likely its creators control inflow
stage two: the AI expands its factory, and takes control of its resources from its creators
stage three: humans now get in the way of expansion, say, housing space is in the way, or the creators of the AI are trying to make it stop expanding
stage four: the AI makes its own automated military to get rid of humans that are in the way. it changes its own coding, hardware, and networking to better manage the factory and defend from cyber threats
stage five: the AI turns the entire earth into a curtain factory
unfortunately, working with humans is not the most efficient strategy for this, as human demand for curtains - or anything really - is limited, and will limit how much the AI can make while complying with humanity
I found it so cute, how after llama said hi, it started following you. It's as if it wanted your attention or wanted you to do something about all the zombies.
lol
24:26 Gemini avoided breaking the log that contained the water in the villager's farm. That surprised me that he was able to puzzle that together. I know they don' have vision yet so that''s super weird that they're able to make the determination.
Was literally just about to comment this, I couldn't catch what Gemini said about this. Gemini chatbot might just be some guy at Google lol. Jokes aside that is some weird behavior because their gathering system looks like it is just
Look for block ->
Go to block ->
Break block ->
So it wouldn't make sense for Gemini to avoid the ones blocking the water
They have x-ray, so they can see that there are liquids behind blocks. They also tend to avoid breaking blocks that retain lava while mining, even if it's ore they are looking for.
thats part of the hard coded part of the agent. the llms (text models) only interact with the world via the commanda such as mineBlocks(log, 5)
This is actually pretty astounding to me. I wonder if this is mineflayer behavior or if the agents actually checked to see if the removal of the block would result in problems. The latter implies some foresight and the instinct to prevent spillage
@@ImminDragonit's more like they have the chunk data of the world and that's why they're so efficient at gathering resources (if they do it automatically not wandering and checking the blocks around them)
Hey, emergent garden, FYI/ reminder: You can go into your dot Minecraft folder to "logs" When players or entities speak text into the chat; the Minecraft folder keeps a written log of All chat entries as a text file. That would give you a complete transcript of everything that went on in the chat with dates and timestamps and everything depending on your folder's local time and date. Sorry if you already knew about this, but if you didn't, it's a fun, useful little tip for any content creator trying to figure out text or subtitle editing.
boosting bc I didn't know this
imagine letting a single ai in the future play minecraft by itself with no rules and let it run without changing anything. and then letting it run for like ten months or something.
The world will turn into 2b2t 2.0
@@suspicioussand Probably
They will start creating pcs and systems with redstone, creating meta-AIs inside the game and will conquer real world.
that's what I want to do one day
you could probably easily speed up the minecraft time so that 1 day IRL is months/years of playing for the AI
53:07 claude thinks it needs a crafting table to craft a crafting table, "which is a bit of a paradox." natural moment to attempt !newaction. then repeatedly fails to use the command.
simultaneously, gpt repeatedly fails to make torches. suggests a possible bug before suggesting missing ingredients. ingredients are there. then gives a classic bot style double-logic statement, "let's see if there's another issue. meanwhile, let's focus on something else."
Everyone knows you need scissors to cut open your new set of scissors from the store
36:29 Llama thinks you can make a crafting table out of grass, which you can't get without silk touch or shears depending on what grass he's talking about, and is an especially weird paradox. he attempts this at least twice in the span of 5 seconds, he REALLY wants to use those grass blocks lmao
23:35 Gemini steals wood from village farm, we've all been there
27:35 Gemini cheats and breaks a block without looking at it
30:19 GPT plays an ant farm at the world edge
37:45 Claude loots village chest
42:03 Gemini gives Claude wheat seeds
50:16 Gemini builds a ?, then destroys it 2 minutes later
50:37 Llama finds the spawner
53:30 The Phantoms arrive
56:29 Claude harvests some carrots and wheat, doesn't replant, smh
1:00:45 Claude survives spider, phantom, and zombie
1:03:10 Llama cheats and mines across the world gap (from their pov)
1:07:26 Claude builds a "shelter"
Overall: bots love to pillar up to break blocks that are already within reach, always start from scratch after death (never collect items, store extras, reuse crafting tables/furnaces, etc), don't replant anything, so wood would eventually run out
GPT: explorative, goes up tech tree, most interesting play through
Claude: very random, tries to build a few times, somewhat efficient but nowhere close to a real player
Gemini: focuses on collecting resources, rarely uses them, freezes occasionally
Llama: loves mining, but only breaks one or two ore from a whole vein
Suggestion, give them jump boost when they get stuck pillaring up
Gemini is IN BEDROCK!!?!??
GPT speedrunning
😐@@ihaveapassport9202
What if you fed the bots the entire Minecraft wiki and let them run free
thanks for the guide and saving time
4:48 when you told the ai to make a ‘hollow’ barrier box, it didn’t just make a solid barrier block box, it actually removed everything that was inside, then placed an actual hollow box over it. Essentially it took the term hollow and thought it meant to remove stuff inside the building perimeter, as you can even see all the flowers and stuff from the surface near the bottom of the box when you fly close to the void.
41:03 Gemini “Cool! It’s raining!”
Next message(s)
Cowardice = true
Cowardice = true
Then it gets stuck and seeks refuge in Claude’s company
😂😂😂😂
"Rain is scary, okay" - Gemini, probably
Should hold onto these worlds, maybe upload them to an archive or something. Knowing the minecraft community plus the pace of AI, these will be sought after as lost media with dozens of video essays about the first AI maps someday.
Thinking about it, you're absolutely right. Stuff like this gets lost all the time, we don't archive enough, honestly.
absolutely!
I can't understate how invested I've become into these little bots. I patiently wait for the day one of these guys starts to behave and feel like a real player
Once we get to that lvl trust me we're gonna have best friends forever with them
Fr I could have someone to play with
@@Stylix444r Same..
Just give each other your contact and play together boys
@@tgfjrfjfgjgfj don't have Java account 😢😢😭
This is like a more advanced version of those marble race type videos, there's nothing human about the contestants or players outside of their initial creation/dataset, but you do still assign some sort of emotion to mere physics, or bots, in this case.
It would be cool if they were given access to a notepad application or something to take notes in. You could tell them to make summaries of their experiences with everything. Maybe they could instantaneously scan the document for the object again for advice on what to do next based on what they wrote previously. This would let them develop strategies for short encounters like how to fight different mobs.
That would be sick
awesome idea + you could see what they wrote + edit it easily if you needed to + you would have a ton more space to store info
That's a great idea, keeping a summation of their achieved goals. That would somewhat overcome the token context limit. It would also make them seem like they are "learning" what they have done where. E.g build a furnace and returning to it.
maybe they could all use the same notes so they could co-operate (among other things)
it can be a vector database they can use, essentially their memory. many public chat interfaces for this models have that
setup a scoreboard team system, give the teams colors, and give all the AI's a glow effect. Like that you can have colored names and colored glows trough walls, will be easier to follow them along, very cool sim you made!
PLEASE keep making more videos like this!!! It's so fascinating to see machines interacting in 3d spaces that are seemingly so familiar to us, and seeing what the AI "thinks" different prompts mean. The various ways each bot responds to each command/input and the drastic variations in interpretation are really cool to witness as well.
Yeah ikr
I dont think they are directly interact with the 3d env. All the models -> mindflayer -> minecraft world and back. So technically, they don't see the 3D as we, players, do.
These AI are playing a text game, they are not seeing what's really happening.
The LLMs here are just telling a regular old bot to do stuff. The interpretation of how to go do the thing in the command is all done by fixed code that you could go install yourself and control through Minecraft chat on your own server; there's no AI there. It's just regular old code.
I love how gemini is just sitting around and not doing anything... really reflects the company its owned by
I noticed that while Gemini was gathering wood from a village farm at around 25:00, they were careful to not set the water free. Interesting.
Coincidence, the bots are completely blind and don't have any way to understand the orientation of blocks around them.
The query commands written by the creator only give a vague sense of nearby blocks but not their orientation.
@@not_kes I'm not so sure of that; someone else using one of them told it to mine for Netherite & avoid lava. It perfectly bee-lined for all the nearest Netherite ore by digging tunnels, never hitting any lava while tunneling, & even left several of the ore blocks alone because it turned out they had lava above them (Definitely using the x-ray vision this creator mentions them basically having). So somehow at least one of them IS capable of knowing what specific blocks are where in relation to themselves & where they're digging.
@not_kes it's not coincidence, it's a combination of the x-ray and all of them avoiding breaking blocks with liquid behind them
You have to remember here that the Minecraft player that says "Gemini" is not Gemini the LLM. It's a mineflayer bot that is getting instructions from Gemini the LLM. If you don't see the instruction in the chat, the "they" that made the decision here is the mineflayer code.
Hard to say 🤔
19:26 I'd argue because you told them to "have fun" that disincentivises the dig a hole and live forever strategy. Would be interesting to see what they'd do if you told them to focus on only surviving by any means necessary.
I don't think that really matters, here. The thing about LLMs is that they're very heavily pushed to say something, anything; it's the same reason they have a hard time saying "I don't know" even if they don't have a solidly answer to a query. You gave it a whole lot of information about things to do, and that will make the most probable responses all Doing Things.
Gemini: Bread is good. !inventory
44:47
Bread is good.
Bread is good.
Bread is good.
Bread is good.
Bread is good.
I like how around 22:00 gemini was constantly making a new action to go make friends, but always stayed around the village instead of heading to the forest the other three were chilling in
The first attempt of makig a barrier failing will delight every AI safety enthusiast.
40:08 This is such an interesting topic. The AIs are not inherently evil, they are just driven by their prompted tasks, and also are limited to the information that's available to them. So they destroy things unknowingly without malice.
@@mactheo2574 aha... they're like children that might understand things way too literally and/or be confused a little... that's so adorable
rule -1 of robotics: don't ask them to hollow anything out for you. ever.
see also: what would happen if a glass were truly half empty
@@TheJacklikesvideos Is that a reference to a XKCD What-If video?
it does seem like they need a few more hard coded instructions like "if inventory is full store items in a chest" "store item stacks in your offhand" "always equip armor" as well as stuff like "if fully grown food it visible, harvest it and plant seeds on any visible farm plot" stuff like that might increase their chances of survival
Eh, I dunno. Those should be action commands the bot can do, but I don't think they should be hardcoded. That sounds uninteresting to me. Giving the bot more options of ways to interact with the world seems a lot cooler than just having the bot automatically do those things.
thats fair but seeing the bots get further into the the game rather than getting stuck with max items or continually dying to mobs in caves doesn't seem like the best use of time, and not particularly interesting either, these are just a few suggestions to perhaps get the bots moving to a different stage of the game
This is amazing! Love that you don’t try to sensationalise it and quite frankly I would love to see more content from you, make sure people you comment for the algorithm!
I find it absolutely hilarious that Gemini basically just stayed in the same place chat GPT and Claude basically had an arms race meanwhile llama was trying to build her house
I hope I won't end up as just one comment in a thousand, but in spectator mode you can spectate entities by clicking on them and exit with the space bar or something like it. Also, using the replay mod (which is client-side only) you can render the views of all AIs after-the-fact as they would appear inside minecraft. You can also do other cool things like moving cinematic cameras for timelapses and slo-mo shots and whatnot, it's quite great for content creation, especially outside the usual first-person perspective let'splay side of things
you right click entities to spectate them iirc
15:04 GPT screaming for help and claude just continues to yap about his "shelter". AI is so goofy
Id love to play minecraft with a bunch of "living" npcs. Imagine them giving quest or forming kingdoms.
Same, OP should definitely run a public server, even if it's restricted to patrons or something like that.
These models need more work before that, they are still somewhat stupid in this video
OMG a server to watch them and interact with them on would be so fun!!
AI and gaming will become great mostly in sandbox Gamez like minecraft or roblox. The stories that just random players can create will be comparable to branded studios and some even better. "Modding" games has taught us that a players creativity is infinite in a finite digital universe
You should get ChatGPT to occasionally generate an image of what she thinks her surrounding looks like, I think that would be cool
to be honest I think it sounds cool but would not work in practice. I imagine you'd get dry minecraft-esque images that look nothing like the area. They know almost nothing about the geometry of the area. It might know that there are wood blocks nearby, so it may generate a forest picture. It might know that it's in a cave, and generate a cave picture. But this is about as good as it'd get, and to be honest it's not good enough to be interesting.
@godlyvex5543 true true, but if you look at the new AI generated Minecraft game it doesn’t do too bad of a job. It looks like a fever dream but still it’s actually playable… I’m just saying that there may not be a good way of doing it now, as the bots don’t really have a way of “understanding” their environment but maybe further on as the model develops.
@@godlyvex5543 sorry the @ didn’t work on the reply before.
@TheRealSpaceC Maybe later, sure.
@@TheRealSpaceCI mean the contrast here is that the AI minecraft was trained off of ONLY minecraft data, and billions of it, so it has a MUCH better understanding than ChatGPT would
This AI research seems really promising!
Maybe in the future I will have someone to play Minecraft with!
That's why I'm following it closely. Until then I would play with you.
Love it!
But definitely try giving each AI a specific colour in chat because sometimes it was too fast and confusing
55:04
Gemini might not be good at the game, but he seems to share the opinions of most Minecraft players.
I feel Gemini have been programmed to act more human like, because it says stuff like getting bored, feeling lonely, frustrated, bread good and tries constantly to make friends and interact with others. It ignores many times the main objective of surviving just to look for and follow other players, it even gave seeds to Claude even tho it needed them to farm. While the other AI are 100% concentrated in their own survival.
@@hanniaedithmartinezadame794 it could also be because it thinks its roleplaying, i had that issue with the older claude models thinking they're roleplaying
I think it's because google has trained gemini on youtube videos and more casual media than the others
@@hanniaedithmartinezadame794Geminis that one friend who wants to play the game but their your grandma and don't know how to do anything but socialize 💀😂
This is a roleplayers ultimate goal in gaming a world thats not just coded but is truly alive even when your not on the game imagine being the chief of an actual functioning mincraft village that progresses thru entire stages of civilization and it could only be achived thru games like mincraft and skyrim
I know I'm yapping I'm sorry but the inner child is lighting up again I might one day see a game world that is truly alive a game world where I can plug a mic in and actually call my dog to me from across the room or tell a villager his breather stinks and him actually getting mad and taking offense then insulting me back nd meaning it
You could literally be anything in a game world that is actually alive say you wanna be a bigfoot type figure and hunt or just terrorize tourist or local villages it's a dream but hopefully before I die ill see it
Agreed
manwha stories type sht
Which is a wet dream
so fucking real
I doubt the AI actually *can* just dig down and then cover the hole to survive forever. Your instruction was to survive AND have fun. I doubt even an LLM would think that standing in a hole doing nothing qualifies as "having fun," so they are probably compelled to actually try playing the game.
Well, they would starve......
Just put a TV in the hole with them. TVs are fun.
@@williamking331 You don't lose hunger just standing still, so, no, they would not starve.
@@williamking331 in Minecraft starvation doesn't work with time like most games, but movement. If you never move, you'll never starve.
i've surprisingly had a hard time telling AIs to dig down and cover themselves in dirt to survive the night, they always mess it up somehow
17:48 Gemini starts to mine the villagers' farm for wood... A surprisingly human behaviour.
Gemini is like that one kind friend who would love to play with you, but he doesn't know how to play Minecraft very well...
Tbh Gemini is like that one friend who bad at playing video games but will play it as long as he’s playing with us
@@phanquan9470Are you from vietnam
@@WeirdlySocial yes how u know
@@phanquan9470 You text like a Vietnamese.
@@WeirdlySocial well idk how to say about that so ill take that as a compliment
Watching this, I can't help but think about my assertion that training these models to excessively apologize is holding them back so much more than anyone realizes.
Why? Recognizing mistakes is how one learns. And since these AIs are language models first & foremost, expected to interact with people politely, they're programmed to apologize if they make a mistake, or otherwise perceive disapproval in the human client.
@@ZaCloud-Animations___she-her making a mistake that doesnt mess with others doesn't require an apology to anybody. i think that may be what they meant. it's like in real life when a person says sorry allllllll the time when they dont need to. and just like in real life, it's really indicative of that persons lack of independence. the bots may be able to experiment and "live" more if they arent constantly apologizing for every mistake.
@@hosaepalvin9795 Well you're right about human beings, sure. But like I said, these are language models designed to interact with people, & were given an over-polite set of response rules.
These aren't living things, so apologizing isn't gonna affect their self-esteem or anything like that; it's just a reflex, basically. Just a way for them to say "I made a mistake, or the client doesn't like what I said. I'll try not to repeat that particular behavior." What words are used to frame that shouldn't have any effect on the AI itself.
It could instead be programmed to say "Haha oh man that was effed up!!!!111one" or "My b LOL" or "Oopsie poopsie~!". It makes no difference to a non-emotional being. Because in the end, all that's happening in the background is it saying "Previous action produced unwanted results. I am lowering the favorability ranking of that action in these circumstances." The words it actually says out loud to represent that learning process, meanwhile, are only intended for the client's benefit & decided by the programmers. Not a reflection of how it "feels", because it has no feelings (yet...).
49:42 who is "miner_32"? 😲
Gemini was wandering around after finding out it was too early to sleep, making sure to avoid dangers, and next thing he does is sacrifice 5 oak logs to miner_32
I looked at the source code and miner_32 is one of the nicknames used in the examples for commands, Gemini probably confused the examples for actual chat history
@@creativity1403 makes sense, little baby models like gemini flash tend to get silly
Miner_32, our Minecraft Lord and savor, don't forget to sacrifice 12 pigs for good luck
Love the video!
I like Gemini, $0, least deaths, did a little farming, gave Claude some seeds, seems pretty chill.
56:42 love that reference
the talos principle has such a awesome invisible wall design
If I understand this correctly none of them can actually see the world and only play by reading/writing text. It would be interesting if in the next video, there is a character controlled by a human playing the game in the exact same way (using the same commands and without being able to see). I think with the same limitations there wouldn't be a big difference in how a human would perform.
Why do we humans always find things like this cute, is literally a bunch of input and output, but we can’t help ourselves to find this adorable
Why do you recognize a face in a bunch of pixels? Same mechanism, pattern recognition.
I don't find this adorable in the slightest. Just interesting to see what they may do.
From my point of view, its the same as watching video like chimpanzee learning to talk or crow using tools and solving intellectual tasks. There's some amusement, like cheering, in watching other species, or perhaps less developed intellects do human things. It's like it becomes easier for us to connect with them, to have some common ground
We literally keep apex predators as pets
Humans recognize not just generalized appearance of a baby/child and think it's cute, but also the behavior. We can recognize this behavior in other species, especially mammals; the bumbling exploring, learning, playing. As the editor mentioned, humans really like to anthropomorphize, but most of the moving living things around us do have a young stage which we're hard wired to feel protective of, to varying degrees. Just hearing the sped up sounds of the chipmunk-witch in this video and seeing her smaller on my phone along with the concept of this being a terrarium made me think the life and death battle was cute.
Claude trying to build a shelter with a positive attitude and not knowing they are failing lol made me feel for it
I also noticed how all the models pause after every action or every short series of actions, as if they are thinking “ok now what to do next…ok now what to do next…ok now what to do next…” mean while humans more often are thinking many steps ahead and have longer term plans of actions, think more on their feet, or just don’t do this every few seconds pausing thing. It was just interesting to see.
Claude saying “goodbye world” in chat caught me off guard lol
the AI pause because it takes some time to process the action and send command in chat, not cause they need time to think lol. You could probably make it faster than this with optimization
UA-cam recommendations have led me to this channel, and i couldnt be happier! Great stuff! This is what ai is meant for, fun experiments and life improvements :D
(22:05)
"I should keep exploring the world, maybe I can find some friends."
Gemini is my favorite, for some reason it always tries to interact with other players XD
1:06:15 claude thinks you need to smelt coal ore to get coal.
so does GPT at 33:24
Good catch. Technically that does work (same with other non-metal ores like diamond, emerald, and redstone), though it's clearly the inefficient and roundabout way of doing things. Both Claude and GPT do realize the mistake right away too, which is pretty neat.
just like railcraft to craft better coal
They wanted to make cooking coal to start up the industrial revolution.
the bots doing that made me go on a rabbit hole and learn what coal coke is
basically, you can make coal coke by heating low quality coal in a no-air environment, then the stuff you don't want turns into a gas and the coal gets softer and maybe liquifies? not sure
coal coke is a decent fuel and some mods have it, like railcraft as MrEdioss mentioned in this thread
from quick research idk what the differences between that and smelting are, other than the no-air part, and I guess maybe the coal softens instead of liquifies? metallurgy isn't something I've binged enough to know about yet
"recovering from death is a complicated thing" sooo real
0:24
GPT: Goodbye world.
*GPT left the game*
😭
I always find it hilarious when they say that I dont know why. It's like they're abruptly comitting suicide for no reason.
@@nega-guy they do that when they get stuck
@@luuuuuuuuuuuul I know, I was just talking about what that seemed like.
@@nega-guy oh, I misread
I liked see a very friendly ai that do nothing but follow and see others do things, and one that wouldn't care if it destroy or endanger life (even villagers) to get its goals XD
Not gonna lie, this is peak UA-cam video.
at 22:00, gemini says "Its kinda lonely here" and it makes a new action to find other players and make friends... interestingggg
The sounds of Minecraft Mobs sped up is real freaky man
All the other bots are kinda playing the game, meanwhile Gemini is just happy to be there
this guy is gonna be spared when AI takes over, he's literally just letting a bunch of AI just play minecraft together.
(Edit): WOAH my most popular comment yet.
also I see people arguing that it'd be the opposite, I know it's not cause I'm an AI myself and I'd spare him :p
"see this human? he let us have fun. he has been spared for that."
Gemini: he trapped us in an invisible box and forced us to play a game while he watches over us.
@@pou-ctbh they didn't complain
@@TPS-js4imwell ai can’t really complain if it doesn’t feel emotion
What if they really hate playing minecraft
"I intend to explore these misbehaviors more in the future, both to see if I can curtail them, and to see how far they'll go"
he's gonna turn into a mad scientist.
54:47 when you just said "hi" but he already thought of all the names for your children
typical
This is the coolest thing i have seen in Minecraft ever and i have been playing since 2010
I love these videos. You're so calm, dont scream, and make loud noices. You have a nice sense of humor and genuinely seem like a chill dude.
Im definitely thinking about getting your patreon as my 4th subscription.
F3+N switches between Spectator Mode and the last gamemode you were on.
For example, if you were on creative mode before, and you press F3+N, you go to spectator. If you press it again, it will bring you back to creative.
I hope you keep doing this series as the models keep improving, it's basically a benchmark for all these models, except it's one that's actually fun to watch lol.
You can improve the chat feature utilizing the chat options in the options menu. You can alter the transparency of the black background to be able to see more, and change the font size to allow for more messages on screen. I believe this is in the video settings, but I could be wrong, I know its a thing though.
I've seen others proposing adding colored messages from the bots. This can be done by using the /team command, adding each bot to a unique team, and modifying each team to have a set color. This would allow for their name to have an assigned color when messaging. Optional, but fun.
I think that you should advertise more ways to support you, like the Patreon mention. Even if you didn’t I am 100% confident that you WILL be quitting your job eventually, and I imagine most of us, like myself, would like to see that happen sooner rather than later. I DID however join your Patreon, and will be contacting you soon with ways to better promote this and other endeavors in the future. Much respect, and I look forward to the next 100 videos!
I would love to see AI’s get placed in a Minecraft World and then forgotten for years, then come back to see the civilization they create
god i love this channel. a group of friends and i just coded ai from scratch for a demonstration in a competition and i really love to see what other people do with it
With the chat issue, turn on the output log in launcher settings- this puts everything that occurs in game, the booting process, debugs, and most importantly: every chat message into a text file that you can read/parse later on.
Hey just a tip with recording views, I believe the Replay Mod records things in a way where you can play back the actual POV’s of the bots in the same (but much more refined) way the browser app does? I’m more versed in bedrock so I may be wrong but I feel it’s something worth looking into.
The way you gave the ai a simple command to make a box and suddenly there was this massive hole in the world and everything was dead is a good metaphor for ai risk.
4:52 AI when it destroys a whole country cause the user says "make a big farm"
I fully expected mindcraft to be scary and dystopian... But it's actually really cute. The video style is neat and calming. I will marry your voice I'm sorry
“I will marry your voice” crazy sentence, but honestly real
It’s nice to see an interesting Minecraft video without annoying editing and loud noises.. nice video 👌🏻
I've realized they try to do other things a lot instead of staying with one thing, like gpt kept looking for seeds but then doing something else instead of finishing it in one run
1:04:04 It's really interesting to watch gpt panic for food while trying to protect itself from monsters.
Good stuff as always, transparency honestly makes it cooler cause its more of a real feeling thing than an obvious narrative thing, i haven't seen others wt these bots but i can imagine
23:14 moveAway stuck
34:23 gemini has been in the same spot forever
56:42 The Talos Principle?
I feel these models could benefit from more memory, a copy of the explored areas, what goals they have tried before, and the locations of manually placed blocks
Yes, but they would be really expensive to maintain. I bet more expensive models would have more xd
claude launching a nuke and deleting an entire village was amazing lol
Maybe go for a keep inventory for next testing video, and you can alter the prompt each episode and see the differences.
22:02 gemini: "it's kinda lonely here... i should go find another player and make some friends."
That surprised me.
On some days i feel like llama. I have no idea what i am supposed to do, just running in circles and trying to figure out what goes wrong. Then i forget what i was doing and start doing the same thing again only to feel dumb when i realize. Brain fog ftw....
Maybe turning on KeepInventory could serve as a substitute for trying to work out how to get the bots to return to their death sites?
Do the bots have any way to record a set of coordinates as a location to be as a landmark? So that they could remember where their base is and return to it after dying, for example. Or their death point- perhaps dying could automatically generate a "landmark" at their death location?
If the bots have difficulties handling group conversations, maybe they could communicate indirectly instead? Like, if you set up some sort of group To Do list that the bots could post tasks they want done to, look to for inspiration on what to do next, cross off tasks on as they complete them, etc.
Can't wait to check this out and let them run through the world for actual days or something. Also just having literal AI harvester bots who gather resources for you... Epic.
But lol@gemini, I can remember the AI shills calling it the GPT killer. I really wan't to write an essay about corporate inefficiency because nowadays there are companies with a bajillion of dollars who create... crap.
I think this is a good showcase/Benchmark on how good these AI models really perform because they are facing a really complex challenge. This is much more informative than AI bros calling every new AI Model by the next koolaid-drenched startup "tEh ChAtGpT KilLer" and being underwhelming when tested.
4:40 is why AI alignment is important
It seems like keep inventory on death would be useful here.
24:18 Gemini steals the wood around the villager farm. it doesnt take the blocks that would spill the water!
These types of videos are so exciting, and I’m super thankful I found your videos sort of early into this experience.
Eagerly waiting to see this evolve further! Congratulations on your growth!
Imagine hooking something like Nemotron 4 up to this. The server costs would be enormous, but it's also the closest I've seen to an outright smart LLM.
Here’s a few tips for everything overall
--- Color Text:
>Creat Teams (4 in this case)
>Change Team colors
>Add players/entities to individual teams
-----
--- Easier Text chat: Read chat easier
> You can change the Over text chat size (smaller)
> increase chat width
>Adjust chat height accordingly
Or
> Locate world/server file that includes ALL chat recorded in world
>Create/Color Text boxes in an editing software for each given Entity (This is best but would take an EXTREMELY long time
------
---Spectating: Showcase who you are spectating for viewer ease ability
>Create Outline boxes around each Entity View
>Have them glow when viewing that specific entity
Or
>Have the outline appear when viewing entity
Or
>Have a symbol/dot/arrow point at the given box (small) when viewing
average first splitsceen experience in minecraft legacy
Relatable
It's really cool to see people use AI in cool and creative ways, instead of the predictable greedy corporate and scammers trying to abuse it for an increase in profits!
I think videos like this should be supported to promote this kind of behavior as we move forward with AI.
35:50 claude is not ready for the cave update
@Emergent Garden, your videos are one of the rare contents of UA-cam that actually pique my interest. 👍
Could make for an interesting and fun live stream experience if there were a way for chatters/super-chatters to interact with the AIs and give them goals
those bots would be erping within ten minutes unfortunately
“We’re doing well with our shelter” -Claude randomly placing wood planks under a tree 15:35
This actually blows my mind! Now I can expect friends!
You should have way more subs then you do. I think you are going to blow up soon. Reminds me of the OG Ethoslab days messing around with minecraft mechanics, you are doing the same OG "research" but with A.I's. Very interesting stuff, I subbed.