Dr. Zsolnai walks through OpenAI's O3-mini, a free reasoning model that outperforms the previous flagship. It handles complex tasks like physics simulations with 100 balls and Minecraft world generation. GPT-4O beat expert humans in biology experiments, and O3-mini is available free with usage limits.
OpenAI's O3 mini AI is here. This is the Reasoning and Thinking AI. And I waited a little before publishing this video to see how you fellow scholars are using it. Which is not very good for business and views, but it's better information for you fellow scholars and that's what matters. So we do that. And I have to say the results are incredible. The classic bouncing ball experiment that could not be reliably done with the previous OpenAI01 system could be done with the free Deepseek R1 that you can run at home self hosted. So what is OpenAI's answer to that? Let's see. Oh yes, this one can finally do that. But now let's dream bigger. Let's graduate to 10 balls. Still works. Fantastic. So can we do it? Can we ask for even a hundred balls? Let's see. It works beautifully. Loving it. It can also create better, more interesting Minecraft worlds compared to the previous vers. There is quite a difference. And if you want to specify an input image and write a program that creates this beautiful ASCII character art from it, you can do that too in just one prompt. And get this, it is now available for free for all of us. You can hopefully also try it by clicking on the reason button when it appears. If you're interested, you should probably try it right now. But the good news continue. So I hear you asking, do we have a paper? Oh yes. Fantastic. And it has some really surprising results there too. More on that in a moment. I love looking at these comparisons. For instance, this falling ladder experiment shows that other systems have also caught up and are performing really well too. So is O3 mini any good? And the answer is yes. When looking at benchmarks, things are going up and to the right in an almost unprecedented manner. I mean, just look at that. But it gets better paper time. Wow. I don't even know which the crazier news is. The paper reveals that 1, even the older GPT4O is better than the baseline expert humans in biology experimentation. Or that even the mini version of O3 outperforms the previous flagship thinking AI and it is so cheap to run that it is now free to all users with some limits. Yes, all of us. Super cool. Note that we have no business relationship with OpenAI. Never had. And there is an even bigger and more surprising insight in this story. For many, many years we thought that whichever company gets first to something that kind of looks like a generally intelligent AI system is going to be unbeatable, but today it does not seem to be so Completely open source solutions that you can run yourself at home for free are popping up daily by day and for the first time ever, they are able to challenge OpenAI. That is absolutely amazing for us. Just think about it. O1 took just a few months to reproduce and that thinking AI was a paradigm shift. But O3 is not a paradigm shift, it is a better O1. So I think it will be challenged by a completely free variant in less than three months, possibly even less than a month. So how is all this possible? Well, all this is possible. Possible through the power of the papers and open sourcing and open science. People are able to work together all around the world and share their findings with each other. And they put together all their smarts to create something even better. That is humanity at its best. And as a result, we are all going to have intelligence in our pockets for free. Loving it. And that is what we have been talking about here for almost 10 years now. And nearly a thousand videos. And now suddenly everyone is talking about the papers. What a huge win. That makes me very, very happy. So that is the amazing world we live in today. What a time to be alive. Also note that in the system card areas of risk are reported and it's not just sunshine and rainbows. I like how these are reported in such a transparent manner. Respect for that. Also, if OpenAIS01 could become a formidable con artist, I don't know what O3 is because it is so much stronger in that area. I hope that this can and will be used as a shield against this seemingly never ending line of people who have coins and riches to offer for anyone who believes them. Using AI as a shield. Somebody please work on that too. So, conclusions. AI research keeps progressing at an absolutely stunning pace. These systems are getting more and more intelligent. They will be able to help you build a billion dollar company completely alone. And all this is going to run in your pocket for free. I mean, just as I started making this video, they also came up with deep research which improves the already improved benchmark results on a proper data set, part of which is private. So hopefully it cannot be gamed as easily. And it not only improved it, it doubled it. My goodness. And look, not a few weeks, but a few hours later, an open source clone has already appeared. What a time to be alive. Dear fellow scholars, this is two minute papers with Dr. Zsolnai. This is where we celebrate amazing papers and amazing human achievement together. And one more thing, every now and then we are able to invite new sponsors to the show. This time is that time if you are interested, let us know. The link is in the video description. We need new tools for the era of LLMs and weights and biases. Now has weave a lightweight toolkit to confidently iterate on LLM applications, use traces to debug how data flows through each step of your app, and use evaluations to measure your progress. It is the best. Try it out now at WNB Me papersllm or click the link in the description below.
