OpenAI: The Age of AI Is Here!

I must admit I am a bit nervous. I am  nervous because this is our 941st video,   and it might be one of the most important  ones I’ve ever done. I’ll try my best. In   their latest paper, scientists at OpenAI won  a gold medal, sort of, and in the process,   they found something absolutely incredible  about the nature of artificial intelligence.    I have been waiting for this for a long time and  barely anyone is talking about the paper. Crazy. I promise to get to the paper in a moment, but  to understand it, we have a little gaming to   do. You see, earlier, when we wanted to  write a computer program to play a game,   we had to give it instructions by hand. Go  here, turn around, jump up, and so on. Then,   as the first AI techniques appeared, many of them  were able to learn a bit, but they were taught   the rules of the game and some of the classic  strategies. Think about giving a few books of   opening strategies for a chess AI to learn. That sounds like a good idea, we humans have   years and years of knowledge about the  game. Of course. Why not help the AI? Well, that might be a big mistake. Why?   Because if you teach it good strategies,   it might never find the best  strategies. Wanna see an example? This is the “You Shall Not Pass” game, where  the red agent is trying to hold back the blue   character and not let it cross the line.   Here you see two regular AIs duking it out,   sometimes the red wins, sometimes the blue is  able to get through. Nothing too crazy here.    This is the reference case which is somewhat  well balanced. Now, look closely, because here   comes the hacker adversarial agent. Ha! Yes, you are seeing correctly,   this chap it doing nothing. Absolutely  nothing. But it is doing nothing in a   way that reprograms its opponent to  make mistakes and behave close to a   completely randomly acting agent!   This paper was absolute insanity. And that is the key. If you teach  the AI strategies on how to play,   it might find some good strategies, but  it won’t ever have the chance to find the   best ones. And these are things that  we would never have found ourselves. So, here is a strategy: teach the AI less,  and let is learn on its own more. Okay,   but I hear you super smart Fellow  Scholars asking: okay Károly,   this works on this little toy game, does this  concept work on a greater variety of tasks? Glad you asked. Let’s see  together. Have a look at this. So far we have talked about mastering one specific  game. You train an AI that is able to do that.    However, if you wish to play a different game,  you need a different AI. One AI, one game. Now   here is the best part: you can also train one AI  to be able to play many games, and if you do that,   something so surprising happens that it is one of  the most stunning insights of my entire career. What is that? Dear Fellow Scholars, this is Two  Minute Papers with Dr. Károly Zsolnai-Fehér. It is the following experiment. You take a  specialist AI, and let it master one game.    Then you take a generalist AI that kinda  knows this game, but knows other games too.    Who wins? Of course, I say the specialist does  because it played the game so much more. Imagine   the olympics: you have a super muscular  wrestler who did wrestling his whole life,   and you get a scrawny guy who kinda dabbles in  wrestling, swimming, and 20 other sports. Now   these two wrestle. Who wins? Of course,  the specialist, the muscle guy wins. Except, that he doesn’t. Look. The guy  who dabbles in many sports, that is,   the AI that played many games, but  one particular game only a bit,   it can beat a specialist that played this  game a ton. I mean, what? That is insanity! And now, I can finally tell you about  why this fresh new research is a stunner:   OpenAI applied this concept to not just games, but  to programming, and quite possibly, everything. So   what did they do? They started using their  AI to solve challenging programming tasks. So, what is the result? Now hold on to your papers  Fellow Scholars, I shall present to you the Holy   Chart. Their o1 system did really well. So  far so good. Then, their specialist system   did even better. This is a specialist. This is  the muscular wrestler. It was shown handcrafted,   human-taught data and strategies to excel at one  thing. Is it good? It is not good, it is amazing.

Under somewhat relaxed conditions, this is  good enough to win a gold medal. Whoa. That   is stunning, and normally we would stop here, but  it gets even better. They introduced a smarter,   generalist agent, o3 that has no specialized  knowledge and it learned on its own, so in return,   it is probably worse and…what? Are you seeing  what I am seeing? Once again, in a new area,   the generalist beats the specialist. The  scrawny guy beats the wrestler. Wow. But why? Because this AI learns something in one task,  and is able to apply it to another. But wait,   to me, that sounds exactly like intelligence.   To me, this sounds like artificial intelligence   is finally a possibility. Think of  all the good this will be able to do,   from designing new drugs to defeat  previously untreatable diseases   to giving a personalized teacher to every  child on the planet, and so much more. Wow. To get intelligence, we don’t need to teach  sophisticated strategies to an AI. No-no-no!    We’re just holding it back. Instead, get  a smarter AI and make it learn by itself,   and it will do better. If will find the crazy  strategies that we cannot find ourselves. So,   yes it is possible, and it is simpler than  we thought: we need simple algorithms, tons   of compute, and you will likely get artificial  general intelligence, possibly superintelligence. So much so that the o3 AI now ranks among the  best human programmers in the world. And this   is what Two Minute Papers is about. Of course,  it is about the papers, but also about the wider   context around the paper. Something that you  don’t really get elsewhere. And with that,   the age of AI is here. Subscribe and hit  the bell icon if you wish to see more. So,   what would you Fellow Scholars use this  for? Let me know in the comments below.