You find yourself somewhere in rural乡下的 Japan, walking along. You find this little shed, some little structure. A woman approaches向…靠近. In her hands, she has something written on paper. It's Japanese characters性格. She slides滑 it into an input投入 slot狭缝, goes around the other side, waits a couple of minutes, and another piece of paper emerges. She looks at it, smiles and walks away. Then a man approaches向…靠近, and he does the same. He's got something written in Japanese on a piece of paper, slides滑 it in, walks around the other side of the shed, waits a minute, and receives something.
You run up to him and ask, "What are you doing? What is this?" He said, "Well, it's the question box. Anything I want to know, we can just write it down, put in the input投入 slot狭缝, and the answer comes out the output输出 slot狭缝." You see a hundred people wait in line, and all during the day, they keep asking questions, and answers keep coming out. At the end of the day, the door is open. Inside is some American dude that doesn't speak a lick of Japanese.
"How are you doing this?" you ask him. He says, "Come here." And he folds out this massive厚重的 book, and this book has huge巨大的 pages with two columns纵队. He says, "Whenever I get a question, I look for it in the left column纵队 of this page. When I find it, I just draw whatever's to the right. I didn't even know I was answering questions until you told me." Those people felt understood, were they? If you put a fluent流畅的 Japanese speaker in that shed, he would understand them.
But the answers would be the same. Is there any difference at all? What does it mean to be understood? There could be a way to distinguish区别 between the two cases. When you have the guy who doesn't speak Japanese and the guy who does, you have them put up all the pieces of paper on a board, and connect them with red yarn纱, like a conspiracy共谋 board. You may have seen those in the movies like a beautiful mind.
How would the non-Japanese speaker connect the dots? How would the Japanese speaker do it? Certainly differently, right? No matter how they would be, it would be different. This is a little bit like the difference between the connections连接 made in your mind and in an AI. Now, I hear a lot of you are afraid of AI. You shouldn't be, but we need to understand it a little bit. Now, you might be afraid it's going to take your job.
That's one of the questions I get. Maybe every company in the future will have two employees, a man and a dog. Because you can do so much with AI, why have people do all the jobs? The man's job will be to feed the dog. The dog's job is to keep the man away from the computer. GPT is not ready to take your job. I asked GPT, "Please tell me about the early life and great works of the 17th century philosopher哲学家 McGruff the crime dog."
It very happily gave me a confident answer about his early life, his influences影响. He had a philosophy哲学 of positivity toward humanity人性, it was quite controversial有争议的 at the time. But his parents supported him and he did well. Good for him. What we connect in our mind, the shape of those connections连接 is everything here. How do humans connect thoughts? Go home, blue sky. We know those thoughts are connected because they're next to each other. Adjacency communicates something. But there's limits限度 to this.
Jack gave Jill a book. Well, gave is connected to three things. Jack the Giver, Jill the recipient容器, and the book, the gift. How can gave be next to three words at once? You can't do that because every word comes out of your mouth one at a time. Somehow由于某种原因 in your brain you turn that sequence[数][计] 序列 into a tree. A tree like the one you see there. And the links have different colors because there's a giver, there's a recipient容器, and there's a gift and they're all treated对待 differently by your brain.
In English we use word order. In some languages they use suffixes后缀. There's other ways, but every language creates that tree inside your head. Language is tree-shaped形成. AI thought is not tree-shaped形成. Let's look at the structure of AI. LLMs like GPT and what everyone's excited about these days start off with something called self-attention自己. Now I'm going to get a little hand wave. If you're professional in this, I'm skipping跳过 some details. You can thank me later. Attention is like a number between zero and one.
If you put all your attention on one thing, that's like a one. If you're totally ignoring不顾 something, it's like a zero. Well, we didn't have time to teach the computer those trees I showed you. So we let it guess, and it looks like that. Some connections连接 are dark, which means they're closer to one, and some are light, so they're closer to zero, and then you get a matrix矩阵. It's just the words all connected to each other.
That's all it is, but it's just numbers. It's a fuzzy version版本 of a tree. It's no longer a tree. It's a fuzzy amalgamation of trees. Well, they tried this on some text. It didn't quite work, but they liked the idea. So let's try two. It didn't quite work, but they liked the idea. So let's try four. Didn't work, but they liked the idea. So I said, let's try eight. It didn't work yet. But they liked the idea so much, they said, let's stack堆积 it together with another neural network.
We'll call it an encoder, and we'll make a stack of these encoders and a stack of decoders. All of them filled with matrices矩阵 were so far removed from human thought now. That is not even funny. It's matrices矩阵 all the way down. I called a meat grinder[机] 研磨机 for matrices矩阵. When my daughter was three years old, I could have a reasonable合理的 conversation with her. But GPT, the one you use, has been trained on over a million years of English at a natural exposure暴露 of 20,000 words a day, over a million years.
It's not learning the way we learn. It's not like us. It's there's no simple tree structure inside there. So I'm not afraid of AI. As an AI researcher for over a couple of decades, I don't fear AI. Instead, I think about what business I'm in. I'm in the biomimicry business, trying to reverse颠倒 engineer设计 a human brain. No matter what task you're doing, the principles原理 of reverse反面的 engineering工程 apply the same, whether you're teaching a computer to talk, listen, walk.
When you have to teach a computer to do something, you start from scratch. You feel like you're thrust插 into a dark forest of unknowns. And you have to sit there and figure认为 out how do I even get started? It's only then, when you realize just how many untold无数的 lessons mother nature大自然 must have learned over the billions of years that life has been here on Earth. No, I don't fear that artificial人工的 brain. You shouldn't either. Instead of fearing that artificial人工的 brain, you should stand in awe敬畏 of the one you've been given. Thank you.