The World isn't Ready for AI this Capable.. Dive into Open AI o3 mini & Deep Research

Folks, the AI race has a new goal post a new target for all of the open source developers and other companies to hit. And of course this new target is brought to us by OpenAI. Yep, they are king of the hill once again releasing the best models that are the most cost efficient and benchmarking higher than a lot higher than anything else we've seen before. Guys, we are talking about significant improvements here. The this is the takeoff effect of chain of thought and inferencing becoming a new scaling law for large language models. I also think that for OpenAI, this has been a little bit of a reaction to the deep seq R1 models that supposedly kind of took the market down in the US tech space. Last week. OpenAI just couldn't stand to be stomped on by some lowly open source competitor, so they upped their game. This, this is why I say competition is great for us all. In this video, we're going to be playing around with these brand new models from OpenAI, testing them out a little bit. We're going to dive into the blog posts, some community reactions, and of course the benchmarks, which are just absurd. The best model we're taking a look at today is almost 10 times better than GPT 4 omni. 10 times. I mean it's just absurd. The power potential of AI is unbelievable. Let's dive in. First up, let's talk about what was announced from OpenAI January 31. We saw them drop oath. So this was Last Friday and O3 Mini comes in two model variations inside of Chat GPT. We'll talk about how you can access them and what they're capable of in a little bit. On February 2nd, which was yesterday, they introduced Deep Research and I personally think Deep Research is the most exciting thing they announced as this particular model was the one that scored that 10x or nearly 10x benchmark over GPT4 omni. Let's start off with O3 mini. Obviously guys, O3 mini is a successor to the O1 series of models. Back in December, OpenAI teased O3 and what it was capable of, but now it's accessible in chat GPT and on the API. O3 mini is a smaller model in terms of parameter count, which means it doesn't use that much compute, but it does chain of thought thinking and it does it better than 01 while being more cost effective. It has better STEM capabilities, most notably. So that would be science, math, coding and as I said, it's also available in the API and it's also their first small reasoning model that Supports all the developer features so people can start building with it right away. So that's function calling, structured outputs and developer messages. Right now the O3 models are available to plus users, team users and pro users. Enterprise users are getting access. In February, we've also got an upgrade on the message cap, the daily message cap. So it's going from 50 messages with these O series thinking models to 150 messages per per day. Obviously due to the increased cost efficiency. O3 mini also works out of the box with search, but does not work with image recognition. I'm a big fan of image recognition. It clearly has the capability being omni an O series model to do image recognition, but for now they're keeping it out of our hands and we don't really know why. Free plan users can also try O3 mini by selecting the reason option in the message composer. So I like that they are still giving this cutting edge model to free users to test out, even if it is pretty limited. I also can't help but feel that they wouldn't be giving free access to O3 mini without something like DeepSeek R1 existing, providing significant competition because that model can be accessed for free in a wide variety of different websites. All right, let's take a quick peek at some benchmarks. We've got 01 mini and then the full 01. These are the previous models. Oh three mini low. So the smallest size here is still a little bit under O1 mini and definitely quite a bit under 01. But next up with O3 mini medium size, it benches quite a bit higher than 01 mini and is nearly at the same capability level as 01 now. 03 mini high. Here is the best model out of this new series, giving us a new high score of 87.3. And this benchmark in specific is competition Math. So that's AIM 2024 and we're really getting close again to 100 accuracy, meaning these models are going to surpass the benchmark we were testing them on, which is crazy. The development rate of this is insane. We went from 01 mini getting a D ranged score 63.6 if we use like an American school grading scale to 0.3 mini high getting a B plus 87.3. What do you think full O3 is going to get? Probably well into the 90s. Wouldn't be surprised if it was an A plus. Moving on to PhD level science questions. This is GPQA diamond, very similar story. These models are much more leveled off though in general with 0.3 mini high taking the cake. But still we see both 03, mini low and medium come very close to the original 01. For coding though, things change a little bit. The 03 series of models in general just slaughters in terms of coding capability. Previous 01 got a 1891 ELO score. Oh three mini low trails just behind it with 1831. But then 03, mini, medium and high stomp all over. Original 01 getting scores pretty deep into the 2000s territory. Now there are a ton of benchmarks on here and we've got a ton to cover today. So I'm obviously going to link everything below if you want to take a closer look. In general, the models are smarter for their size and faster for their size. Alright, now let's talk about Deep research. Now Deep Research, it's not necessarily a model in itself. It's more of an agentic framework as a whole because it's not only using LLMs, but it's also using other tools to be able to search the web. And basically as they describe it, it's an agent that uses reasoning to synthesize large amounts of online information and complete multi step research tasks for you. Now this one is only available to pro users right now, which is very disappointing. I'm a Plus user currently, as you guys know. Chat GPT Pro is a staggering $200 a month in my opinion. I mean it's just an absurd price tag. Plus is only 20 bucks a month. I find it pretty disappointing that plus users don't have any access to Deep Research. Like we can't even try it without spending 200 bucks. I don't want to say it OpenAI but kind of a dick move. They provide some quick footage showing how the model works. You click Deep Research, you give it a pretty detailed prompt and it goes to town. For this specific one, we're looking for a research report on how the retail industry has changed in the past three years. It actually goes ahead and asks a follow up question which is very rare for a large language model to do in my experience. Anyways, we say we want a global perspective on the retail industry and it gives an overview of what it's about to do and then a loading bar will appear on the screen. It says it's starting its research and it's going to take, you know, anywhere from a few minutes to, I've heard it goes up to 10 minutes in some cases doing all this different research and it gives you a full detailed breakdown report. Now this is the model folks, where this is the framework that was able to achieve an nearly 10 times higher score than GPT4 omni on one of the most challenging AI benchmarks, which is crazy. It's absolutely crazy to think about. We're moving away from necessarily training larger and larger models, but now we're moving into chain of thought thinking. We're moving into deeper research utilizing the capabilities of the LLM much more. The power potential was already there in these models previously, but now we're hooking them up and giving them the right tools necessary to be better. If that makes sense. It can go through hundreds of online sources to give your comprehensive report. They claim it's at the level of a research analyst, and from the examples I've seen so far, that's absolutely true. Now, this framework under the hood is powered by the full big bad O3 model. I assume it's an O3 preview. I don't think they're totally done with O3 just yet, but this one's also more optimized for web browsing and data analysis. Apparently it also does have the capability not to just analyze text though, but also images as well as PDFs. I really like this statement in the blog post as well. The ability to synthesize knowledge is a prerequisite for creating new knowledge. For this reason, Deep Research marks a significant step forward towards their broader goal of developing AGI, which of course is that super loose definition. But OpenAI does kind of give us a benchmark for AGI here. They claim it should be able to produce novel scientific research. New theories actually probably go out and test those theories for themselves and help advance society as a whole. All right, let's talk about some comparisons. As we can see, the comparison is directly with GPT4O. When you're using something like Deep Research, you're asking very complicated questions. You know, this isn't help me make a meal plan for the rest of the week. This is help me find iOS and Android adoption rates, percentage who want to learn another language, and you know, all this other stuff. It's a very specific prompt. Now, to GPT4O's credit, it does give you an answer. You know, a very generalized answer as well as some insights and then finally some recommendations. It's not a bad response, but Deep Research's response is just on another level. It gives us a full table breakdown. Country market share for both iOS and Android, interest in new language by percentage of population very specific and the mobile penetration change over 10 years. So 2013 to 2023. Like I said, we've got so much to cover. So I can't exactly break this example down all the way for you, but it's a lot more useful than the GPT4O answer. Obviously we've got real data here, real insights into specific markets in terms of the changing growth rate between iOS and Android. And the recommendation section is far longer. Based on the above, prime targets for a new ChatGPT powered iOS translation app would be countries with large iOS user bases and strong language learning demand, notably Japan, the us, uk, Canada and Australia as top choices for this app. So that's where you would want to market to. This would be someone's job this. How long would this take a traditional data analys to come up with and then provide this result? I mean it take hours, if not possibly days to gather all this information and get to these insights and this thing can do it in minutes, Minutes. It's on another level, this is just one in in terms of the business sector, we've got a needle in a haystack. This prompt is basically asking for a specific Moment in a TV show, the user can't remember what TV show it's from. GPT4O gives an answer that is just straight up incorrect. Where Deep Research gives us exactly what we're looking for, even down to the moment of the TV show that was specifically mentioned. So its ability to go out on the web and grab the information you're looking for, not having to rely on its database or its training data, I should say. I mean, it's here, it's here. It can really use the Internet better than a lot of you guys probably would, at least in terms of gathering information. We've also got examples for medical research like, I mean, this stuff is all beyond me. Study the patterns and the changes to the protein and corresponding results across the papers and list the top three domains that scientists modify to increase efficiency and why they believe these changes are effective. I'm not a, I'm not a medical research person. I'm just trying to give you guys a good look at the capabilities. But based on the response from Deep Research here, look at how long. This is absolutely absurd. This is the longest by far generation I've ever seen come out of a large language model. But to also have this information be accurate and based on sources, totally insane. Now here's something that's more for.