OpenAI's GPT-4: A Spark Of Intelligence!

Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Reading this paper was an absolute  delight, but also a shocking experience,   I will try to tell you why and you can  let me know if you feel like that too.   This is Microsoft’s assessment of the  capabilities of OpenAI’s GPT-4 AI,   and I am out of words. This  is unbelievable. Why? Well,   let’s see together. By the end of this  video, I think you will be out of words too. Let’s start immediately with what I find to  be perhaps the most fun experiment. Scientists   had the idea to play a text-based game with  GPT-4 where the AI can navigate through a map,   and we make up a little adventure and  tell it what happened. For instance,   it starts in the main room, and it chooses to  move left. We say that this is now kitchen 1,   so it says that it wishes to move up. And so on  it goes. Nothing too crazy here. Right? Well,   check this out. After reaching the goal,  it can build up and reconstruct the map   of the whole game. But wait, how do we know?   This is just a bunch of text. This one, yes,   but don’t forget, this is an assistant, so  we can ask it to, and get this, draw a map.    And the map it has drawn it perfect! And the funny  thing is that no one even asked it to remember the   layout because it will be asked about later.   I absolutely loved this. This is so much fun. Now, remember, this can only answer in terms of  text, however, we can create images by using text.    One example of this is called TikZ, which can  create beautiful images from a text description,   but you can ask almost anyone who has ever  used it, have them describe the experience,   and “joy” is not likely a word you will hear  from them. If only we had an AI that could   do all this for us. You know what? Let’s  try it! Let’s ask it to write the code for   a little person built from the letters of the  alphabet. This is not bad, but we can ask for   some improvements. Yes, and then, add some  shirt and pants, and there we go! Great job. Or, we can also ask for a unicorn as well. And  here we find something super interesting. Hold   on to your papers Fellow Scholars, because it  improves over time. These prompts were run in   the span of a month, and over time,  the system became better and better. What’s more, it can even create a really  simple mockup of a video game. And now,   we can give this to another AI, Stable  Diffusion to embellish it a little. And,   there we go. That almost looks like  a screenshot from a real video game. And while looking through the results,  we forgot that a miracle happened. What   is the miracle? Well, this version of  the GPT-4 AI has never seen an image. This is an AI that reads text. It has  never ever seen an image in its life.    Yet, it learned to see, sort of, just from  the textual descriptions of things it had   read on the internet. That is insane. Let’s test that some more! Have a look at this.   Little AI, imagine that we have a book, 9 eggs,   a laptop, a bottle and a nail. Please tell me  how to stack them onto each other in a stable   manner. Now, the previous version was  off to a great start, and then it said.    “Place the eggs on top of the nail, making  sure they are balanced and not tilting to   one side. ” You know what, little AI? You do  it. I will believe this when I see it. Now,   let’s see the new one. It says: “Arrange the  9 eggs in a 3 by 3 square on top of the book,   leaving some space between them.   The eggs will form a second layer   and distribute the weight evenly. Make  sure the eggs are not cracked or broken,   and handle them gently to avoid dropping them. ”  Now that’s what I am talking about! Checkmark! And there is so much more to talk about, I  don’t even know where to start. When asked,   it can even create a simple little  video game in HTML and Javascript,   even with rudimentary physics. Its coding skills  are so sharp, it would likely be hired as a real   software engineer. And that is perhaps an  understatement. Look, I loved this part:

Time allotted: 2 hours. Time spent:  1 second shy of 4 minutes. Not hours.    Minutes. Holy mother of papers. It crushed  the interview faster than any human would. I was also shocked by its mathematical  skills. It can be given these problems   from the International Mathematics  Olympiad, these folks love creating   problems that look like a piece of cake, but  require considerable mathematical experience   and thought to solve well. But this comes  out almost instantly. And it nailed it. In a different, physics-inspired problem from  somewhere else, the previous version of it just   made something up. Not cool. So, can the new  one solve it? Well, it identified that using   integral calculus is necessary, what’s more,  like a good student, it starts integrating by   part. That is fantastic. And the result is that  it almost nailed it. It made an arithmetic error,   but other than that, it did very well. That is  the perfect metaphor of the whimsical nature   of the AI systems we have today. They understand  how to apply integration by parts in reality,   which is outstanding, and then, it slips  up when it needs to count up a few objects. So, conclusions. Where does all this put us?   Get this - Microsoft claims that GPT-4 might   have a spark of general intelligence. This is  something that most people thought, and perhaps   still think is not possible at all, but even if  it is, surely not in our lifetime. And here we   are. The future is here. This is intelligence like  we’ve never seen before. What a time to be alive! So, this is all very impressive, but is that it?   If I saw that the answer is no, that would be an   understatement. This paper is over 450 pages,  so we have only scratched the surface here,   but I would definitely like to continue this  journey, so if you think that is something   that you would like to see, consider subscribing  and hitting the bell icon to not miss out on it. Thanks for watching and for your generous  support, and I'll see you next time!