# OpenAI Just Released o1 Early.... ## Метаданные - **Канал:** TheAIGRID - **YouTube:** https://www.youtube.com/watch?v=G3xY3ub6drI - **Дата:** 03.11.2024 - **Длительность:** 11:04 - **Просмотры:** 36,912 ## 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://x.com/tsarnick/status/1851910149423927647/video/1 https://x.com/btibor91/status/1852796199562096922/photo/1 https://x.com/legit_rumors/status/1852629240761426398/photo/1 https://x.com/btibor91/status/1852656441020092515 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=G3xY3ub6drI) Segment 1 (00:00 - 05:00) so it seemed like yesterday open AI made a genuine Mistake by releasing the 01 model in full now currently if you guys are familiar with the 01 model it is the model that thinks before it responds now currently we don't have access to the 01 full model the only model that we have had was the distilled model called 01 preview now the reason this is a complete difference you can see there's a complete difference between 01 preview and of course 01 and you can see that across the board 01 completely dominates what 01 preview is able to do so them releasing 01 is going to be a hugely significant upgrade in terms of the capabilities which is why I think they did it by mistake now for those of you who don't believe me I'm going to walk you guys through exactly what happened and then I'll show you guys how I used the model and some of the responses that I managed to get so first we had this tweet called 01 first Contact it said brought to you by me and jcraft 39 this was rather interesting because at the time I didn't understand what I was looking at but then I decided let me go ahead and check what this tweet is and this is where I started to see something really fascinating so we had a user input an image into chat GPT called you know just a simple chat of Bing which is a pretty simple conversation then he says describe this then of course I saw something even more interesting he said thought about image description for 7 Seconds so this is where we actually get confirmation that the full 01 model has image editing capabilities so you can see this user was able to interact with a 01 type model that is able to think about images for certain seconds now I did the exact same and I'm going to show you guys that in a moment but you're about to see that this was actually a completely different model now once he posted this he then posted something else where he further proceeded to test the model on different benchmarks now he said let's test it on a sample from the simple bench now if you don't know what the simple bench is a reasoning Benchmark created by AI explained and he wanted to create this Benchmark so that you could essentially have super simple questions which most humans would get right I think it's around 93% to 96% is what the average human would get but a lot of these advanced reasoning models like GPT 40 and clae 3. 5 Sonic fail at these questions however if we do test the new paradigm of models we're starting to see a huge Improvement of what the results are so the question here for this full 01 model is that a juggler throws a solid blue ball a meter into the air and then a solid purple ball of the same size 2 m into the air she then climbs to the top of a tall ladder carefully balancing a Yellow Balloon on her head where is the purple ball most likely now in relation to the blue ball the correct answer is at the same height as the blue ball and then the responses that we see are quite different so we can see that 01 the full model which is the one that I showed you guys the one that actually gets a lot more on the test results actually manages to get this right you can see it says at the same height as the blue ball and the user then says congrats but if we check this compared to the 01 preview you'll see that the answer is wrong so he actually says above the blue ball number F which is completely wrong so we can see here that 01 passes whereas you know when he tried it multiple times 01 preview consistently fails now we could also see that behind the scenes he actually managed to check which model he was using and you can see here that it says 01 are most capable model great for tasks that require creativity and advanced reasoning now this is pretty crazy because one of the things most people have speculated is that once these models start to get towards that upper bound of humanlike reasoning we're going to start to see some real exponentials in terms of AI and the fact that now this model gets this question right is going to be something that indicates to us that 01 is a little bit more powerful than we did initially think then the next thing I did see on Twitter was the fact that T or blaho someone that regularly checks around the code in open AI websites and manages to find different leaks was able to see that there was actually a new capability in 01 that is of course image analysis now this is really interesting because it was only a few days ago the Sam mman actually said that you know he expects rapid Improvement in that area how will Vision capabilities scale with new inference time Paradigm set by 01 uh ### [5:00](https://www.youtube.com/watch?v=G3xY3ub6drI&t=300s) Segment 2 (05:00 - 10:00) without spoiling anything I would expect rapid progress in image based on it's a bit of a carot down okay so that means that maybe just maybe we might even be getting image analysis potentially next week I know that would be insanely quickly but it looks like they've already managed to deploy this so when you look at the code here we can see that this says our most capable model great for creativity and advanced reasoning you can see the tag is A1 then we can also see is that the attachments type on multimodal and the accepted mime type SL image types whatever you want to call it is image PNG webp GI so we can see it's able to analyze these images now I wanted to go ahead and test this myself at the time I was still a little bit skeptical because I know that sometimes rumors do go around on Twitter and they're pretty see Hefty rumors so you do have to watch out for a lot of fakes but I input the link that they said it was actually an open AI link and I was so skeptical that I input the first message which is what are you then of course it says you know was thinking for a few seconds and it didn't actually tell me which model it was and I was like okay but I wanted to see if I could interact with a thinking model and input an image so then I put explain this is an image of a vision Transformer which is for another video I was working on so I decided that if it could accept this image then it's quite likely that the individuals were telling the truth and there we could see it said I thought about this image explanation for a couple of seconds so this confirmed to me that they actually did release o1 and I'm not sure how this even happened but there was a link where I could just simply click and talk with this model so you can see right here that it says I'm exploring the vision Transformers architecture with a diagram showing patch embeddings class embeddings the Transformer tooda and yada yada and you can see if I scroll down here we can see an entire description of exactly what this Vision Transformer is now I'm pretty sure that if I compared this to a chat GPT response this seems to be a little bit more detailed than what you'd initially traditionally get from those models I'm not entirely good I'm not entirely sure with as to how good these Vision models are but one thing I do want to show you is someone that managed to do side by-side testing of 0's New Image reasoning capabilities one user called Anna GH was able to get the 01 model on the left and compare it with GPT 40 reasoning on the right this was something that I wish I had did if at the time I realized but I was busy making a video so he asked it this popular question that has been on the internet for quite some time which is how many triangles are in this photo and he also asked the same question to GPT 40 side by side once that image is put in we can then see that these models are reasoning about these images now if you are curious about this answer the correct answer for this is 24 these are all of the possible triangle combinations and I just want to show you that before we get into the actual responses from the models and then we could see here that GPT 40 manages to reason quickly and gets the answer wrong it gives the output that there is 19 however on the left hand side we do get 01 analyzing the picture piecing together the puzzle and identifying many different things in this image now interestingly enough unfortunately I believe the 01 does actually get this wrong I'm not sure why it is but I'm guessing that this question is a lot harder than most people think but I think one of the cool things that we do get to see from this video is the fact that if we look exactly at this area you can see all of the different things that it is able to do when it's analyzing the image you can see it says breaking down the triangles analyzing the pyramid examining the patterns decoding the figure breaking down the process I mean there's a million different things that it is doing with its image capabilities and I'm not sure what kind of you know image capabilities are behind the old one model but it's clear that they're pretty Advanced and I think it was you know thinking for about like a minute 30 seconds before we even got a simple answer for the actual model response and then you can see after that we then managed to get a response there and it says how many triangles with the answer being 24 25 or 27 and you know eventually we do get a response and it does actually give the response of 27 which is uh which is you know quite wrong but of course some you ### [10:00](https://www.youtube.com/watch?v=G3xY3ub6drI&t=600s) Segment 3 (10:00 - 11:00) know it actually says that you know the frequently cited number for such puzzles is 27 so I think it just cites the commonly referenced answer even though like in its you know thought process it managed to get right which is kind of weird but um I think what this shows us is that there is Advanced reasoning capabilities when it comes to images now the craziest thing about all of this you know being released like I think potentially you know a day early a week early is pretty crazy because Sam mman has been on a spree of teasing us with the O2 model sl01 model you can see here that Sam Alman basically says that unleash the full 01 and of course he says that it's not that much longer hopefully so it seems that you know potentially next week it's quite likely that we will get the 01 model because of course he responded to someone saying not that much longer and of course there was that interview clip from devday where he actually talks about the 01 image reasoning capabilities I think image is something that is quite underrated in AI but it is something that has a remarkable level of --- *Источник: https://ekstraktznaniy.ru/video/13836*