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!
