# OpenAIs New AI Research Is Incredible... ## Метаданные - **Канал:** TheAIGRID - **YouTube:** https://www.youtube.com/watch?v=ljstxRAfVgA - **Дата:** 04.11.2024 - **Длительность:** 8:03 - **Просмотры:** 26,026 ## Описание Prepare for AGI with me - https://www.skool.com/postagiprepardness 🐤 Follow Me on Twitter https://twitter.com/TheAiGrid 🌐 Checkout My website - https://theaigrid.com/ Links From Todays Video: https://openai.com/index/searchgpt-prototype/ https://www.youtube.com/watch?si=auhCGtYnYp4EXpSQ&v=xIoqmgpYHlw&feature=youtu.be https://www.reddit.com/r/aivideo/comments/1g83ndj/lifeskin_part_2/?share_id=mmb0dtyRBlB-RYpb6OOJx&utm_content=1&utm_medium=ios_app&utm_name=ioscss&utm_source=share&utm_term=1 LEMMiNO - Music https://www.youtube.com/watch?v=b0q5PR1xpA0 https://www.youtube.com/watch?v=xdwWCl_5x2s https://www.youtube.com/watch?v=rlaG7gF7qeI CC BY-SA 4.0 Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos. Was there anything i missed? (For Business Enquiries) contact@theaigrid.com #LLM #Largelanguagemodel #chatgpt #AI #ArtificialIntelligence #MachineLearning #DeepLearning #NeuralNetworks #Robotics #DataScience ## Содержание ### [0:00](https://www.youtube.com/watch?v=ljstxRAfVgA) Segment 1 (00:00 - 05:00) so we do have new research from open aai called simplifying stabilizing and scaling continuous time consistency models now ai image generation has made great strides but the current mainstream methods primarily diffusion models tend to be slow openi has developed a new more efficient and faster method called simplified continuous time consistency models this is a key achievement as scaling continuous time consistency models can generate high quality images in just two steps compared to the dozens or even hundreds of steps required by previous methods thus speeding up the entire process now if you don't understand how traditional AI image generation works these traditional AI image generators rely on something called diffusion models so imagine these models as starting with chaotic noisy image and slowly cleaning it up over time SL over many stages until it becomes a recognizable picture and this process takes anywhere from 500 to 100 steps to generate an image making it computationally expensive and too slow for real-time applications that I've got a new solution SCM now SCM changes the game by taking giant leaps instead of many small steps requiring just two steps to create an image and this actually makes the model approximately wait for a guys 50 times faster than traditional methods for instance it can generate an image in around 0. 11 seconds on Specialized Hardware I don't know about you guys but generating images in 0. 1 seconds 50 times faster is absolutely incredible now if we want to break down this technological achievement we can look at the scale the system is large containing 1. 5 billion parameters capable of working with high resolution images of up to 512 by 512 pixels now of course there is a speed versus quality trade-off typically a faster process results in lower quality but SCM manages to maintain high quality images despite being much faster and moreover it requires less than 10% of the computational power of older models now if you want to know how this actually works instead of the traditional method which is where the model starts with random noise and iteratively removes the noise like unscrambling a jumbled puzzled one piece at a time the new method SCM instead of gradual noise removal SCM takes a shortcut jumping directly from noise to the final image and this is akin to having a blueprint of the complete puzzle and assembling it very quickly so we've got the real world impact here guys of this is that the speed of SC M could enable realtime image generation making it possible to edit photos instantly apply Real Time video effects or quickly create images for apps and games and that's what I'm going to get into in a second now do you remember this research because when we take a look at SCM we understand that this thing is absolutely crazy it can generate images ridiculously fast but this brings me back to some research earlier this year that was developed by Google called Genie now if you don't remember Genie developed by Google Deep Mind is a generative interactive environment so imagine being able to create and explore entire virtual world just by giving simple prompts like text images or even sketches Genie can take these inputs and generate interactive environments that users can control and explore frame by frame which is quite similar to how one might play a video game it's like a genie that can make a new world for you hence the name now Genie okay and just bear with me for a second is a system that learns from over 200,000 hours of video data available online and the interesting part is that it learns without labels or instructions kind of like a child watching and understanding how things move and interact without explicitly being told what's happening now the craziest thing about this is that if we do take Genie and SCM together the simplified continuous time consistency model described earlier it's actually super fast at generating images taking only two steps to generate high quality results and genie on the other hand focuses on generating entire video like interactive environments what if we were to combine the speed of SC with Genie's ability to create interactive scenes we could potentially develop realtime immersive experiences for example we could get instant interactive World Imagine typing in a few words and a rich virtual world ### [5:00](https://www.youtube.com/watch?v=ljstxRAfVgA&t=300s) Segment 2 (05:00 - 08:00) appears almost instantly for you to explore or play in we could have you know realtime video game creation developers or even players could design levels or scenarios on the fly as they play like building their dream game or dream world in seconds or we could even have augmented reality with the speed of sem adding virtual objects or creatures into the world could become nearly instantaneous making AR more engaging and more fluid and if we managed to combine This research we could see a leap towards realtime virtual environments the users can interact with as easily as snapping a photo turning static image generation into Dynamic playable words but how we just discuss artificial reality what it's actually augmented reality take a look at what this Reddit user called Jessa 722 posted he actually shows the way that augmented reality could be in the future imagine you have goggles on and this is the kind of thing that you're able to see I think this kind of editing and this kind of you know visual display shows us exactly the kinds of things that should be available to us in the future I'm not sure if I would switch mine to horror I don't think I need to be any more scared as I already live in a pretty scary apartment so this is something that I think is pretty cool but what if we wanted to use these filters for something else this user on YouTube nikkan has actually shown how you can transform different video games into remarkably realistic footage you can see right here that GTA San Andreas a popular version of the GTA series actually looks like real life Graphics when prompted with Runway recently released some filters in which you can use them to basically change any kind of content into another style of content so if you want to do this yourself you can they do have a free plan but this is the kind of thing that these you know consistency models might be able to enable in the neartime future if we can get models that are able to you know produce tons of images at real time we could get Virtual Worlds virtual games augmented reality and much more now I pose this question to you if you do have this what will be your largest application for these realtime models will you be in AR in your own cust custom environment will you be playing your own custom game as your own custom character in its own custom world with your own custom AI filter and will you be generating multiple AI images just for whatever might please you I mean there's going to be a million different use cases in a million different areas and it's quite likely that some of them we won't be able to predict let me know what you guys think about opening eyes research I think it's really cool that they're now starting to be a little bit more open and I think this kind of shows us that open AI is probably once again in some areas still a lot far ahead --- *Источник: https://ekstraktznaniy.ru/video/13828*