OpenAI Just Released o1 Early....

So it seemed like yesterday OpenAI made a genuine mistake by releasing the O1 model in full. Now, currently, if you guys are Familiar with the O1 model, it is the model that thinks before it responds. Now, currently, we don't have access to the O1 full model. The only model that we have had was the distilled model called O1 Preview. The reason this is a complete difference. You can see there's a complete difference between Zero1 Preview and. And of course Zero1. And you can see that across the board, Zero1 completely dominates what Zero1 Preview is able to do. So them releasing Zero1 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 use the model and some of the responses that I managed to get. So first we had this tweet called 01first contact. It said, brought to you by me and jcraft39. 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 ChatGPT 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 zero one type model that is able to think about images or 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 these simple bench is. 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 GPT4O and Claude 3.5 Sonnet fail at these questions. However, if we do test the new paradigm of models, we're starting to see 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 two meters 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 as 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, so he actually says above the blue ball number F, which is completely wrong. So we can see here that 01 passes, whereas when he tried it multiple times, O1 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, our 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 human like 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 O1 is a little bit more powerful than we did initially think. Then the next thing I did see on Twitter was the fact that Tibor Blaho, someone that regularly checks around the code in OpenAI websites and manages to find different leaks with, 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 that Sam Altman actually said that, you know, he expects rapid improvement in that area. How will vision capabilities scale with new inference time Paradigms set by O1? Without spoiling anything, I would expect rapid progress in image based models. 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. 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 01. Then we can also see is that the attachments type are multimodal and the accepted MIME type image types, whatever you want to call it is image png, webp give. 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 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 OpenAI 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 third thinking model and input an image. So then I put explain this. 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. It said, I thought about this image explanation for a couple of seconds. So this confirmed to me that they actually did release 01. 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, transformer, Discoder and yada 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 ChatGPT 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 O1's new image reasoning capabilities. One user called Anna GH was able to get the O1 model on the left and compare it with GPT4O 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 GPT4O 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 GPT4O 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 O1 does actually get this wrong. I'm not sure why that 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, 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 image capabilities, capabilities are behind the O1 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 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, which is, which is, you know, quite wrong. But of course some, you 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 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. It's pretty crazy because Sam Altman has been on a spree of teasing us with the O2 Model 01 model. You can see here that Sam Altman basically says that unleashed 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 Dev Day where he actually talks about the O1 image reasoning capabilities. I think image is something that is quite underrated in AI, but it is something that has a remarkable level of.