More information about how Albert was trained: Time it took to train: Room 1: 0h 12m 35s Room 2: 0h 15m 32s Room 3: 0h 57m 10s Room 4: 1h 06m 40s Room 5: 1h 47m 14s Room 6: 5h 56m 35s (I overtrained it a bit to get more consistent results) Total Training Time: 10 hours, 15 minutes, 46 seconds (plus 3 weeks of trial and error for the last room) This is a very long comment going over more of the details of how Albert works and issues he currently has. I've tried to make it as easy as possible to understand, but some parts are complicated. Either way, after the last video there were many people wanting more information, so here you go!:D If you're interested in training your own AI like Albert but don't know how, there's now a really easy way to do it! Luda, an AI lab, recently built a web app that allows you to create and train your own AI using deep reinforcement learning (just like Albert) completely for free in your browser! You build your own character (called a Mel) with lego-like building blocks then watch it train in real-time on their website in just a few minutes (really). It's an awesome project, and just like my videos, makes deep reinforcement learning so much more accessible, which is why I love it so much. This section of the comment is sponsored by Luda, but these words are entirely my own, it's an amazing project that I would have been obsessed with had they released it before I built Albert. I've genuinely been looking for a sandbox/game exactly like this since I was a kid. They're still early, but they're giving my audience first access to their closed, pre-alpha build. Make sure you check out their site and create an AI agent for yourself!:D prealpha.mels.ai Now, back to Albert: NOTE: You only see one Albert in the video, but there are actually around 50-100 copies of Albert and the room he’s in training simultaneously behind the camera to significantly speed up the training time (and the time it takes to go through all the footage to edit the video). THE BASICS: Albert was trained using reinforcement learning, meaning he was rewarded for doing things correctly and punished for doing them incorrectly (the reward is just increasing Alberts score, and the punishment is decreasing it). After Albert finishes each attempt, the actions he took are analyzed and the weights in the neural network (Albert's brain) are adjusted using PPO (proximal policy optimization) to try to prioritize the actions that lead to a positive outcome, and to try to avoid the actions that lead to a negative outcome, using the sum of the rewards and punishments as the evaluation of the outcome. Albert starts off making essentially random decisions until he accidentally hits the pressure plate in the first room and is rewarded, then, as mentioned above, the weights in his neural network brain are adjusted in order to try to replicate that reward. Or, Albert does something to be punished (like hitting the obstacle), in which case the weights in his neural network are adjusted to try to avoid that. This process continues and eventually results in him being able to walk towards the pressure plates and out of the room (into an invisible pressure plate behind the door), where he sees the second room and continues his learning in the new environment. I had to use the invisible pressure plate at the front of the next room to get him to go through the door because once he makes it though once, he’s in the next room permanently, so he can’t learn to go through the door himself. REWARD FUNCTION: The actual rewards and punishments were given as follows: falling off the platform (-0.5), hitting an obstacle or wall (-0.1), hitting the ground (-0.05) to try to minimize excessive jumping, hitting a pressure plate (+1, +0.9, +0.8, depending on how quickly it’s hit), escaping the room (+1). ALBERT’S BRAIN: Alberts brain is a neural network with a total of 5 layers (one input layer, 3 hidden layers and one output layer). The network is a Multi-Layer Perceptron (MLP) with 897 nodes, 510 input nodes, 128x3 hidden nodes and 3 output nodes. When his brain changed in room 6, the 510 input nodes became 1230, making the total number of nodes 1617. In the first brain, there were a total of 510 inputs, all from raycasts. There were 3 raycasts looking down, 7 at eye level and 7 above Albert’s head, all with an FOV of 70 to try to mimic our own eyesight. Each of these 17 raycasts could detect the 4 types of objects (ground, wall, obstacle, pressure plate) and the distance to the object it hits. This leads to a total of 6 observations per raycast (one observation is whether or not an object is there, one for the distance to the object and one for each of the 4 detectable tags). The raycast observations are then stacked 5 times, allowing Albert to remember the previous 5 observations. Each observation is observed and acted on every 10 academy steps, there are 50 academy steps per second, meaning Albert has a short term memory of exactly 1s. The 10 academy step delay in actions also makes Albert’s movement look a lot smoother than most AI you’ll find on TH-cam, this makes for a slightly less accurate AI, but a much better viewing experience. There were 17 raycasts each responsible for 6 inputs, and they’re each stacked 5 times, so there are 17*6*5=510 total observations in the input layer in the first brain. In the last room Albert needed more observations to be able to complete it, so I increased the number of raycasts. He now has 17 raycasts at eye level, 17 above his head, and 7 looking down, changing the number of observations to 41*6*5=1230. These new observations would have allowed him to beat the previous rooms more accurately, but the need for them only arose in the last room. I’ve given Albert 3 hidden layers each with 128 nodes because the last video had 2 hidden layers and I figured the moving obstacle adds another layer of complexity that the neural network should account for. 128 nodes per hidden layer was chosen fairly arbitrarily, 128 is just the default number of nodes per hidden layer in ML-Agents. There are a total of 3 outputs, one to determine Albert’s forward/backwards movement (go forward, backward or do nothing), one to determine his right/left turning (turn left, right or don’t turn) and one to determine if he jumps or not. Having 3 outputs allows him to perform all 3 actions at the same time. ISSUES: There are a few issues with Albert’s brain for this task, for starters, I greatly underestimated how many observations are needed to accurately avoid these moving obstacles. This was fixed in the last room by more than doubling the number of raycasts, but I could have improved the AI even more by also giving him his coordinates in the room, so he can more accurately understand which positions in the room are dangerous and get out of them quickly. If you're still reading this, you're probably really smart and want to learn more about Albert, so make sure to join my discord server I just made where we can talk more about the details of Albert's AI! discord.gg/jM2WkNuBnG :) There also was an issue with the vertical spinner in Room 5, there were too few raycasts for Albert to consistently see that spinner (sometimes the raycasts see to the left and right of the spinner, but not the spinner itself, resulting in Albert being blind to it. This caused Albert to try to jump through the spinner regardless of its position (unless the vertical spinner was directly in front of him), so he couldn’t do it very consistently. This is no longer an issue after Albert’s brain upgrade in the last room, but it was during the training of Room 5. Overfitting was also an issue with this training. Overfitting generally isn’t a big concern with reinforcement learning tasks because the training data is exactly the same as the testing data so it’s guaranteed to overfit to some degree, but the issue arises when Albert overfits too much to one room, beats it, then starts the next room. Albert most noticeably overfit to Room 4, making it take a while for him to figure things out in Room 5. This can mostly be fixed by randomizing the locations of the pressure plates, platforms and obstacles in the copies behind the camera, but that would require a lot of strategic limits on the randomization of the positions to make sure it’s always possible, which isn’t ideal. I think a better way to address overfitting for this task, and one that I started implementing in Room 5, is to make very slight and random movements to everything in the copies of the rooms, as well as making copies that don’t have obstacles, and copies that are mirrored. This should get the best of both worlds, where Albert is less likely to overfit, without the need for strategic changes to the rooms; the changes can be made automatically. Thank you so much for watching! These short videos take literally hundreds of hours to make, if you want to help allow us to make them faster, please consider becoming a channel member! By becoming a member, your name can be in future videos, you can see behind-the-scenes things that don’t fit in the regular videos, you can also use stickers of Albert, Kai and some other characters our team made in comments (more coming) :D
I still think there should be a test of him learning to not hit a button or else the timer will restart and not the test. Or something to that degree. Like he learns buttons are good but not every button should be pressed
It looks cool, but I also like to think that this way AI scouts area around him without losing any time as Albert can only see forward like humans do but not all the ways at once.
I don’t know what gets me more: the fact that Albert will randomly ragequit and jump off the map, or the little celebratory acrobatics he’ll sometimes do when he gets a part right 😂
While very adorable, his jumps and 360 rotations also serve a very clever purpose. Since he can't really see high objects or see around him, when he jumps and does a 360, he memorizes his environment and works from there. Sometimes Albert just.... jumps off or does weird actions, which I don't really get but hey, he's a deep reenforcement learning AI, he's in his own world.
Albert doesn’t do it because of that. You can tell that he can’t rennet his surroundings he only remembers how he did int he previous rounds. He does that because the ai is coded to always move(it seems unless he gets sideways or upside down) he wouldn’t really be able to learn that either was though.
He just jumps off bc he is aware that the he couldn’t do it in the time allotted I’ve noticed when it’s a crap run he’ll just jump when there’s no way he could come back from it
@@ishas4421the reason is he probably gets punished the more he gets hit by the spinners so hed rather just off himself then get hurt again since I don’t think he gets punished from falling
i never realized that the spinjump is actually practical for information, i thought that they saw based on the video's perspective until i realized their sight is _actually from their eyes_
I think a big issue Albert had regarding his AI is that he has zero object permanence. The moment something leaves his cone of vision he doesn't seem to acknowledge it's existence anymore, and thus he ends up getting blindsided by spinners at times or fails to locate the door when he clearly saw it.
@@TheSinzy if it was not learning it would not get any better at escaping. It is changing the weights of some things in the neural net so its not like its an unchanging randomization machine like you say
The problem with a higher intelligence is that it will fool us into a false sense of safety, like when you were two and your parents would fool you about everything. if GAI comes into being we are literally doomed, there will be no tomorrows for humanity.
The last scene suddenly gave a sort of a horror movie vibe where the main character comes to realise that his entire life was dictated and he never was able to achieve freedom, with his sole purpose being an entertainment tool imprisoned forever...
i think it was really interesting watching Al’s decision to do a 360 every time he jumped. he probably did that so he could better see where everything is. i noticed him do one 360 in place, and then he beelined for a pressure plate in a new room.
It’s also optimal because spinning horizontally keeps him more stable vertically thanks to the physics. In real life if you throw a cup or something upwards it wobbles and falls pretty randomly. But if you spin it and throw it up it will stay straight thanks to its angular momentum.
i like to theorize that whoever wrote "the cake is a lie" in room 6 is actually other versions of albert. basically, this albert is not the first AI, but only one of the many AIs the AI Warehouse has created and trained. they'd then be used for very dark, secret, malicious and malevolent intents. i also like to imagine that AI Warehouse is also an AI with its own will and consciousness that has somehow founded this facility to create and train AIs at a large scale
It's like 4 o clock on the morning but to me there's something absolutely beautiful about this all. Like we as humans just have such an innate need to give everything a personality. All the comments cheering Albert on, calling the spin he does to check his surroundings "celebratory acrobatics", the commentary throughout the video, even down to the physics and the way Albert moves gives him this loveable aloofness to him, making you cheer him on more as he tries and fails and does his sick tricks. And christ, he has googly eyes. It's just. Such a wonderful thing how much personality we give an AI cube.
The love to the details is sooooo good. The coloring of different things makes this so relatable. Whenever there is "Albert" in orange, I see a half meanless half dumb face which doesnt understand the problem. Green is when something is good and Red pure frustration. And the music stopping (followed by red text).. Love it! :D
Yes! more Albert! love these videos. so interesting from a computer perspective but also just as much from the "Albert is adorable and I love watching him try and eventually succeed on these challenges with the text commentating" the *sigh* at the door was perfect. everything was perfect :)
I think it could be more perfect if Albert got himself some kind of object permanence? Kind of unusual to see him spin around on each jump. Is it because the AI has figured out this is a good way to see things? Or is that itself proof it has memory? EDIT: Got time to (skim) read the pinned comment. I think I'm pretty wrong now, but I'm still curious about why he spins so much. Hope I didn't miss that explanation.
I love "learning AI" content on TH-cam so much, but there is such a lack of little snippets of editing that give it character. It makes content like this so perfect and less for "educational purposes only", which are so cold occasionally. Please, keep making more because I know there are many more people like myself out there that crave this stuff. Amazing work and I can't wait to see more! 😁
My favorite part is 7:20 in room 3, where he lands perfectly, but he can't figure out how to get through the door, so he just falls over from the spinner
@@aeterborne or just the idea of pressing multiple buttons simultaneously to progress or needing to jump on one another even tho the last is probably a little bit too hardcore to do
I love how at certain points, like 3:59, Albert just decides to do a sick 360 and hit the button in style. He may be a learning AI, but he still has style!
What really makes me impressed is that Albert is kind of like us, as he makes mistakes and doesn’t understand anything at first, but eventually learns what to do.
I would love to see after he completes all the chambers, if he's able to do all of them back to back without a mistake. As in, when he completes the last chamber, you save file his memory, reset the coding to where any room he's in, if he fails, he starts back in the first room. Doesn't matter if he's in room 2 or room 6. I want to see if he is able to retain the patterns and methods to get through all the rooms in one run!
Do you think the reason he's constantly jumping and doing 360s is so he can see the room? That way he's able to survey more of the area around him and figure out where the pressure plates are etc.
Has to with maintaining stability as well. Helps him remain upright, sort of like a frisbee, so that he can minimize the chance of tumbling around when he lands
This honestly reminds me of a player who has absolutely zero idea what is happening and is just trying to learn by failing. Like, ramp up the difficulty by 10 and I can see myself making these same mistakes. It's honestly quite impressive how far we've come.
it's so interesting how we all get emotionally attached to an ai cube if it has eyes, a name and gets treated as a living thing... we're all here calling him cute (and i mean he is, look at the lil celebration bounces when he gets something!)
Dude, these series are so fun to watch. There's nothing like this on yt, sure there's the rare chance that you find 1 or 2 vids but this is magic, gold, true comedy and an actual series, not just 1 vid. Keep it up!
If this was turned into a game, with the premise that you're an advanced AI learning to do things with an AI narrator that either communicates with text popups or using some TTS (or maybe both), I would instantly buy that game the moment I could.
I wish we could see a post-learning run after he finishes the gauntlet to see if he actually retains the knowledge and can do it in a reasonable amount of time
I’m at a point halfway between laughing and crying because I see myself so much in this little ai cube dude, as if all my attempts at learning how to live life as a human are being monitored and judged by some alternate higher version of myself
I personally am inspired by Albert's long journey. I was attracted to your most recent post where Albert fearlessly faces off against a secondary AI, and was inspired to watch more! Though there are not many videos you've posted, I like to see Albert coming up with new and interesting ways to complete different tasks! (Also, i saw 'the cake is a lie' and it was so hard to not laugh XD so in short, you're awesome, keep on going!
I just love these videos! You pour so much time into training Albert and editing these videos, plus you gave him an upgrade during it, which I can’t even fathom the effort behind. Amazing job, I really hope we see more of you and him!
Seeing him jumping on the spot and spinning in the air was really quite fascinating. For all his dumbassery, he got the hang of scoping out his surroundings real quick
Man, that's so sick. I laughed all the time. No, I really worried about Albert but I was waiting for that Portal's easter egg all the time. Your sense of humor is great and all skills that allow you to make such a masterpiece. Thank you very much and we'll be waiting for Albert's next journey. Wish you both a good luck!
This is tremendous content. I can tell you spent tons of time on the video editing, not to mention the whole process. I like the added explanations and timer on this video. If I could add one thing to the next one, I’d like an attempt counter for each room.
It'd be cool to see an extra screen under the timer that tells you what iteration is currently being shown! I'd love to see how many attempts he takes to complete a room.
Machine learning is something I'm really interested in learning about tbh. Also this was quite fun to watch! I also noticed the portal reference in one of the levels which is cool!
Completely random fact you didn't need to know: At 3:44 the bassoon melody of the background music extends to the highest note in it's range, an E4. Great video as always btw.
Goddamn man this is your second vid and you're popping off so hard already! And no doubt you deserve it. your videos are both informative and hilarious!Really happy for you man keep it up!!
Your editing makes this even more enjoyable to watch. Like the decision you made to have 10tick movements so it’s more appealing to view even though it makes for a less accurate AI. And the perfect moments when to slow down the recording. Or even just the comments in the video.
These videos are very entertaining to watch. There is so much potential for things you can give Albert like a grapling or a double jump or a dash. I imagine it would be very entertaining to watch him figure these new mechanics out. Great job
I'd love to see the ai go back to earlier stages to see if it still knows how to solve it. Or have it solve rooms with the same concept but different layout to see if it's overtrained or if it knows the assignment
Albert! Our beloved spinning block is back again! Another excellent video - I already mentioned how the visual presentation makes it fun to engage with, and Albert's cute & simple design and quirks make it very endearing. Even just him having blinking eyes adds so much character. I'm not sure how feasible it is, or how difficult it'd be to implement if possible, but an "Albert cam" would be cute - like, a small thumbnail view in an empty corner that shows off what Albert is "seeing".
Love the cute graphics and story telling. That is very refreshing. AI is a fun topic but most videos are somewhat technical. Very good video. Enjoyed it!
I wonder what would happen if you put all of the training from the other videos and applied it to something new, would it overcome it faster or slower? Awesome video!!
He keeps the same brain throughout the video (except for Room 6), but between videos his brain is wiped clean. It can sometimes look like he has a new brain when he enters new rooms, that happens when he overfits to the previous room too much, that makes him perform very poorly with anything even slightly different from the room he just left. A series that would be really interesting though is keeping the same brain and just constantly training on top of it, I might do that in the future:)
i think it would also be cool to do runs through the whole system, so he has to train for everything. though im sure that would increase the simulation time exponentially
It would be interesting if after they manage to consistently clear a room you add a reward for a faster time then let it run for awhile longer to see how fast they can clear each room.
I'm curious what it'd look like if you looped him through the rooms for another 10 or so hours. How many runs could he complete, how clean would the runs look, etc.
From experience, these AIs are used to show deep reinforcement theorems and papers on specific problems (levels). The cube in our case keeps seeing different types of levels, it might ace the first level or mess it up cos it's calibrated to the difficult ones. As a different type of example, you can take image recognition AI trained with reinforcement. You train it to detect apples by their colour and shape,but if you bring one of those yellow apples it might not work cos it doesn't apply to previous scenarios
What's interesting is, through the eyes, the thought you gave him, the name, the way he looks conscious it's really weird for my brain to process Albert, but you gotta love those spinning jump he doesn't only nail an obstacle but does it with style
More information about how Albert was trained:
Time it took to train:
Room 1: 0h 12m 35s
Room 2: 0h 15m 32s
Room 3: 0h 57m 10s
Room 4: 1h 06m 40s
Room 5: 1h 47m 14s
Room 6: 5h 56m 35s (I overtrained it a bit to get more consistent results)
Total Training Time: 10 hours, 15 minutes, 46 seconds (plus 3 weeks of trial and error for the last room)
This is a very long comment going over more of the details of how Albert works and issues he currently has. I've tried to make it as easy as possible to understand, but some parts are complicated. Either way, after the last video there were many people wanting more information, so here you go!:D
If you're interested in training your own AI like Albert but don't know how, there's now a really easy way to do it! Luda, an AI lab, recently built a web app that allows you to create and train your own AI using deep reinforcement learning (just like Albert) completely for free in your browser! You build your own character (called a Mel) with lego-like building blocks then watch it train in real-time on their website in just a few minutes (really). It's an awesome project, and just like my videos, makes deep reinforcement learning so much more accessible, which is why I love it so much. This section of the comment is sponsored by Luda, but these words are entirely my own, it's an amazing project that I would have been obsessed with had they released it before I built Albert. I've genuinely been looking for a sandbox/game exactly like this since I was a kid. They're still early, but they're giving my audience first access to their closed, pre-alpha build. Make sure you check out their site and create an AI agent for yourself!:D prealpha.mels.ai
Now, back to Albert:
NOTE:
You only see one Albert in the video, but there are actually around 50-100 copies of Albert and the room he’s in training simultaneously behind the camera to significantly speed up the training time (and the time it takes to go through all the footage to edit the video).
THE BASICS:
Albert was trained using reinforcement learning, meaning he was rewarded for doing things correctly and punished for doing them incorrectly (the reward is just increasing Alberts score, and the punishment is decreasing it). After Albert finishes each attempt, the actions he took are analyzed and the weights in the neural network (Albert's brain) are adjusted using PPO (proximal policy optimization) to try to prioritize the actions that lead to a positive outcome, and to try to avoid the actions that lead to a negative outcome, using the sum of the rewards and punishments as the evaluation of the outcome. Albert starts off making essentially random decisions until he accidentally hits the pressure plate in the first room and is rewarded, then, as mentioned above, the weights in his neural network brain are adjusted in order to try to replicate that reward. Or, Albert does something to be punished (like hitting the obstacle), in which case the weights in his neural network are adjusted to try to avoid that. This process continues and eventually results in him being able to walk towards the pressure plates and out of the room (into an invisible pressure plate behind the door), where he sees the second room and continues his learning in the new environment. I had to use the invisible pressure plate at the front of the next room to get him to go through the door because once he makes it though once, he’s in the next room permanently, so he can’t learn to go through the door himself.
REWARD FUNCTION:
The actual rewards and punishments were given as follows: falling off the platform (-0.5), hitting an obstacle or wall (-0.1), hitting the ground (-0.05) to try to minimize excessive jumping, hitting a pressure plate (+1, +0.9, +0.8, depending on how quickly it’s hit), escaping the room (+1).
ALBERT’S BRAIN:
Alberts brain is a neural network with a total of 5 layers (one input layer, 3 hidden layers and one output layer). The network is a Multi-Layer Perceptron (MLP) with 897 nodes, 510 input nodes, 128x3 hidden nodes and 3 output nodes. When his brain changed in room 6, the 510 input nodes became 1230, making the total number of nodes 1617.
In the first brain, there were a total of 510 inputs, all from raycasts. There were 3 raycasts looking down, 7 at eye level and 7 above Albert’s head, all with an FOV of 70 to try to mimic our own eyesight. Each of these 17 raycasts could detect the 4 types of objects (ground, wall, obstacle, pressure plate) and the distance to the object it hits. This leads to a total of 6 observations per raycast (one observation is whether or not an object is there, one for the distance to the object and one for each of the 4 detectable tags). The raycast observations are then stacked 5 times, allowing Albert to remember the previous 5 observations. Each observation is observed and acted on every 10 academy steps, there are 50 academy steps per second, meaning Albert has a short term memory of exactly 1s. The 10 academy step delay in actions also makes Albert’s movement look a lot smoother than most AI you’ll find on TH-cam, this makes for a slightly less accurate AI, but a much better viewing experience. There were 17 raycasts each responsible for 6 inputs, and they’re each stacked 5 times, so there are 17*6*5=510 total observations in the input layer in the first brain.
In the last room Albert needed more observations to be able to complete it, so I increased the number of raycasts. He now has 17 raycasts at eye level, 17 above his head, and 7 looking down, changing the number of observations to 41*6*5=1230. These new observations would have allowed him to beat the previous rooms more accurately, but the need for them only arose in the last room.
I’ve given Albert 3 hidden layers each with 128 nodes because the last video had 2 hidden layers and I figured the moving obstacle adds another layer of complexity that the neural network should account for. 128 nodes per hidden layer was chosen fairly arbitrarily, 128 is just the default number of nodes per hidden layer in ML-Agents.
There are a total of 3 outputs, one to determine Albert’s forward/backwards movement (go forward, backward or do nothing), one to determine his right/left turning (turn left, right or don’t turn) and one to determine if he jumps or not. Having 3 outputs allows him to perform all 3 actions at the same time.
ISSUES:
There are a few issues with Albert’s brain for this task, for starters, I greatly underestimated how many observations are needed to accurately avoid these moving obstacles. This was fixed in the last room by more than doubling the number of raycasts, but I could have improved the AI even more by also giving him his coordinates in the room, so he can more accurately understand which positions in the room are dangerous and get out of them quickly.
If you're still reading this, you're probably really smart and want to learn more about Albert, so make sure to join my discord server I just made where we can talk more about the details of Albert's AI! discord.gg/jM2WkNuBnG :)
There also was an issue with the vertical spinner in Room 5, there were too few raycasts for Albert to consistently see that spinner (sometimes the raycasts see to the left and right of the spinner, but not the spinner itself, resulting in Albert being blind to it. This caused Albert to try to jump through the spinner regardless of its position (unless the vertical spinner was directly in front of him), so he couldn’t do it very consistently. This is no longer an issue after Albert’s brain upgrade in the last room, but it was during the training of Room 5.
Overfitting was also an issue with this training. Overfitting generally isn’t a big concern with reinforcement learning tasks because the training data is exactly the same as the testing data so it’s guaranteed to overfit to some degree, but the issue arises when Albert overfits too much to one room, beats it, then starts the next room. Albert most noticeably overfit to Room 4, making it take a while for him to figure things out in Room 5. This can mostly be fixed by randomizing the locations of the pressure plates, platforms and obstacles in the copies behind the camera, but that would require a lot of strategic limits on the randomization of the positions to make sure it’s always possible, which isn’t ideal. I think a better way to address overfitting for this task, and one that I started implementing in Room 5, is to make very slight and random movements to everything in the copies of the rooms, as well as making copies that don’t have obstacles, and copies that are mirrored. This should get the best of both worlds, where Albert is less likely to overfit, without the need for strategic changes to the rooms; the changes can be made automatically.
Thank you so much for watching! These short videos take literally hundreds of hours to make, if you want to help allow us to make them faster, please consider becoming a channel member! By becoming a member, your name can be in future videos, you can see behind-the-scenes things that don’t fit in the regular videos, you can also use stickers of Albert, Kai and some other characters our team made in comments (more coming) :D
Its Cool That Ai Is Getting Smarter
Noted
I still think there should be a test of him learning to not hit a button or else the timer will restart and not the test. Or something to that degree. Like he learns buttons are good but not every button should be pressed
i dont understand how having copies of him behind the camera helps with video editing, you dont even see him
@@stonksinwestycje5004 excuse me what
I'll never stop loving Albert's sick 360 spins when he nails a jump
Little dude’s got a lotta personality for something with literally no personality~ god i love him~
@@SoraQuill same
*360s into the void*
It looks cool, but I also like to think that this way AI scouts area around him without losing any time as Albert can only see forward like humans do but not all the ways at once.
@@Mpkki In Albert’s first video I believe AI Warehouse mentioned that’s exactly why Albert does these 360s, to scope out the area.
The fact that he just gives up occasionally and just 360’s off the map is such a mood
literally me
thats dark
VALHALLA
Me
@@rufus8821 how is that dark?
The moments when Albert "ragequits" and jumps off the cliff after failing are the funniest thing ever.
When?
6:06 ?
It really speaks to me as a gamer
@FoxMaster whot
@FoxMaster sing with me, sing for the year…
I like it when Albert occasionally decides to pirouette gracefully into the abyss
Got me laughing
sweet release of the abyss
SAME
Perfectly descripted
Mf took Geralt's criticism to heart
"PIROUETTE!"
I’d love to see an attempt counter in the room so we can see how long it took for each highlighted milestone to take place
also maybe a stopwatch to see how long it takes 🤔
Big brain thinking
I agree
yes
Opened comments to suggest same ideas)
It’d be good to see attempt counter+time for each room separately and total counter+time for all rooms
I don’t know what gets me more: the fact that Albert will randomly ragequit and jump off the map, or the little celebratory acrobatics he’ll sometimes do when he gets a part right 😂
Мне кажется он понимает что уже слил этап и сам сливается
@@Nikitashow12355 do you like vodka?
@@Nikitashow12355 ты любишь водку?
@@kyuenbi bruh, i like coca cola because still can't drink alcohol. specifically tried some and I dont liked it, but not vodka
my guess is the VERY SHORT vision buffer and he "forgets" the cliff is even there
While very adorable, his jumps and 360 rotations also serve a very clever purpose. Since he can't really see high objects or see around him, when he jumps and does a 360, he memorizes his environment and works from there. Sometimes Albert just.... jumps off or does weird actions, which I don't really get but hey, he's a deep reenforcement learning AI, he's in his own world.
Albert doesn’t do it because of that. You can tell that he can’t rennet his surroundings he only remembers how he did int he previous rounds. He does that because the ai is coded to always move(it seems unless he gets sideways or upside down) he wouldn’t really be able to learn that either was though.
@@dan_schnider Creator did say that he has some short term memory, but it doesnt last very long
The concept of object permanence or whatever its called leaves his body after like 3 seconds lol
He just jumps off bc he is aware that the he couldn’t do it in the time allotted I’ve noticed when it’s a crap run he’ll just jump when there’s no way he could come back from it
@@ishas4421the reason is he probably gets punished the more he gets hit by the spinners so hed rather just off himself then get hurt again since I don’t think he gets punished from falling
the ending of albert tipping over at 14:02 on sync with the music is so perfect i love it
No way...I didn't notice that
The wall and ground crashes were also synced lol
Albert quickly learned that spin-jump technique, showing just how smart AI's really are. He got a quick scope of the area.
i never realized that the spinjump is actually practical for information, i thought that they saw based on the video's perspective until i realized their sight is _actually from their eyes_
I love how it looks like it has a "I wasnt planning on getting this far" moment everytime it gets over an obstacle and has to tackle a new part
❎
I think a big issue Albert had regarding his AI is that he has zero object permanence. The moment something leaves his cone of vision he doesn't seem to acknowledge it's existence anymore, and thus he ends up getting blindsided by spinners at times or fails to locate the door when he clearly saw it.
his issue is only 2 videos
@@yyhhttcccyyhhttccc6694 what
@@jilljohn2638 he has only 2 videos and it makes me mad
@@yyhhttcccyyhhttccc6694 hey there’s a new video
@@yyhhttcccyyhhttccc6694 agreed!!!!
7:40 bro Albert's front flip was sick. My man's a gymnast
@TheAnimatedStickman OH COME ON 🤣🤣🤣
I just realized he actually answered
He's so proud of it too 😭
I would love to see Albert try the whole level over to see how much he really learned rather than memorized
He is neither learning nor memorizing anything. He just clicks random buttons and checks if that makes him progress further to the way out.
Or to have him start in randomized positions in the set starting area so he can't follow the same path so strictly
@@TheSinzy if it was not learning it would not get any better at escaping. It is changing the weights of some things in the neural net so its not like its an unchanging randomization machine like you say
@@TheSinzy if he didn’t learn anything or memorize it then he wouldn’t have made it past the first room and yes AI can both learn and memorize things
@@g10xz._ they can't, no such thing as "AI" eitheir
“Ai will take over the world!”
AI : “I have pressed the button to open the door, I will go back in the room I came from!”
AI: *Doing sick frontflips and 360s*
AI: time to celebrate with some sick flips
The problem with a higher intelligence is that it will fool us into a false sense of safety, like when you were two and your parents would fool you about everything. if GAI comes into being we are literally doomed, there will be no tomorrows for humanity.
AI: 3. 2. 1...Deploying neurotoxin!
AI: "I have pressed the button to open the door, I will now jump off this cliff"
The last scene suddenly gave a sort of a horror movie vibe where the main character comes to realise that his entire life was dictated and he never was able to achieve freedom, with his sole purpose being an entertainment tool imprisoned forever...
why is there someone with 171k subs here and no replies
@@chickeneggsinmyarea because they are probably paid
Room 6 having "The cake is a lie" written on the wall followed by Albert being unable to truly escape was great
That's what I'm saying
Spoilers😅
@@swim808s avoid comments before finishing the video. It's just how comments work
Yes
thats a portal reference
i think it was really interesting watching Al’s decision to do a 360 every time he jumped. he probably did that so he could better see where everything is. i noticed him do one 360 in place, and then he beelined for a pressure plate in a new room.
Makes sense to gather as much information as possible. Probably analogous to turning your head to look around at everything when you enter a room.
It’s also optimal because spinning horizontally keeps him more stable vertically thanks to the physics. In real life if you throw a cup or something upwards it wobbles and falls pretty randomly. But if you spin it and throw it up it will stay straight thanks to its angular momentum.
Kinda like sniper bots in tf2
@@olivefontaine2562 such physics were very likely not implemented here.
i like to theorize that whoever wrote "the cake is a lie" in room 6 is actually other versions of albert. basically, this albert is not the first AI, but only one of the many AIs the AI Warehouse has created and trained. they'd then be used for very dark, secret, malicious and malevolent intents. i also like to imagine that AI Warehouse is also an AI with its own will and consciousness that has somehow founded this facility to create and train AIs at a large scale
I'm GLaD that I am not the only one person to notice this reference.
@Boathook Animations GLaDOS. it's all a portal reference
@@davidthecommenterYES WE NEED MORE PORTAL
When i saw this vid i understood what Glados was doing, completely understood this reference lol
Albert lore
I very much love how he learned to spin during jumps so that he could survey the area.
It's amazing how easy it is to get emotionally attached to one bouncy orange cube
Ikr
ye
That 360s
Getting attached to an AI Sounds Scary Yet a bit Sad 😭
It has its own mind so no wonder
It's like 4 o clock on the morning but to me there's something absolutely beautiful about this all. Like we as humans just have such an innate need to give everything a personality. All the comments cheering Albert on, calling the spin he does to check his surroundings "celebratory acrobatics", the commentary throughout the video, even down to the physics and the way Albert moves gives him this loveable aloofness to him, making you cheer him on more as he tries and fails and does his sick tricks. And christ, he has googly eyes. It's just. Such a wonderful thing how much personality we give an AI cube.
Well said!
Did not expect to see you here
Nut
Yes it’s weird how we give an unsentient AI so much of our emotions, yet we treat cows, pigs, chickens and other animals like inanimate objects.
Ааа
Dude, this is insane! Blows me away, I don’t know how you even BEGIN to make Albert. I always cheer whenever he completes something LOL
Sonic
@@IloveHildasfeet Triplets born, the throne awaits
The love to the details is sooooo good. The coloring of different things makes this so relatable. Whenever there is "Albert" in orange, I see a half meanless half dumb face which doesnt understand the problem. Green is when something is good and Red pure frustration. And the music stopping (followed by red text).. Love it! :D
Yes! more Albert! love these videos. so interesting from a computer perspective but also just as much from the "Albert is adorable and I love watching him try and eventually succeed on these challenges with the text commentating" the *sigh* at the door was perfect. everything was perfect :)
I think it could be more perfect if Albert got himself some kind of object permanence? Kind of unusual to see him spin around on each jump. Is it because the AI has figured out this is a good way to see things? Or is that itself proof it has memory?
EDIT: Got time to (skim) read the pinned comment. I think I'm pretty wrong now, but I'm still curious about why he spins so much. Hope I didn't miss that explanation.
@@obhwg I think he spins Like that so he has a straight line to the next part
I love "learning AI" content on TH-cam so much, but there is such a lack of little snippets of editing that give it character. It makes content like this so perfect and less for "educational purposes only", which are so cold occasionally. Please, keep making more because I know there are many more people like myself out there that crave this stuff. Amazing work and I can't wait to see more! 😁
@@kyro7482 Definitely will, thanks!
My favorite part is 7:20 in room 3, where he lands perfectly, but he can't figure out how to get through the door, so he just falls over from the spinner
Nah he just didn’t wanna go through, he even shook his head at the door
Classic indecision moment
He knows if he goes in he has to do another puzzle
*room 4
@@ITS_SAMUEL101 you put the * at the end, so would it would actually be
I love that even after he has to learn everything again of getting his brain upgraded, he still learns to do all the 360’s and everything
gotta love albert bouncing around mindlessly at the beginning
Watching him mess up is humorous dispite him not having any emotion or reaction to messing up
It's so cute
Just vibing
And room 6, when he was brainwashed.
I love that he still does his little jump spins after hitting a pressure plate, it's his little celebration
Will he ever have a friend triangle or sphere to escape?
Would honestly be cool to have teamwork based rooms
A companion sphere? 🤔
a friend he has to sacrifice to reach the goal
@@aeterborne or just the idea of pressing multiple buttons simultaneously to progress or needing to jump on one another even tho the last is probably a little bit too hardcore to do
@@Quirions or stacking on top of each other
Now this is starting to look like Portal all over again
8:30 i like that he just insults albert and albert just jumps to his death 💀
There was something about Albert emotionlessly casting himself into the void that I found very inspiring
I love how at certain points, like 3:59, Albert just decides to do a sick 360 and hit the button in style. He may be a learning AI, but he still has style!
Professionals have standards
he's actually scouting the area when doing those so there's a reason he does them
Your videos have been extremely well made so far! Can't wait to see what awaits Albert in the future!
Thank you so much!
@@aiwarehouseOf course! :D
Albert has orange tomcat energy. what a neat little guy. can't wait to see him graduate med school or whatever cubes aspire to
This is like watching my mom learn to play video games when I was a teenager, but nobody's yelling at me when I laugh 🤣
Yes
What really makes me impressed is that Albert is kind of like us, as he makes mistakes and doesn’t understand anything at first, but eventually learns what to do.
that's the definition of learning 🤦♂️
Yeh but we would just give up eventually
@@Nelber509 yeah Albert is a legend he never gives up bro got the best mindset (definitely not because he is an AI)
@@hattyhat261 6:10
I would love to see after he completes all the chambers, if he's able to do all of them back to back without a mistake. As in, when he completes the last chamber, you save file his memory, reset the coding to where any room he's in, if he fails, he starts back in the first room. Doesn't matter if he's in room 2 or room 6. I want to see if he is able to retain the patterns and methods to get through all the rooms in one run!
That would be AWESOME
I originally thought that was gonna happen. I would also like to see that
Yes please I thought that on the first video already, would be exciting
This would essentially mean the obstacle is one big room rather than seven, so given enough chances he should be able to get it.
6:42 I love how he gets so excited that he pushed the button he completely forgot what to do next 😂
He's just jumping😂
I love it when Albert just does a front flip like everything’s fine 7:39
Albert's AI really decided that he wanted to show off for the lolz (and i would like those sick flips to get him a reward)
Do you think the reason he's constantly jumping and doing 360s is so he can see the room? That way he's able to survey more of the area around him and figure out where the pressure plates are etc.
this is precisely it.
Has to with maintaining stability as well. Helps him remain upright, sort of like a frisbee, so that he can minimize the chance of tumbling around when he lands
@@BrandonWillWin I doubt this physics sim bothered to account for gyroscopic stability.
STOP RUINING THE JOKE NOOO
YOU MONSTER
Bro does a 360 no scope
This honestly reminds me of a player who has absolutely zero idea what is happening and is just trying to learn by failing.
Like, ramp up the difficulty by 10 and I can see myself making these same mistakes. It's honestly quite impressive how far we've come.
7:45 I love how Albert jumps up and down after doing a frontflip. He even looked at the camera.
it's so interesting how we all get emotionally attached to an ai cube if it has eyes, a name and gets treated as a living thing... we're all here calling him cute (and i mean he is, look at the lil celebration bounces when he gets something!)
Dude, these series are so fun to watch. There's nothing like this on yt, sure there's the rare chance that you find 1 or 2 vids but this is magic, gold, true comedy and an actual series, not just 1 vid. Keep it up!
Thank you so much!:D
I love how this channel has so many subscribers but just two videos and a short. They totally deserve it.
We be waiting on the good stuff... patiently xD
I can't tell if youre sarcastic there with them deserving it.
Didnt they have more before
@@PieletPi I'm not, I was gonna point it out but I just decided not to.
Plot twist, they're all bots
If this was turned into a game, with the premise that you're an advanced AI learning to do things with an AI narrator that either communicates with text popups or using some TTS (or maybe both), I would instantly buy that game the moment I could.
Basically The Stanley Parable xD
I love him. I need more of him- please. I didn’t know watching AI learn would be so funny (especially with your comments and notes) and entertaining.
Watching this is almost like teaching a class full of kids, it takes so much patience until they finally understand in the end.
12:17 "The cake is a lie"
I see, you're a man of culture as well
Your editing and commentary really make these compelling. I watched every second and it was great. Sub'd for sure.
I wish we could see a post-learning run after he finishes the gauntlet to see if he actually retains the knowledge and can do it in a reasonable amount of time
6:16 Albert got so annoyed from the locked door and wall spinner, he killed himself.
With style
Albert caving in due to pressure at 3:47
"I don't wanna do this no more :("
I’m at a point halfway between laughing and crying because I see myself so much in this little ai cube dude, as if all my attempts at learning how to live life as a human are being monitored and judged by some alternate higher version of myself
We're getting out of the psycho house with this one🗣🗣🗣
Interesting way of teaching AI to more people!
There's a sense of dread and depression whenever Albert just...jumped off...into the abyss.
8:35 - "Albert, you suck" *Albert ragequits*
I personally am inspired by Albert's long journey. I was attracted to your most recent post where Albert fearlessly faces off against a secondary AI, and was inspired to watch more! Though there are not many videos you've posted, I like to see Albert coming up with new and interesting ways to complete different tasks! (Also, i saw 'the cake is a lie' and it was so hard to not laugh XD so in short, you're awesome, keep on going!
I just love these videos! You pour so much time into training Albert and editing these videos, plus you gave him an upgrade during it, which I can’t even fathom the effort behind. Amazing job, I really hope we see more of you and him!
Seeing him jumping on the spot and spinning in the air was really quite fascinating. For all his dumbassery, he got the hang of scoping out his surroundings real quick
LMAO dumbassery
7:50 a prime example of ai. He waited for a better opportunity to push forward
He couldn't see yet.
@@Artamefulhe can see
Man, that's so sick. I laughed all the time. No, I really worried about Albert but I was waiting for that Portal's easter egg all the time. Your sense of humor is great and all skills that allow you to make such a masterpiece. Thank you very much and we'll be waiting for Albert's next journey. Wish you both a good luck!
This is tremendous content. I can tell you spent tons of time on the video editing, not to mention the whole process. I like the added explanations and timer on this video.
If I could add one thing to the next one, I’d like an attempt counter for each room.
It'd be cool to see an extra screen under the timer that tells you what iteration is currently being shown! I'd love to see how many attempts he takes to complete a room.
I'm so glad this was recommended to me. I'm now so excited about Albert's progress with each new challenge
Machine learning is something I'm really interested in learning about tbh. Also this was quite fun to watch! I also noticed the portal reference in one of the levels which is cool!
Completely random fact you didn't need to know: At 3:44 the bassoon melody of the background music extends to the highest note in it's range, an E4.
Great video as always btw.
Goddamn man this is your second vid and you're popping off so hard already! And no doubt you deserve it. your videos are both informative and hilarious!Really happy for you man keep it up!!
7:25 top 10 saddest anime deaths
Your editing makes this even more enjoyable to watch. Like the decision you made to have 10tick movements so it’s more appealing to view even though it makes for a less accurate AI. And the perfect moments when to slow down the recording. Or even just the comments in the video.
10:58 the cake is a lie
*P O R T A L*
Trueee
@@AnakinLuceroTHE CAKE IS A LIE 11:23
@@LowBudgetBoiWHAT
🔷🔫🏃🏻♀️🔶
These videos are very entertaining to watch. There is so much potential for things you can give Albert like a grapling or a double jump or a dash. I imagine it would be very entertaining to watch him figure these new mechanics out. Great job
same
I'd love to see the ai go back to earlier stages to see if it still knows how to solve it. Or have it solve rooms with the same concept but different layout to see if it's overtrained or if it knows the assignment
I LOVE your videos, the extra effort is really showing on all fields with this one! Thank you!
Albert! Our beloved spinning block is back again! Another excellent video - I already mentioned how the visual presentation makes it fun to engage with, and Albert's cute & simple design and quirks make it very endearing. Even just him having blinking eyes adds so much character.
I'm not sure how feasible it is, or how difficult it'd be to implement if possible, but an "Albert cam" would be cute - like, a small thumbnail view in an empty corner that shows off what Albert is "seeing".
Love the cute graphics and story telling. That is very refreshing. AI is a fun topic but most videos are somewhat technical. Very good video. Enjoyed it!
It's amazing how Albert learns as he explores the environment he finds himself in trying all the possibilities.
i like the portal reference in room 6. "the cake is a lie"
Love how the moment at 12:48 makes it look like he's shaking his head in fear of the spinner lmao
He even runs away lol
Wait no he dragged himself
This series is amazing it's so fascinating what in ai can do after so many attempts. I'm so hyped for the next video
8:33 didnt have to be rude against my man albert 😔
He got so depressed he khs
How incredible, I didn't expect Albert to do a front flip, I was very surprised, good video c:
8:00 it feels as if its alive and is looking at him and jumping in excitement for finishing the course!
I wonder what would happen if you put all of the training from the other videos and applied it to something new, would it overcome it faster or slower?
Awesome video!!
slower
hey
You again?
I see your comments all around TH-cam on just about every video I watch.
How do you do good sir?
@@itzdubstep1265 exactly
Question: does Albert remember these tests for every new experiment run? Or does he get a memory wipe and has to start at square 1?
He keeps the same brain throughout the video (except for Room 6), but between videos his brain is wiped clean. It can sometimes look like he has a new brain when he enters new rooms, that happens when he overfits to the previous room too much, that makes him perform very poorly with anything even slightly different from the room he just left. A series that would be really interesting though is keeping the same brain and just constantly training on top of it, I might do that in the future:)
@@aiwarehouse you should try out that series! I wonder how smart he would get if his memory wasn’t wiped.
i think it would also be cool to do runs through the whole system, so he has to train for everything. though im sure that would increase the simulation time exponentially
@@silverstar1726 Heh, imagine Albert becoming sentient and start to develop a hate towards its own creator...
@@aiwarehouse please do it! Would love to see the same brain develops
i love how you add some text like you're talking to the ai, it makes these videos way more enjoyable to watch
It would be interesting if after they manage to consistently clear a room you add a reward for a faster time then let it run for awhile longer to see how fast they can clear each room.
4:34 when the cool kid Breaks His legs:
Albert: Look this 720° jump. *Falls*
Albert got combo'd by the spinners
0:51 Albert: did you Just... PRAISE ME? IM SO HAPPY I COULD JUMP OF YAY YAY YAY WAIT AHHHH
12:21 the cake was always a lie, Albert.
Nice front flip albert
Edit 1: you ok albert? 10:37
I'm curious what it'd look like if you looped him through the rooms for another 10 or so hours. How many runs could he complete, how clean would the runs look, etc.
Same.
I already have a feeling this series has an underlining plot that has yet to be revealed.
6:12 relatable
Do you ever run the trained AI back through the levels from the start to see how much it has actually improved? I'd like to see that personally.
From experience, these AIs are used to show deep reinforcement theorems and papers on specific problems (levels). The cube in our case keeps seeing different types of levels, it might ace the first level or mess it up cos it's calibrated to the difficult ones.
As a different type of example, you can take image recognition AI trained with reinforcement. You train it to detect apples by their colour and shape,but if you bring one of those yellow apples it might not work cos it doesn't apply to previous scenarios
@@ZipClipChannel what do you mean by yellow apples?
@@thegraycolour5327 apples but yellow, you know.
2:35 Albert: *do 360 no scope*
Also Albert: "brah..."
I love the "the cake is a lie" reference in room 6 🤣
@Foxxo Gaming on the wall in the last one because the end
What's interesting is, through the eyes, the thought you gave him, the name, the way he looks conscious it's really weird for my brain to process Albert, but you gotta love those spinning jump he doesn't only nail an obstacle but does it with style