DeepMind's GNoME Creates Materials | Schmidhuber Claims Q* | TLDRAW is out of this world!

Jurgen Schmid Huber so Jurgen Schmid Huber claims to have invented principles of meta learning the generative adversarial Network Transformers back in 1991 and in that same year also very deep learning according to Elon Musk Schmid Huber invented everything everything you've ever heard of and now Jurgen Schmid Huber it seems like he's saying he invented qar and he's got an image to prove it don't take this thing 100% seriously and we're going to cover it at the end in case you're interested but there's much bigger things that are happening right now and here's one so here's Demi Hobbies maybe arguably the number one person pushing AI forward outside of open AI so here's Google deep mine their sort of blog millions of new materials discovered with deep learning this to me is so strange conceptually to think about so the AI the neural network is kind of like like a brain it structures you could say similar to the neural structure of the human brain but instead of being made out of fat and protein and placed in our heads it's digital and they get trained so they become useful when we feed it tons of data we feed it tons of texts and it becomes something like Chad GPT where it's able to answer questions and maybe even do certain reasoning tasks if we take a neural network and we show in a bunch of images interestingly it starts creating excellent images there's even some research showing that it kind of builds these mental models potentially about how to create those images it almost has like a 3D representation of the world even though it's trained entirely on 2D images it starts figuring out like hey this object is closer to me this object this further from me Etc this is why it's so exciting is it seems to be like this sort of Technology we uncovered it goes much deeper than we would have thought 5 10 years ago and so Google deepmind in 2020 so not that long ago figures out how to guess at the 3D structure of proteins so proteins is basically these little life factories that make everything go around and function that you think of as life probably is on one level or another controlled by proteins they're made up of amino acids and those are pretty easy for us to sort of read right kind of like a letters in a sentence we can kind of see what those are but then those things fold into a 3D shape and that shape is insanely complicated the number of different unique 3D positions that that shape can be there's more of those possibilities than there are atoms in the observable universe so it's kind of a big number right can we say that it's not even astronomical it's like beyond that but slowly through time we were able to figure fig out the structure of these proteins so we collected a number of them that we slowly tediously painstakingly figured out and so we have like a database of them we train Alpha fold on that database Alpha fold begins correctly guessing at the 3D Shape of the protein it even is able to discover new proteins whose shape we did not know before which again to me is kind of strange like if you think about it too deeply because it's not guessing it's not like it's going through all the possible variations and it just keeps guessing until it gets the right one it looks at how all the other ones are shaped and then it figures out what this one is going to look like one that we haven't seen before but the point is that if you think about so we as humans we generate certain data we generate textbooks and images and then we feed it into the AI and it gets really good at whatever that thing is at producing images or producing text we feeded protein data and it gets good at guessing more protein data that we can't even understand how it's doing that so the kind of next big obvious question is okay what else does that apply to what other things what other data can we feed it to where it's going to come up with these insane solutions that we haven't even that that we can't even begin to approach and so I think a lot of people were waiting for this next shoe to drop and well I think today is the day or yesterday rather 29th of November because the same AI the foundation between a lot of these models from what we've seen from Deep mine they're saying they a lot of them are the same so the AI that plays Starcraft the AI that plays chess the AI that plays go that the AI that optimizes is machine code to make it faster all those breakthroughs are out of deep mine like they're separate sort of AIS but the the way that deep mine is building them they're they're similar and so this is yet another AI tool gnome finds 2.2 million new crystals including 380,000 stable materials that could power future Technologies and so they're introducing graph networks for material exploration genome or gnome are new learning tool that dramatically increases the speed and efficiency of Discovery by predicting the stability of new materials with gnome with multiply the number of technologically viable materials known to humanity meaning that this thing is able to sort of predict on paper right theoretically these materials that are stable meaning they can be used to build stuff with which means if we're able to synthesize it if we're able to make it in a lab we can potentially use those materials we we know what their sort of characteristics are going to be so we could potentially use them to replace materials that we've been using for a long time because these new fan materials are going to be potentially much much better so they have the potential to develop future transformative Technologies ranging from superconductors powering supercomputers and Next Generation batteries to boost the efficiency of electric vehicles so it looks like external researchers in labs around the world have independently created 736 of these new structures experimentally in concurrent work and they've published a paper in nature which is this one right here and they're referring to it as an autonomous materials Discovery platform and similar to Alpha fold they're also sort of creating a database where people can go where researchers can go and see those new materials and use them for their own research and stuff like that and this this is the picture that to me kind of really I think kind of sheds light on this concept that is so kind of mind-blowing I think so this deep blue circle that's human experimentation so that's all the stuff that we've been able to find using our best thinking our best brains Our Best Equipment right and then that's 20,000 right and then with computational methods we were able to find this sort of secondary Circle 48,000 so more than than double the amount that humans alone could find then comes this AI cuz keep in mind this AI a lot of these ideas like neural Nets they've been around for a long time but there hasn't been too much progress until very very recently and then when that progress came it kind of exploded it it hit that inflection point quickly and it seems like it's still I mean it's pretty obviously still rapidly accelerating so gome found 421,000 gnome expands the number of stable materials known to humanity to 421 thousand and so when people say that AI could be the greatest scientific discovery that Humanity ever achieves I think this helps understand why because what we're seeing is this next sort of evolution of AI it it almost teaches itself it learns from itself and it's beginning to make scientific progress so much faster than humans could so if we keep on creating this AI that eventually starts improving itself and learning more and more about science I mean one perspective is that it's will be pushing scientific progress forward much faster than than the humans alone so we might have invented AI that will then invent everything else and not only is the scale of this kind of Staggering but you also have to consider sort of the recursive nature of this the compounding so here they say for example 52,000 new layered compounds similar to graphine that have the potential to revolutionize Electronics with the development of superconductors so the question becomes okay so once this AI discovers all these compounds and we figure out how to synthesize it probably with the help of AI and then we use those sort of improved compounds to create better computers better batteries better better microchips that in turn will drive better AI the data it produces will help it predict and get better at producing more stuff the breakthroughs here will help us discover other avenues where we can use this so for example once Alpha fold Pro that we were able to guess accurately at the 3D structures of the proteins like this is this is an amazing Discovery but we kind of almost like expected more stuff like this there's genetic companies that are doing something similar with the human genome for example can we take this massive amounts of data about the human genome about our DNA can we feed it into an AI and have some sort of insight some sort of solutions maybe cures to certain diseases from that AI so at this point I I think it's kind of safe to expect more and more of things like this rolling out so this is kind of a low graph of how that neural network works you have the structural pipeline the composition Pipeline and this repeats for rounds of active learning and it comes up with these AI recipes so here's here's some things that we think might work the AI predicts that this might be useful must might be stable and then this robot goes to work and so here's that paper in nature.com so they're saying that the synthesis and characterization is done by a robot and so there's that Central robot arm made by Mitsubishi it's sitting there with powder dispensing and mixing and a heating station with four box furnaces and two collaborative robot arms that help kind of transfer samples and stuff and they have their material screen so they're just making sure they don't make something that blows up right so there's certain things that they're not touching that potentially could be dangerous but so here's what that thing kind of looks like so I'm assuming it sounds like they have it like behind glass it's its own little room just in case you know just in case uh maybe a few compounds don't mix properly and there's a big boom that happens but yeah I mean this is this is what's happening now we have research done by digital brains being carried out by robotic arms but that that's nice but is it working is it real well in 17 days of closed loop operation the alab that's that little robotic arm that moves around in a slow cage performed 355 experiments and successfully realized 41 of 58 novel inorganic crystalline solids with diverse structures and chemistries the unexpectedly High success rate 71% by the way it sounds like here they're using machine learning and active learning algorithms so they're using AI on that back end to to improve the synthesis so it's AI all the way down and so they're saying it seems like it's pretty easy they're going to bump up the success rate 274 with just a few easy simple tweaks and that autonomous research agents can marketly accelerate the pace of materials research this rapid Discovery points to a vast landscape of opportunities for me as someone who's been in Ecommerce since 2011 and have spent quite a bit of money on Google ads I'd like to think of myself as someone who helped fund some of these experiments that are leading now to a greater future you're welcome I'm kidding of course I didn't really have anything to do with it other lighter news here's how the next generation of humans here's how they will code take a look imagine you wanted to create a little timer and so you drew out how you wanted to look what the different buttons do right reset Stop Plus minutes per seconds and you're done here's that app it took about 5 seconds but let's say the computer got wrong so it should be counting down or whatever so you make a note and boom it fixed it you debugged it by writing that's wrong fix it and it's fixed here's creating a game where you jump over little obstacles and you get a plus one score every time you manage to successfully jump over it you draw it out make some notes click make real and so that took about 15 seconds or so there you go so you're jumping over obstacles and each time you manage to jump over it you get plus one to the score so it looks like the life system doesn't work but that's an easy fix here you put a photo of the celebrity here and then it's it's like hangman so you where you guess the celebrity right so you can see so add a visual keyboard modify the screen so they're adding one of the time sort of just adding different versions and building it up until it's finally complete and it looks the way that you want it to look and here's the final version of that so you're trying to guess who that is this person is guessing Jay-Z or I guess that was correct I'm going to guess Tom Hanks correct you're able to create an invoice generator so just sort of write out what you want on that page and within seconds it gives you this create invoice sort of sort of form you put in kind of what the invoice is and it will generate it for you by the way this generates HTML so you're able to take that code and go put it somewhere where it's going to run you can do it as an app on your desktop you can put it on your website whatever so if you wanted to try it for yourself so if you know what tldr is like too long to read so meaning like a summary so this is tldd draw.com so when you go to TLD draw.com this by itself is a pretty good sort of diagramming tool where you can draw and it's it's I got to say it's got really good drawing abilities just just by default kind of makes everything look kind of good which is very cool so you're able to draw you're able to do little text box like this and point the arrows to it right you want like a red arrow over here and add some text I got to say just just this alone is very well made for creating little diagrams or sketches or whatever but the real magic is you have to know the secret URL link I'll leave the link in the description and so it does require your API key so you head over to platform. open.com ai- Keys all right then you add your API key in there and so let's say I wanted to create a little timer just to start off with something simple and so I'll make a little box like this and I'll Point here and that's where the timer goes here start at 0 0 default show minutes seconds and let's say you forget what comes after you can say the next unit of time and we'll say timer counts down and then we'll have two buttons this button will be start I need brighter colors here how about this red and this will be reset actually this will be pause and we'll change to reset when pause is we also need like a plus one minute button let's put that right here add one minute to timer all right then we select all of this and we're going to hit make real and so this little box pops up and it starts generating the thing that we so masterfully drawn all right there it is so I like the fact that honestly well so okay so it didn't do pause didn't do the pause button the layout is a little bit different but I feel like it's because I kind of like screwed up the layout because I've seen other people when they do a layout that makes sense it shows up exactly correctly so add 1 minute start and it starts counting down so that's pretty good I think the I think the problem was more how I I set it up not not anything wrong that it did so let me do this so timer starts at 00 0 user can add time timer counts down in minutes seconds and milliseconds this will be add 1 minute to timer this will be add 10 seconds to timer this will be start timer and this will be reset timer all right let's try that I think that's going to work better all right so there it is I think that took about 20 seconds or so so let's add 1 minute so the buttons here don't work let me see what the issue is I think the timer units are throwing it off the milliseconds so add 1 minute so double click on it to interact add 1 minute perfect add 10 seconds perfect start timer and it starts yeah everything's perfect so add 2 minutes and 30 seconds start timer boom yeah had the milliseconds were throwing it off but whatever all right final thing let's do a video game let's do a quick game so here we're going to have the player that blob is the player and these things here are the chickens and then we'll make some notes here about the game so the chickens the chickens move in a random Direction when they hit a wall they bounce in a different random Direction the player player moves towards the mouse cursor when the player gets close to the chicken make the chicken disappear and add plus one to the score and we'll put the score here score goes here starts at zero all right so let's select this whole thing so when you click make real whatever you select is what gets ped so as you can see here I can have tons of different stuff going on but whatever I select is the only thing that gets uh created it's own special area here all right so that took I want to say I don't know 20 seconds or so all right so double click to interact uh well okay so that is cool but not quite what I had in mind because the player so wanted the player to kind of slowly move in Direction not not move as fast as the mouse so the player moves closer to the location every frame so let's try that if that makes a little bit more sense all right so that fixed the the cursor this is more kind of like what I had in mind but where are the chickens and why is my score so high start with five chickens located randomly start with five chickens located randomly on screen all right there we go I'll make that larger and let us begin well it's still it's still having some hard time understanding what this player supposed to be doing so let's uh let's spill it out so this is the player and this is the mouse so every frame make the player move closer to the mouse and that's the mouse all right all right so another 20 30 seconds and there we go okay that's exactly what I had in mind so we sort of drag the thing in the direction that we want to go and we can eat all of the chickens this chicken got stuck in a corner I guess okay so that's very good and we just need the Final Touch start the game with an intro screen that has a start button once the player hits the start button the game starts and then we need some sort of background for the intro screen so let's use this so use that as the background for the intro screen all right there it is so that took a little bit longer maybe 35 seconds I'm going to have to eat every Chicken in this room that is terrific start the game so double click to interact start the game and yes yes that is precisely more or less what we wanted this game will be available for purchase in the description of this video just kidding it's not but here's the thing by the time you're watching this this TL draw has already evolved to its next shape to its next form check this out so first of all they sped it up it's now much faster it looks like they're making it easier to create certain diagrams certain sketches you're able to create certain physics and bouncy shape slippery materials and rotating boxes so this is basically you can recreate Angry Birds this is Angry Birds and also they're adding uh something that I'm not even sure what to call it it's as you drag certain images or you create certain images within you know let's say a square it interpreted into something that you said it should do as you're drawing it so I I'll show you a different one it'll probably make a little bit more sense so take a look at this so we take this picture of a tomato and we change it into tasty delicious candy disgusting horrible vegetable like it's kind of rotten futuristic robot car yeah I don't even know what to call that it's like viewing some objects from different lenses from different perspectives maybe you're able to for example draw a letter and have it be sort of recreated in whatever form you want so for example from carved wood if you ever wanted to create your own font I feel like this would make it very easy here's like a little diagram of stuff falling to the floor wow oh didn't expect that dominoes here's a shot of it sort of organizing a diagram into just a be better layout from Messy to just more visual appealing easier to read so I got to say it's incredible how easy it is to make your idea come into life how intuitive it is the speed with which you can build this stuff and I apologize I forgot to show you one last thing so here's our little so our little game that we just created so we can copy the URL to the clipboard open it and uh send it to somebody else to play or we can copy the HTML to our clipboard post the HTML into notepad and save that as let's say chicks. HTML save sa that as all files save and so there's so there's that and you can run it locally in your computer so it's absolutely incredible to think where all of this is going I feel like development is going to get so much easier because if you think about what code is code is just whatever is between what the user wants and the final output it's just the stuff that translates sort of our thoughts what we want into the computer doing that but with AI more and more I think it's going to get just a lot more intuitive so it's not going to be like if you want to create something you have to code and this is how you code it's going to be more like how do you want to make this do you want to make it by sketching do you want to make it by just talking into a microphone maybe some sort of a back and forth conversation with AI do you want to do it through text do you want to point to something that someone else has done and say hey just kind of start there and let me iterate on top of it like start with Angry Birds but then I want to add my own twist on it Microsoft just came out with something they call task Weaver where now it seems like I haven't read the full paper yet but it seems like what they're trying to do is to have the AI kind of troubleshoot everything to take whatever code they produce for you and then go run it to figure out what worked what didn't and then kind of self-reflect on it before giving you the final answer so in that situation where we're building that little chicken game or whatever like it could probably take a look at and say hey this is not quite what I think we wanted let me try that a different way right maybe there's like a second AI that kind of like critiques it and maybe tries to make it better now you're still going to need to have the highlevel engineers and software Developers for like a lot of like the advanced stuff but if you wanted to create an app or a simple game we're getting to the point where I mean and child is going to be able to do it very quickly and finally that brings us to Jurgen Schmid Huber so I'm going to be honest I don't know too much about what's happening here there's sort of a dialogue happening here that I'm a little bit out of the loop because it seems like he kind of comes in and maybe a little bit aggressively accuses a lot of the prominent researchers such as Hinton Yan laon Etc of maybe citing each other and kind of like giving credit to each other but but he's saying but wait I've discovered everything before you guys and you know Ela musk comes in and says you know shmid Huber invented everything and more and more you hear this idea of Schmid Huber invented everything and as I'm kind of understanding it what's happening is that everyone is saying the same thing he's saying but everyone else is saying it ironically like they're kind of making it a joke and he's D was like no no no for real you guys I invented everything and certainly it seems like he's got a lot of clout so here he is in 2010 with Jensen from Nvidia here he is if you ask I guess if you ask the lamu model which Lama was developed by Facebook and Yan laon who is one of the researchers that you're going kind of maybe talks about as maybe not giving him enough credit you know so if you ask llama about Schmid Huber you're going Schmid Huber llama's like oh that's not a very good person I don't want to talk about him so somebody chimes in like looks like an answer that was hardcoded by Yan Lon so I don't know what to make of this thing here I've joked in the past that these people as AI gets better at writing lyrics and music and stuff like that they should create their own like diss tracks like the rappers used to do back in the days right so yuran comes in and does a distract by Yan Lon and then Yan Lon fires back they sort of have Chad GPT write out the lyrics with the main points you know lay down a fat beat and just release it for the rest of us to enjoy it's just a thought obviously I'd prefer all our AI researchers kind of got along and and played nice together but a part of me kind of wants to see that and just as I'm saying that it dawn on me wasn't there a YouTube channel that used to do that yes there was epic rap battles of history so these guys posed as various different historical figures that would kind of like rap battle each other they have Elon Musk and Mark Zuckerberg I know they have Bill Gates and Steve Jobs here somewhere there it is Steve Jobs and Bill Gates so maybe we're not too too far off from something like that happening we'll see anyways I hope you enjoyed that make sure you're subscribed we have a lot more stuff coming out it's going to be amazing AI is really heating up used to be I could do one video about one subject and as long as I did that a few times a week we would cover sort of the most important subjects now it's getting collapse to the point where I have to do multiple huge stories in one video just to kind of at least have a chance of being able to cover them and talk about them and I think that's what the singularity was supposed to be is when we get to a point where the progress is happening so fast that it's impossible to keep up and so you know maybe the singularities here is just not evenly distributed by the way if you're curious about who won in the battle between Steve Jobs and Bill Gates Linux comes in and just kills them all I thought You' knew and with that that's it for me my name is West rth and thank you for watching by the way this was like 11 years ago this video was like 11 years ago

Channel: Wes Roth

Views: 140,369

Keywords:

Id: UObruJtikrA

Length: 23min 20sec (1400 seconds)

Published: Fri Dec 01 2023