OpenAI Releases Smartest AI Ever & How To Use It

Okay, so OpenAI made a big move. They released a brand new model titled O1. This is the first model in a brand new series next to their GPTs and what it specializes in is reasoning. Okay, what does reasoning exactly mean? Well, I would define it as thinking about something for more than a few seconds. This definition works really well in using ChatGPT, as you'll see in a second, because this new model takes a bit of a different approach. We'll talk about that in a second. But first things first. How can you access this, who can access this and what are the limitations? Let's move over to the screen recording here. So first things first, this is available to all ChatGPT and Teams users. As I'm on a teams account, I already have access to O1 preview and O1 mini right here. There's big limitations on it though. If you go through ChatGPT, O01 preview, 30 messages per week. Okay, that's per week. So be careful with using this.01 mini, 50 messages per week. The API access is unlimited, but it only has been rolled out to people with $1,000 or more of spent. You need to be in the tier 5 category with OpenAI. So not everybody has the API access. So those are the base facts. Now look, let's talk about this reasoning word capability, okay? Because up until now we were familiar with models like GPT4O and a lot of competitors that kind of do the same thing. And if you tell it to write you an essay about penguins, it will just do that without really thinking about what makes a good essay about penguins. And it just does it. Now look, with prompts like this, I'm not saying the new model works better. I'm just saying it's going to spend some time thinking about the answer before it gives it to you. And this is nothing new. Anybody who's been following the channel closely or following the prompt engineering or large language model space, gen AI space closely over the past year or two will have ran into this technique called chain of thought. And chain of thought is easily described as a different way of prompting that includes a little bit more reasoning, a little bit more thinking. So you would just, just add think step by step and then you would get improved results on reasoning related tasks. Now, what are reasoning related tasks? Let's go a little deeper into this. They're tasks in the domains of science, math and coding. Okay? So something like this, like writing an essay is not going to be improved. This is not some magic bullet. You don't just switch over to this new model and it's amazing, but there's a lot of things that it's amazing at. My hope is that by the end of the video you will understand them a little more closely. So as it is super efficient and improved in these domains of and I'll have to briefly regenerate this as it is improved in the domains of science, coding and mathematics. What if you don't work in science, coding or mathematics? Does this even matter to you? Well, I would make the case that if you could master well or at least get to a decent level of understanding and application of math, coding and science without any effort, would it be useful to you? I think that's the real question to answer here because initially people will be like, oh, I don't need math in my everyday life. I don't need a math PhD in my everyday life. Well, but if you had one without any effort, no cost, would it be useful to you? And that's a question to ask here because these are the tasks that you can now perform. They claim it does PhD level mathematics and they have various benchmarks to back that up. The most interesting one being this graph, this comparison. So as you can see here on competition, math GPT4 all scored 13 questions. 13% of the problems GPT01 preview. The model that we have right now in the web interface, this one right here, the preview model scored 56%. And the full 01 model, which is still not full, not released, this is coming up in the future, 83.3%. I'll directly quote something from their blog article here. In a qualifying exam for the International Mathematic Olympiad IMO, GPT4O correctly solved only 13% of the problems. That's this right here. While the reasoning model scored 83. Okay, so this is not a little change. This is not an improvement in reasoning and thinking. This is a massive change. And there's also massive changes in how this processes your requests. Because if you write an essay like this, the result might be the same. But if you run this inside of the new model, it will take a whole lot longer. I prepared a few examples here in advance. And as you can see, this simple business plan here with a $2,000 budget, it took it nine seconds to answer that. If you run that through GPT4. Oh, it just starts generating right away. This was thinking for nine seconds and then it started generating because it actually goes ahead and creates a little bit of a plan. Multi step reasoning. This is really a big step in this agentic future that we've been Talking about on this channel, we. Quite a bit. Because before you execute something, you think about it, right? Think about it just from a human perspective. This is. I think this is really the important point here. If you are told, hey, could you please help me prepare a business plan for a new T shirt brand with the name and slogan, where's the voice Assistant? Which, by the way, I still wonder. But let's say somebody tasked with this, you know, creating a business plan for a new T shirt brand with a $2,000 budget. Well, this is not the type of question that you would just come up with an answer to like this. This is the type of question where you would say, sure, I can do that. Let me get back to you. When do. By when do you need it? Right? Implying that there is multiple steps in thinking this through. You have to think about the marketing plan, you have to think about the finances. You have to break it down. What do we need? Do some research, then think about that research, let it sink in, then come up with a business plan. Right? But from GPT, as our assistant this far, we always expected it to just give us the answer right away. And then we were disappointed when the answer was not so good. And fair enough, we're early on in the development of this technology, but this changes now. And I also want to make the point that this is not just chain of thought built into it. This is not just them saying things step by step at the end of every prompt. Because if that was the case, then we could have solved some of these harder problems early on. Some of these problems that require a little more reasoning, and that wasn't the case. I can give you a good example here. I played around with it for a bit here. This palindrome example is a really good one because GPT4 was just terrible at this. Palindrome, if you're not familiar, is if the phrase is the same from front to back and back to front. So here it just said, create a palindrome related to cats or fats. And then it fought for 18 seconds with all of these different steps in between. And then they gave me, I saw a cat a hat attack. I was. I. Now that doesn't really, you know, make sense. So I said, okay, but make it like a common sentence for common people. And then it fought for 31 seconds and said, okay, here's a palindrome that reads like a common sentence. Was it a cat I saw in front? Back to front? Was it a cat I saw? This is awesome. But it excludes the hat, then it goes ahead and admits that, okay, I couldn't really fit the hat in meaningfully. But there you go. This one incorporates both. So it's at least self aware. If you ask this and throw this into GPT4O, you will be very disappointed because it will do something. Eva, can I see bees in a cave? Yeah, there you go. This has nothing to do with what I asked. Right? So that was initially very interesting. And look, this is not really coding science or mathematics. Right? But this is also, on the other hand, this is not something that you would use in your everyday life. I would, I would argue. Now let's talk about that here in a second. Like how can you actually use this model in your everyday life? But before that, I really do want to show you some of these change capabilities here because another interesting one is translation. Translation. Let me tell you, as somebody who speaks three languages at the level of English, it's just something that has been, you know, granted to me and like my life has evolved in a way where I didn't really have a choice but to learn all three of them. But I speak English, Slovak and German in no particular order, all at this level of both reading. I suppose my Slovak writing is way worse than my English, but essentially I'm trilingual. And let me tell you, when you translate things back and forth, there's a lot of complexity involved and there's a lot of context involved and certain things only make sense in a certain language and cannot really be translated word by word. So there's actually this one phrase from German that I just threw in here right away. And I was kind of blown away by the result because it's a ridiculous phrase. It's a phrase where people often make fun of German for having phrases like this, but German is a bit crazy like that. It has a lot of very peculiar phrases that are commonly used. And German speakers don't even realize that as long as you don't context switch it to other languages. But it's this one. It's davita hunde vaneferrukt. It's what this is, what this literally translates to is that's when the dog goes crazy in the frying pan. Okay, so it's a completely nonsensical and weird phrase, but it is a phrase that is used in German and it's hard to translate. Right? But GPT401 preview actually went in and started looking at idioms in Hindi and because it was curious to see if mixing literal and idiomatic translations works best. And then it came up with well, I'll be darned. And let me tell you, this is perfect. This is how this phrase is actually used in German. Now let me just see what we would get from GPT4O without doing these extra reasoning, these extra thinking steps. But my guess would be that this is not going to go as well. That's unbelievable. That's crazy. Almost. Almost, but not quite. Let me tell you, not quite. I'll be darned. Well, well, I'll be darned is just as German speaker just feels more accurate. And also you don't get this long answer, you get this short concise answer. So look, these are two simple little use cases that are completely unrelated to science, coding and math, which is is what they advertise it as. They say this model is for advanced reasoning and it will be useful for science, coding and math. But I would challenge that a little bit because these are just my first impressions here. Okay, But I also did a thing here in the background where I actually went ahead and I ran the same prompt twice. And this would be my argument here why I think we're going to find over the next days and weeks that this model is way more useful for than what it might seem like initially because it's interesting. But I don't think we've seen any like game changing use cases as of yet. But I think with prompts into this direction, look, I carefully compared the results. It's not night and day, but there is a difference. So create a business plan for a new T shirt brand with the name and slogan where's the voice Assistant with a $2,000 budget GPT4.0 on the left, GPT01 preview on the right. Okay, there we go. Now this is the part that I was curious about because this is so good at math, it's good at money. And being good with money is sort of a useful thing to people outside of the sciences, right? Or coding or math. So I gave it this prompt and let me tell you, I was actually extremely surprised by 4.0. They must have changed something under the hood because look, it includes a table, it does the calculations, right? This is 4.0. This is not 4.0 preview, the new one, right? This is the original, the model we had. But something must have changed under the hood because let me tell you, a few months ago this result would have not been accurate. It does the math correctly here. This adds up to 2k. Then it gives you a budget and cost estimates for everything. And in the end if you scroll all the way down. I'm sorry if I'm scrolling a bit too fast here, but I believe I had a follow up prompt here. Yeah, it gives you a conclusion and it's decent. It's good. Now look, when I ran this through the new model, it's slightly better. I would say it's slightly better. It gives me different. No, this was actually my follow up prompt. Apologies for that. But it is slightly better at the estimations, at the calculations. It is better. But 4o was precise and that was the surprising part to me. So yeah, if I just look at these conclusions, I think this is more accurate and I would consider this higher quality myself. Also the overview and all the processing is structured better. I really dislike the fact that here it really gives you too much detail in the, in the various steps and here it kind of breaks down what you need. And then I followed up with this prompt over here. Check this out. What is the optimal level of spend to launch this brand? Okay, it's kind of a tricky follow up prompt because like, what do I mean by optimal and what does it take to launch this brand that has to figure all that out? Well, this model fought for 10 seconds and then it gave me an answer and 4.0 just gave me an answer right away and it kind of landed on, you know, you need 5 to 10k would be the sweet spot. But that is sort of an estimate. I kind of like this conclusion over here more. That gives me, hey, these are your free options and here's a summary of why that would be good. I just, I just personally think this is a higher quality answer. So I think my first intuition, and I will test this more extensively in a report back, but my intuition is that whenever money is involved, this is going to be more useful. And I don't think that's a trivial conclusion. I don't think that's a little prompting trick that you might or might not use. Like heck, whenever you need to do financial calculations, just use this new model and let it think about it for a second. Whenever you do something more complex, like a marketing plan, like a business plan, maybe something that would take a little more thinking. Think about it in the way I presented at the beginning. If a human would need to think more than 3 seconds about it or 10 seconds, then your AI assistant most likely will too. And I have one more thing I could go into many interesting kind of like facts and many interesting points from the presentation here. I think particularly, I think particularly this one video over here was very interesting. I would recommend to Check this out from OpenAI. This is. They basically talk about how they built it and there was this one interesting fact where they tried to map human thinking and then use the ways human think inside of the model and then they tried letting AI do the same thing and that worked way better. So I thought that was just interesting. But beyond that, what's the point here and what's the conclusion here? So look, I think for everyday users, is this going to be useful? We shall see. I'm not exactly sure. I think I see a slight improvement in these money related prompts. But let me tell you, in my everyday tasks, even with ChatGPT, I don't use many science or math related things. I do use a bit of coding, but if you don't have coding skills then you know the improved coding ability of this, which now supposedly is better than Sonnet 3.5 and which was state of the art right now for code generation. So for code generation, apparently this is best in class. I'll have to do some further testing to confirm that. But if you're not, My point is if you're not using one of these, one of these, if you don't work in science or, or math or co generation, then it's questionable if this will be useful. But beyond just thinking about this in a way, where do I need more than a few seconds to answer this? Then I use the model. Or would a human need more a few seconds than I need the model? I would also bring up a second point and I would like to round it out on this and a few prompting tips to it that I found. Well, up until now with ChatGPT, you really had an AI assistant. You threw it a task and the AI assistant did what it could. It presented you with the findings and if it didn't have some context because you didn't provide it or because it's not available in the training data, it just hallucinated it and it just did everything it could to give you an answer. Well, with GPT01 Preview, it's more like you have multiple AI assistants. You spawn three of them and they talk to each other before they talk back at you. Okay, so originally you talking to an assistant, assistant doing what he or she can and replying. Now you telling what you want and multiple assistants showing up behind you, them talking to each other, figuring it out. That's what this thinking would be, that's what these multiple steps would be and then they reply to you. Look at that, 31 seconds of thinking of how to rearrange this palindrome to make sense. Okay, So I think that's a really good way of thinking about it. And I'll just leave you with some final prompting tips that were discovered by Cognition Labs over here. Okay, so Cognition, the builder of Devon, you might be familiar. This was the first kind of like AI agent preview. We still don't have it, but they're saying with this model, it works way better than anything before. So that's going to be interesting. But. But they have some prompting tips here. So first of all, never again tell this model to think step by step or think out loud. It works better if you don't. Okay, so that's one thing, because the model is trained on this. It's set up to behave that way. Your AI assistants are set up to talk to each other. Secondly, if you have more clutter in your prompt, it will perform worse. So traditional prompting, as they say here, often even in, you know, Sam, the prompt creator, the free tool that I made accessible and built with the AI Advantage community, that thing really fleshes out the prompts and refreshes certain instructions and gives them multiple times and reinforces them because that's how traditional prompting works. This prompts a little differently. Less is more. In the past, I referred to this as goal based prompting rather than, you know, fleshing out all of the context. Whereas goal babes prompting is better for agents and this is this agentic direction. So it's better for this. What's the difference between normal prompting and goal based prompting? Well, normal prompting would be like, write me like you are an expert in, you know, children's literature. Write me a children's book about, you know, topic xyz with this turn and this story arc and, you know, make it, you know, the 10 pages, whatever. I just invented all that. That would be normal prompting. Second one would be write the best children's book possible for specified target audience. And then the multiple agents, the multiple assistants, they figure it out amongst themselves. It's a different approach to prompting. And as you can see, the second prompt is way shorter. And this has been the case with all agentic experiences so far, experiments. And it seems to prevail. So keep your prompts shorter and simpler. And I would say prompt based on goal rather than based on every single little detail. You can always add the details in follow up prompts. I mean, you know, too bad that we only get 30 messages per week, but that's why we have channels like this. We'll be experimenting with the API and reporting back. But one more thing. Is that it doesn't have tools. That's the last thing I wanted to point out in this video. Okay, but it's on the roadmap, so it doesn't have code interpreter, it doesn't have web browsing, it doesn't have image generation, it doesn't have image upload, none of these. But they did mention that it's coming along with eventually ChatGPT just auto selecting the models and the tools and everything. So right now you have to watch videos like this to have the information of hey, should I be using 4.0? Should I be using 400 mini? Should I be using 01? Should I be using a competitor's model? In the future, it's going to figure it out on its own. You're going to give it a goal and it's going to look at all the tools, it's going to have full understanding and it's just going to make the decisions. Hey, use this over here, use that. Here's your result. User. That's it. That's the future of this technology. This is a major step towards it. Check. Check out the different videos on OpenAI's channel, for example, and follow this channel. Hit subscribe to hear more about this and how to use it in the real world. Because I think there's some hidden. There's some hidden. There must be some hidden gems in here and I'll do my best to discover them. All right, I'll see you soon.