ChatGPT can now generate playable games, run physics simulations, and reproduce PhD-level astrophysics code in minutes. The demo shows it building 2D and 3D games, managing 25 agents at once, and solving constraint puzzles through chain-of-thought reasoning.
Project Strawberry
25
One hour
35 seconds
120
OpenAI's new ChatGPT AI01 was previously called Project Strawberry. They say that it is much smarter than a regular StrawBerry and has PhD level knowledge, can write video games, solve tricky brain teasers and achieve world peace. Well, maybe not that one, but we will make sure to have a look at all the rest and find out if it should be called Dr. AI or not. So what can it do? Well, it can quickly write up a snake game in 2D. But now how about a Snake game but instead in 3D? Let's see. Just one prompt goes in and look at that. This is the game as it came out on the first try. Goodness. This not only seems just fine, but I see a little polish already. For instance, it also checks when you hit an obstacle and knows when the game is over. Now more games. It can also write you a chess game real quick, but with a twist. Look what is happening here? Who is he playing against? Well, against an AI. Wait, so this is an AI that just programmed another AI? That is, I think, insanity. What a time to be alive. So what about the two minute papers special? And that is of course physics simulations? Well, it can't create these incredibly sophisticated systems that we talk about here. However, a rudimentary little program. Not a problem. I will tell you about a PhD level astrophysics example that is a hundred times more difficult in a moment and we'll see if it can do it. A previous research work made a video game where 25 of these AIs were given an identity and they organized a Valentine's Day party. Or they made a video game. Yes, a video game inside a video game. It's like Inception, but nerdier. How deep can this go? Now imagine putting 25 of these brainy little AIs in a video game and and getting them to work together on something amazing. Or maybe they would just start to argue whose code is better. The AI version of a family reunion. Good times. So how does all this happen? Well, this is an AI that is given a little more time to think when asking for a futuristic Tron type game. It gives you code, code, code. Okay, but not everyone is a programmer. And look here, it has also attached the thought process to help out. And from this a playable game comes out that you can play right now if you stop this video here and check out the video description for the link. It even has camera controls so you can see yourself better as you lose. Excellent. And the key here is that this is an AI that can finally do planning. For instance, when we ask a paradoxical question. Like how many words there are in the answer to this prompt. Not in our prompt, but in your answer. Tricky. Hmm. It starts thinking, recognizes the severity of the situation. Don't fall into the trap, Little AI. And ha. Nailed it. Perfect. This planning can also help it write a poem. But not any poem, but one where the first letters of odd lines spell this and more. It is a nice little brain teaser with tons of constraints. Then it thought, for how long? Only 35 seconds. Nice. And it was able to create that poem. And now, as promised, some say that it has PhD level knowledge. But does it? Because if it does, we should not just test it on simple questions like these. Let's do the real deal. Oh yes, Here is a PhD thesis on astrophysics on how to measure black hole masses. And now hold on to your papers, fellow scholars, because the scientist who wrote it says that there is a piece of code for an experiment that took him an entire year to write. And ChatGPT was able to accomplish that in. Are you serious? Get this. It did all that in one hour? Wow. By the time the next version drops, we may have to give it tenure. Okay, so how smart is it? What are its limits? Well, there are IQ tests out there for humans. And don't forget these AI assistants can see now. So can they take a test? Yes, they can. So how well did previous techniques do on some questions? Not so great would be an understatement. If they answer at all. They scored well under 100. The average human score. And the new one? Whoa. They measured an IQ of 120. That is really far beyond previous techniques. And while I say Bravo OpenAI, I would also like to note that this is not peer reviewed research. It is a little speculative result. So take it with a pinch of salt. And what would you fellow scholars use this for? Let me know in the comments below. And by the way, our next episode will be on this work from Nvidia, so make sure to subscribe and hit the bell icon to not miss it. If you're looking for an Nvidia GPU cloud service for large scale AI projects, check out One Click Clusters by Lambda. Featuring up to 512 Nvidia H100 GPUs with InfiniBand 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.
