14:38
Incredible piece of research from MIT. I wanted to make a video about it because it is a potential game changer for AI and I see people talking about it in the media and it seems a bit misunderstood. Okay, what happened? Well, three things. One, scientists used GPT-6 Astra to read the mind, the weights of a neural network. Now, that is normally extremely difficult or impossible. This is a big soup of numbers that roughly gives you the neurons and their connections in a digital brain, an AI system. They work great, but they are often completely incomprehensible. Now, hold on to your papers, fellow scholars, because surprisingly, after looking at another
AI, Astra was able to identify what games this AI system was trained on. It's Breakout, yeah, or Pong or Pac-Man, simplified versions. It recognizes all that from just a bunch of numbers. That is insane. Two, it can fix a failing AI system by editing these weights, kind of like brain surgery for AI's, changing the numbers so it is able to drive smarter. Before the edit, the car initially got stuck on the ramp. Whoo. Now, let's see after the edit. Hmm, look at that. It now successfully goes through. So there was reading, there was fixing, and now three, improving. You give it Quen, one of the crowd favorite open weights
AI systems that you can run at home, and before the edit, it could solve about 8% of the tasks on a math benchmark. Now, after the edit. Whoa, that's super nice. About 48 % that is an insane jump. So how did it do that? Well, Astra recognized that Qwen was not necessarily bad at math, it just got into these crazy thinking loops. It keeps repeating itself and going around in circles, kind of like I do rolling around in bed at night. It tested the other AI model repeatedly and observed what was going wrong. Then it recognized which parts of the neural network are responsible for the loop and edited those to suppress this kind of behavior. But not so fast.
Misunderstanding one that I see going around is that Astra taught this AI mathematics. Not quite. Instead, it essentially helped unblock it. And there are still plenty of limitations. And this is where the coverage that I see elsewhere gets a little carried away. Real talk here. I'll try to clear up what it can do and cannot do in reality. But I am also just a student who is trying to learn, so I might be wrong. Another misunderstanding is that this one can read the mind, can fix the mind, and improve the mind of an AI. This does all three, yes. But people say that it can do all three at the same time. That is not the case. It does these three
things, but in separate experiments, mostly on small AI systems, each with its own custom setup. That is very different. And some say that only OpenAI's Astra can do that. That is not true. There are independent experiments already doing the same, reading which game a neural network can play. And they do it with smaller, free and open models like Qwen that you can run at home yourself. And there are even more things misunderstood about this work. Dear fellow scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér. Here's another one. They say it always works and AI systems are automatically improving each other using this.
Not really. Astra tried improving a different language model, Gemma 3, and it couldn't master any meaningful improvements. So it can read, it can fix, and it can improve sometimes, in narrow cases. And I think that is insane enough. An amazing piece of work. But there are outlets that I feel overstate the capabilities a bit. And wait, it does something more. It can even write a game playing neural network from scratch, which is very surprising because we typically train these with lots and lots of trial and error called reinforcement learning. That is incredible and could be an absolute game changer for AI. Do they know?
Let's see. Of course they know. They are building a full benchmark for this kind of weight space programming as we speak. So this might become a thing. Brilliant work. And yes, we should apply the first law of papers which says do not look at where we are, look at where we will be two more papers down the line. What a time to be alive. I use Lambda to reproduce AI research papers often in minutes. It's great to train your own models, fine tune an existing one, run inference, text to image or video, easy peasy. Running a deep seek chatbot or agent, super fast, super reliable. Lambda gives you powerful
Nvidia GPUs to run your own experiments. I test ideas from papers I cover and moments later, results. Love it. Seriously, try it out now at lambda.com/papers.
