Other UA-camrs would worry only about the amount of views that they get in their videos. Károly is the only one I've seen that sacrifice views in order to give his audience a better understanding of certain topic in which he's not an expert. Respect!
"It's not about winning this particular game, it's about playing in such a way that you get invited to the maximum number of games." -Jordan Peterson (Maps if Meaning lecture "Story and Meta-Story")
Attaching these nice graphics to the environment is the best PR stunt in reinforcement learning Deep Mind did so far. I'm happy these graphics make RL more interesting to the general public. Thanks to you for enabling further dissemination!
@@GabrieleNunnari I often wish all academic departments would have a team focusing on scientific infographics and helping their researchers with beautiful figures.
@@AICoffeeBreak I do also wish that academic departments would have a team working on that, but in research there is already the bare minimum amount of money to make research itself. A proper presentation would require another big investment, or a very talented researcher that is willing to take his work to the next level and spending his valuable time in learning how to do it.
@@atlascove1810 _"Heh, as if such an absurd thing could happen."_ - Homo Sapient (evolutionary carbon lifeform) *No.[EXPUNGED],* Earth (former of New Terra)
Not it didn't. To figure out the concept of destroying other beings to defend yourself, you would first need to understand what "beings" are, have a concept of self, and figure out destruction. This agent doesn't understand anything. Don't anthropomorphize an optimization algorithm.
@@rantingrodent416 I'm not anthropomorphizing, he has to destroy the other agent to keep winning right, its purly optimising. a purly "optimising agent" advanced enough can be threat to humanity just because we are inconvinenant to its goals. Its expected but concerning that this behavior is already being seen in prototype AIs.
@@rantingrodent416 it kind of does understand it needs to defend itself, but at the same time it doesn't understand that it itself is a being, since it's just an AI with no sense of self.
Just the fact you essentially said ‘I want to talk about this, and it would get a lot of views, but I don’t feel I’m at the level of knowledge/understanding to give it a proper video yet so I’m gonna hold off’ is amazing and got my subscription. Keep up the great work!!!
@@OnEiNsAnEmOtHeRfUcKa eh, card games are a lot different controlling a character in 3d or even 2d space. You might be better off following chess and go AIs
You give them simplistic, generalised functions like "maximise your points". The AI throws stuff at the wall until it notices something that increased points and then hones in on that, and repeats until optimisation has occurred.
@@PinataOblongata is that the actual number, only a thousand? Normally I thought reinforcement algorithms take millions of games for complicated stuff like this
I keep finding myself holding on to my papers. I think even if I only had a passing interest in machine learning, I'd still be watching this channel. You've created a great portal into piquing interest and promoting learning in your field. It's a great time to be alive. And we'll see you - next time.
My favorite part of the OpenAI hide and seek experiment was when the seekers learned to exploit the physics engine by grabbing the box that they stand on, or sandwiching themselves between the wall and a ramp.
Justs when i thought i couldn't like your channel anymore you admit your limitations in reference to covering alphafold, which I consider the height wisdom. Keep up the great work and thanks for the inspiration.
I have a feeling that there's soon going to be an A.I. (through its process of elimination) that will finally be able to always generate many trillions of very beautiful unique images from learning what each of us love.
I think it's worth pointing out that the main novelty of this work is a method to generate huge variety of unique worlds with "smooth" transitions between them. Each world is defined by 3 independently generated parts: environment, players, reward functions. This allows agents to be trained on a big variety of unique tasks and this is how trained agents succeed at 0-shot holdout set of manually crafted tasks (like hide and seek and capture the flag). In some sense 0-shot learning in XLand is similar to GPT-3.
It is amazing to think, for millions of iterations in evolution. Like the AI that runs around aimlessly for millions of games, millions of generations of creatures have died due again and again to be able to benefit the entire species through gene selection.
As a minecraft parkourist, (someone who plays parkour in minecraft alot) I wonder if this ai can find ways to do jumps in minecraft with set distance to gain momentum, and distance to end goal. We've never been able to do a 5 block jump using flat momentum (like only a floor, and nothing else) and have proven it using bare force and math, untill recently when we found a common mod called optifine had an exploitable speed miscalculation (on some old versions of it) which can give a speed boost by turning over 900million degrees, resulting in an about 1.03% speed increase, making a 5 block jump possible on flat ground.
They teach this in Cognitive Science classes in Cognitive Psychology too; these are Agent-Based Simulations, they can be designed from multiple softwares, but this one's with a learning A.I. which strategizes between past knowledge, but also what the agents are pre-configured and how the maps can let them do or not. Example: the maze experiment is a basic one, where the programmed agent can move in all 4 directions, has limited field-of-view so when it hits a wall, it will either try another direction or backtrack and find another way.
So incredible the progress made. I can only imagine the potential this kind of AI would have within video games. It would make video games far more immersive and interesting.
As a structural biologist, it makes me pleasantly happy that you are giving alphaFold and the protein solving problem its due diligence! We've already started using AlphaFold in our lab to start asking some tricky questions that we had only small amounts of experimental data on. AlphaFold is certainly not perfect but boy is it impressive!
How do agents "see" in these games? I mean, is there an image-recognition progress so they can understand they see each others by checking pixel color? OR, is there a data coming from game engine (like ray-cast result from Unity, Unreal Engine or OpenGL)?
The AI bending the rules to get closer to the pyramid is a perfect example of how asking AI to “end human suffering” will result in ending human existence 😅
"It's not about winning this particular game, it's about playing in such a way that you get invited to the maximum number of games." -Jordan Peterson (Maps if Meaning lecture "Story and Meta-Story")
Very thoughtful of you to not speak about something you think you don't know enough about, too many people spread false information these days in an attempt to seem clever. It takes a wise man to admit ignorance.
I wonder if this AI could work in a different modality, say, in language instead of 2d environment. It'd be a nice addition to some NLP neural net, like GPT-3, acting as some implementation of long-term planning.
Kudos for knowing when there is a topic that you are not ready to cover yet. It is an admirable ability which appears to be in short supply on the interwebs. People who can't recognize that about themselves clog up the airwaves and make it harder to find the knowledgable folks.
The fact that the AI in King of the Hill has learnt to master eliminating its opponents could escalate in real life! AI has no intent to harm, but if it is given a goal, then it will probably find ways to beat humans in the real world!
@@danielyuan9862 Also it could literally be designed to have intent to harm so thats not necessarily true either. It can be anything you tell the computer to be...
when they prune these sort of neural nets, typically how much smaller can the memory/computational foot print become over the initially trained neural net in terms of retaining say 99% of useful functionality, or are they just not that optimizable in terms of these things because they de facto self optimize resource usage?
Great video! I would love to hear your thoughts on some of the things revealed in Tesla's latest AI day. A lot of it goes over my head but when you get excited about something I know I should be too. I believe they even mentioned the Photorealism Enhancement paper you made a vid of a few weeks back. Do you have a platform where you may share such thoughts as its not exactly a paper?
How about you just interview one of those experts on AlphaFold? You may not have fully grasped the thing, but your 1 million subscribers will be delighted to see you having a chat with somebody about the topic. Just geeking out on all the technical stuff. Make it a "Two Minute Paper - Long Form Edition"
Great video, really fascinating what these generalized agents can do. ps. Friendly neighborhood Biochemist here if you want to ask any questions about proteins that you are struggling to understand happy to help
Teaching AI the concept of the strike before struck is hard. It's easy to just give them "rewards" when they attack something, but it could lead to just to them camping the enemy's spawn point and not fulfilling the objective. Surprised the AI already passed this point considering how complex this is. Could also be they're brutal, like Anakin.
Can you do summary of last few years of progress of AI and near future goals and regroup of more specific games achieved by ai deepmind atari, chess etc..
Thumbs up for maximizing meaning.
Other UA-camrs would worry only about the amount of views that they get in their videos. Károly is the only one I've seen that sacrifice views in order to give his audience a better understanding of certain topic in which he's not an expert. Respect!
Mad respect
Lets see Paul Allens Meaning.......
His credibility was greatly enhanced by acknowledging the extent of his present abilities.
"It's not about winning this particular game, it's about playing in such a way that you get invited to the maximum number of games."
-Jordan Peterson (Maps if Meaning lecture "Story and Meta-Story")
Attaching these nice graphics to the environment is the best PR stunt in reinforcement learning Deep Mind did so far. I'm happy these graphics make RL more interesting to the general public. Thanks to you for enabling further dissemination!
I was just thinking the same. A nice graphic does allow to understand what is happening and is also "pleasing" to the eye
@@GabrieleNunnari I often wish all academic departments would have a team focusing on scientific infographics and helping their researchers with beautiful figures.
If it wasn't for videos like this (and yours) hobbyist programmers like me wouldn't be interesting in trying AI projects.
@@AICoffeeBreak I do also wish that academic departments would have a team working on that, but in research there is already the bare minimum amount of money to make research itself. A proper presentation would require another big investment, or a very talented researcher that is willing to take his work to the next level and spending his valuable time in learning how to do it.
@@justinwhite2725 :blush:
"And you would think that the Starwars references would end here, no.
Not even close, look(Luke)"
That was smooth! 3:45
I thought he was about to point out the rotating agent, and say "Ah, let's try spinning - that's a good trick".
4:00 "grabs his lightsaber, and takes the high ground"
"What a time to be alive!" could have a whole different meaning in the future when we are the hiders.
^^^^ This
I swear every robot uprising joke will be used against us.
@@atlascove1810 _"Heh, as if such an absurd thing could happen."_
- Homo Sapient (evolutionary carbon lifeform) *No.[EXPUNGED],* Earth (former of New Terra)
bump
The fall of the global economy will come first, global anarchy will be first.
5:58 AI already figured out the concept that you need to destroy other beings to defend yourself, this is earlier then expected.
Actually, it was taught that
Not it didn't. To figure out the concept of destroying other beings to defend yourself, you would first need to understand what "beings" are, have a concept of self, and figure out destruction. This agent doesn't understand anything. Don't anthropomorphize an optimization algorithm.
@@rantingrodent416 I'm not anthropomorphizing, he has to destroy the other agent to keep winning right, its purly optimising. a purly "optimising agent" advanced enough can be threat to humanity just because we are inconvinenant to its goals. Its expected but concerning that this behavior is already being seen in prototype AIs.
@@rantingrodent416 it kind of does understand it needs to defend itself, but at the same time it doesn't understand that it itself is a being, since it's just an AI with no sense of self.
@@bronzehd6212 bruh its not smart enough to do anything you just said lol
Finally, true gamer AI. cant wait to see their steam libraries
What a time to be a gamer!
@@4GdaTim 😂
I'm just throwing an idea in the air.
Do you think someone at DeepMind would be interested in helping with the video on their protein prediction tech?
That would be kind of cool. I don't think I've ever seen him invite guest speakers for topics he isn't confident in speaking about himself yet.
Great idea! I would love to see it
With a million subs I think he can get who ever he wants for a 5-minute video.
Year 2030: 2 minute papers uploads are now 10-hour documentaries 👀
That are produced by AI.
2030. No man alive
Yep, just give the AI a 2 minute video as a starting point and it extrapolates the rest.
@@Naxt366 Oh my god... Women took over the whole world?
All that exists are AI bots that produce videos and farm views from other AI bots watching to generate their own relevant content
Have no idea whether to be excited or scared by these incredible advances!
Not long until AI makes all human decisions
Be both.
I remember that hide and seek!
I love how they made them smile and laugh while playing. It's kinda adorable.
@@webx135 Adorable until its your turn to hide
Loved how they broke the engine at some point
Get the experts to guest in your videos. Maximum meaning! Thanks for the great videos.
I find the military potential of AI frightening, the human innovation of AI downright fascinating.
I have no mouth and I must scream warned us about this
Great respect for optimizing for meaning and teaching, instead of views. Props to you!
Just the fact you essentially said ‘I want to talk about this, and it would get a lot of views, but I don’t feel I’m at the level of knowledge/understanding to give it a proper video yet so I’m gonna hold off’ is amazing and got my subscription. Keep up the great work!!!
My hope for this eventually working in Super Mario 64 keeps going up.
I'm excited for when it learns to play card games.
@@OnEiNsAnEmOtHeRfUcKa eh, card games are a lot different controlling a character in 3d or even 2d space. You might be better off following chess and go AIs
Make it rediscover all the A press saves from scratch and see how long it takes
Took a while for another video, glad youre back!
When you're a doctorate and your videos are top quality it makes sense life will delay these masterpieces of information.
Interesting video, but I would've liked to know how these agents were trained
Yea, I would be curious to know how many games they were shown
@@devanmallory5304 They run through thousands of iterations.
You give them simplistic, generalised functions like "maximise your points". The AI throws stuff at the wall until it notices something that increased points and then hones in on that, and repeats until optimisation has occurred.
@@PinataOblongata is that the actual number, only a thousand? Normally I thought reinforcement algorithms take millions of games for complicated stuff like this
@@devanmallory5304 depends on the thing, could be thousands or millions
Training artificial intelligence is definitely my favorite topic as of right now. Thank you for the awesome videos. love what you do.:)
I keep finding myself holding on to my papers.
I think even if I only had a passing interest in machine learning, I'd still be watching this channel. You've created a great portal into piquing interest and promoting learning in your field. It's a great time to be alive. And we'll see you - next time.
"These agents are not preparing for an exam, they are preparing for life" ... or dear, we are doomed :D
We got an 8 minute paper today!
Thanks for Maximizing Meaning! DeepMind is exploring the impossible and it's inspiring to see.
Considering the tag video was my favorite video so far this is even better since it’s improved so much
My favorite part of the OpenAI hide and seek experiment was when the seekers learned to exploit the physics engine by grabbing the box that they stand on, or sandwiching themselves between the wall and a ramp.
4:06 Don't do it Red Agent, Green Agent has the high ground!
Justs when i thought i couldn't like your channel anymore you admit your limitations in reference to covering alphafold, which I consider the height wisdom.
Keep up the great work and thanks for the inspiration.
I have a feeling that there's soon going to be an A.I. (through its process of elimination) that will finally be able to always generate many trillions of very beautiful unique images from learning what each of us love.
0:20 that was the video that made me sub to your channel.
I love how you made a two minute long summary of another two minute paper.
I think it's worth pointing out that the main novelty of this work is a method to generate huge variety of unique worlds with "smooth" transitions between them. Each world is defined by 3 independently generated parts: environment, players, reward functions. This allows agents to be trained on a big variety of unique tasks and this is how trained agents succeed at 0-shot holdout set of manually crafted tasks (like hide and seek and capture the flag).
In some sense 0-shot learning in XLand is similar to GPT-3.
It is amazing to think, for millions of iterations in evolution. Like the AI that runs around aimlessly for millions of games, millions of generations of creatures have died due again and again to be able to benefit the entire species through gene selection.
We are exactly the same except our environment is more complex and we experience time differently.
Phenomenal finding and equally spell binding narrator. Keep up the great work. Meaning will prevail.
This is so cool! Thank you for sharing in such a clear and understandable manner :)
Deepmind's really putting out some magic recently. I can't wait to see what this kind of research means for game agents!
These videos make my day!
The AI catch paper is already two years old?! How time flies!
I'll read the paper but I'd love a bit of information about how these AIs were trained and how the new problems were presented to them
Can't wait for this exact footage to make up the first five minutes of the next RL video!
As a minecraft parkourist, (someone who plays parkour in minecraft alot) I wonder if this ai can find ways to do jumps in minecraft with set distance to gain momentum, and distance to end goal.
We've never been able to do a 5 block jump using flat momentum (like only a floor, and nothing else) and have proven it using bare force and math, untill recently when we found a common mod called optifine had an exploitable speed miscalculation (on some old versions of it) which can give a speed boost by turning over 900million degrees, resulting in an about 1.03% speed increase, making a 5 block jump possible on flat ground.
that'd be cool
General rl ai vs baritone bot
900 million degrees quick scope
What are you smoking. You can totally do a 5 block jump, that's the furthest you can do though. And optifine is client side only. Wtf
@@Grocel512 it makes a weird calculation and just breaks it, giving more speed than usual lol
I am so tired my eyes feel like they are about to fall out. But I need to watch this video before I sleep
Thanks... for making... this informative...video!
Haven't seen interesting stuff around YT in a while, very nice
Remember making sense is also mental. So working together is one single sense where every atom in the Universe is a sense.
4:00 Takes the high ground while also spinning for a good trick, Uses both sides of the force this one does.
They teach this in Cognitive Science classes in Cognitive Psychology too; these are Agent-Based Simulations, they can be designed from multiple softwares, but this one's with a learning A.I. which strategizes between past knowledge, but also what the agents are pre-configured and how the maps can let them do or not.
Example: the maze experiment is a basic one, where the programmed agent can move in all 4 directions, has limited field-of-view so when it hits a wall, it will either try another direction or backtrack and find another way.
Ahh... I really love these kinds of stuff.
I hope more these kinds of game emerge and want to see their creativity!
can't believe its been 2 years since that open AI video.. its what got me into this channel lol
So incredible the progress made. I can only imagine the potential this kind of AI would have within video games. It would make video games far more immersive and interesting.
You can invite for an interview! And let the creators explain in a short amount of time, that would be a nice experiment for the channel!
Maximizing meaning? You just maximized my heart with that line man
As a structural biologist, it makes me pleasantly happy that you are giving alphaFold and the protein solving problem its due diligence! We've already started using AlphaFold in our lab to start asking some tricky questions that we had only small amounts of experimental data on. AlphaFold is certainly not perfect but boy is it impressive!
TMP: "look... Boom"
Blue agent: *Disintegrates*
Is Alpha Fold different from Auto dock Vina?
How do agents "see" in these games? I mean, is there an image-recognition progress so they can understand they see each others by checking pixel color?
OR, is there a data coming from game engine (like ray-cast result from Unity, Unreal Engine or OpenGL)?
guess the 'games' backbone is standard CG. Just the decision making is AI/ML based but I might be wrong.
Were do you find these papers?
The AI bending the rules to get closer to the pyramid is a perfect example of how asking AI to “end human suffering” will result in ending human existence 😅
I recommend Bad Writing Advice's video on evil AI
The hide and seek paper video was the first one I saw from you
This guy single handedly made me interested in this type of things. And I think I've seen this scene before in a video.
"It's not about winning this particular game, it's about playing in such a way that you get invited to the maximum number of games."
-Jordan Peterson (Maps if Meaning lecture "Story and Meta-Story")
Oh god. Every game is democracy
Very thoughtful of you to not speak about something you think you don't know enough about, too many people spread false information these days in an attempt to seem clever. It takes a wise man to admit ignorance.
And I thought that Hungary had nothing going for it in the youtube scene, but here I am watching this video!
Maximizing meaning. Thank you.
Subscribed for maximizing meaning.📈
I wonder if this AI could work in a different modality, say, in language instead of 2d environment. It'd be a nice addition to some NLP neural net, like GPT-3, acting as some implementation of long-term planning.
6:20 When deep mind is open minded and open mind is narrow minded xD
All hugs and puppies until the AI realizes that “stop the ball from touching the red floor” is best achieved by destroying either.
Kudos for knowing when there is a topic that you are not ready to cover yet. It is an admirable ability which appears to be in short supply on the interwebs.
People who can't recognize that about themselves clog up the airwaves and make it harder to find the knowledgable folks.
I was waiting for this, bless you.
The fact that the AI in King of the Hill has learnt to master eliminating its opponents could escalate in real life! AI has no intent to harm, but if it is given a goal, then it will probably find ways to beat humans in the real world!
+Suraj Kothari No intent to harm? Opposing countries militaries and people with nefarious motives... hello?!
@@eyeofhorus1301 "AI has no intent to harm" means an AI does not naturally have an intent to harm anyone. It has nothing to do with people.
@@danielyuan9862 Since it has no intent itself its intent is dictated by the people who use it you're only half right
@@danielyuan9862 Also it could literally be designed to have intent to harm so thats not necessarily true either. It can be anything you tell the computer to be...
Maximizing meaning. God I love this channel
when they prune these sort of neural nets, typically how much smaller can the memory/computational foot print become over the initially trained neural net in terms of retaining say 99% of useful functionality, or are they just not that optimizable in terms of these things because they de facto self optimize resource usage?
Great video! I would love to hear your thoughts on some of the things revealed in Tesla's latest AI day. A lot of it goes over my head but when you get excited about something I know I should be too. I believe they even mentioned the Photorealism Enhancement paper you made a vid of a few weeks back. Do you have a platform where you may share such thoughts as its not exactly a paper?
Lex Fridman has made a video of some of the highlights from AI day,
Sounds like you might find interest in it!
What the time to be alive!! Said Skynet before getting read of mankind...
6:27 I love the enthusiasm
how can you differentiate learning from memorizing (or perpetual trial and error) when you run millions of trials??
THANK YOU for everything you do!
1:55 the other guy is helping to bring one of the boxes closer for his buddy.
Awesome vid bro!
Yes, like other comments said: thanks for maximizing meaning
I'd like to see DeepMind play Mini Motorways
Can't wait for VR worlds in the metaverse with our AI NPC companions.
What a time to be alive!
I started beliving in AI with alpha go. If you play a bit of go you realize how amazing it is.
Spacetime bends on this channel. Two minute paper takes more than 8 minutes!
What about the methods used?
What humble comments at the end of the video. (I believe you could do the AlphaFold ! ^^)
Thanks for this awesome video, do you have the link to the protein structure prediction paper plz? :)
How about you just interview one of those experts on AlphaFold? You may not have fully grasped the thing, but your 1 million subscribers will be delighted to see you having a chat with somebody about the topic. Just geeking out on all the technical stuff. Make it a "Two Minute Paper - Long Form Edition"
Good training for Terminators!
Great video, really fascinating what these generalized agents can do.
ps. Friendly neighborhood Biochemist here if you want to ask any questions about proteins that you are struggling to understand happy to help
This is... Insane... This seems like the closest to general intelligence I've seen with games... Absolutely incredible
I have been waiting for a new video like this every day since the hide and seek paper. ! Thank you
Teaching AI the concept of the strike before struck is hard. It's easy to just give them "rewards" when they attack something, but it could lead to just to them camping the enemy's spawn point and not fulfilling the objective. Surprised the AI already passed this point considering how complex this is.
Could also be they're brutal, like Anakin.
Can you do summary of last few years of progress of AI and near future goals and regroup of more specific games achieved by ai deepmind atari, chess etc..
i have waited a long time for a follow up on the hide and seek video and this is great!
Awesome as always!!
The humility of this guy recognizing he don't have the knowlage to talk about proteins is just great.
ai is so incredible makes me excited for what is possible
YAY YOU MADE A VIDEO ON THIS ONE 🥳🥳🥳
5:54 I see red is doing a victory dance