Dr. Zsolnai demos OpenAI's O1 model, which reasons step-by-step better than GPT-4 but knows less factual content. It solves math systems and generates working code, available to paid subscribers with weekly limits.
OpenAI has just unveiled their new AI assistant called O1. And this is what we are going to talk about. Holy mother of papers. Or wait a minute. Breakthrough performance in some areas and surprisingly worse at others. How is that possible? Is this even good? Well, let me try to explain. Dear fellow scholars, this is two minute papers with Dr. Zsolnai. You see, the previous GPT 4:08 is not that smart. However, it knows about almost everything. It read nearly the whole Internet and The new model O1 does not know a great deal. However, it is able to reason to learn from very little data and it needs some time to think. Look at the chain of thought here. Goodness, it never ends. So what happened here? Let's see. GPT4O, the previous technique was given a ciphertext and it is trying vigorously, but ultimately we don't get a result. However, the new technique, after careful deliberation. Look at that. There are three R's in Strawberry. You learn something new every day. And when given a crossword puzzle, the previous technique shows some signs of life. But ultimately this is a big mess. This is especially difficult. Why? Well, because as you find one answer, you get more information about another question. So yes, this requires reasoning a little more thought. So can the new one do it? Oh my. Now we're talking. Absolutely fantastic work. So that is why it is better at some things and worse at others. It knows less, but can reason better. This is my absolute dream. Just imagine being a teacher, which student would you want to work with? One who read all the books but can never apply that knowledge to anything truly new. So basically a book or someone who is mega smart and just started learning the basics, but can derive the whole theory of pretty much any topic. So, Einstein in a box. That is what I want. And it is finally coming. Wow. So how much of an Einstein is it? Well, geneticists, quantum physicists and many other scientists are using it right now. Goodness. Note that this is also super interesting because it is a combination of neural networks and reinforcement learning. Two different ways of doing AI. And they are finally coming, coming together. I absolutely love this. And the resulting technique requires some time to think and reason. It thinks step by step. Is that good? Because we can ask previous techniques to do that too, but O1 is trained to do that from the ground up. This is essentially an AI implementation of the two modes of human thinking, and that is thinking fast and slow. Thinking fast is about quick instinctive responses, like reading something. But thinking slow involves deliberate, logical and calculated decision making. This can do both as well as, hmm, how well Exactly. Well, look at that. This insane jump in the GPQA dataset and more results tell me that in cases it can do it better than than some of the smartest humans. Einstein in a box. I am absolutely stunned because this is not just the next version of a capable model. No, no. This is a complete paradigm shift in AI. So what else can it do? Absolutely astounding things. Get this. If it would be a human and would compete at the International Olympiad in Informatics, it would have won a gold medal. Wow. Now let's look at this deceiving little problem. Hmm. If you taste the question a little, you feel that it tastes like a system of equations. However, if you feel like guessing, have a look at this. This is how you approach it. Once again, look at that beautiful chain of thought. It truly is a sight to behold. And at the end, it gives us not just one solution, but all possible solutions. Bravo. Let's do one more. You're going to love this one. Let's ask it to write a Snake game. Alright. So does it run? It even has a start screen and it works. My understanding is that this was the first try and if it was, that is very impressive, but we are not done yet. Now let's make it a bit more fun. Add some obstacles there. Well, here is the new code which we will scrutinize line by line. Just kidding. Let's play it. Oh, this is so cool. I would have loved to get the opportunity to give it a try and have it write up a beautiful physics simulation at the OpenAI lab. And I can't wait to give it a try. This is perhaps the first AI technique that finally has the capability to push research forward and find out truly new things. And I am super excited about that Einstein in a Box. So when can we try it? Well, it is hopefully available by now for all paid subscribers. Note that there are some weekly limits. Let me know in the comments what experiments you tried and how it went. I'd love to see you super smart fellow scholars starting to give it a spin. So for now, let the experiments begin. If you're looking for an Nvidia GPU cloud service for large scale AI projects, check out One Click Clusters by Lamda. Featuring up to 512 Nvidia H100 GPUs with Infinity Band networking. And hold on to your papers because with one click clusters you can access all these gpus on demand with reservations starting at just two weeks. Try it out@lambdalabs.com paper or click the link in the description.
