Why Self Driving Cars Will Never Work | Prof. Jiang Xueqin

追加: 2026-05-21

[00:00:00]All right. So, in other words, it's actually no different from a Google search. The only difference is that it's taking the Google search, figuring out what the most popular answer is, and then presenting to you in a way that makes you think that it's talking to you directly. All right? The trick, and this is really important to understand, guys, is it's trying to trick you.

[00:00:23]All right? It's not trying to teach you.

[00:00:25]It's not trying to tell you the truth.

[00:00:27]is trying to trick you into believing it. That's what we call a hallucination.

[00:00:38]Okay? You have you guys have to understand this idea. There's nothing truthful about what um Chhatt says. All it's trying to do is trying to manipulate manipulate you with words or pretty words into believing that it knows what it's talking about. But it itself cannot judge what it's doing. Okay.

[00:01:04]All right. Any questions so far? Are we clear? Okay. All right. So now the question is how does it do that? Okay.

[00:01:13]And um so I'm going to teach you a little about artificial intelligence and please stop me if I'm not being clear about how AI works. Okay. All right. So AI doesn't exist.

[00:01:28]What exists is we call supervised supervised machine learning.

[00:01:36]This is a technical term. Okay.

[00:01:46]All right. Supervised machine learning.

[00:01:47]Okay.

[00:01:49]And how it works is this.

[00:01:53]Before how computer programs would work is we would write the program the algorithm and then we would give the input and it would produce the output. Okay.

[00:02:10]So the output may be a plus b we get the input one one the output will be two.

[00:02:15]Okay, very simple. How supervised m machine learning works is okay this is fine for simple problems but there's certain hard problems that humans cannot figure out okay and one hard problem is the idea of facial recognition technology facial recognition how do I separate faces okay and so the problem is this I have about a million faces. 1 million faces in a database.

[00:02:53]Okay.

[00:02:58]And I don't know how I can best differentiate these faces.

[00:03:05]Now what I do know is that there are certain characteristics about the face that allows me to differentiate. Okay.

[00:03:13]All right. Right? So certain variables weights.

[00:03:17]Okay. So for example, eye, for example, uh nose, chin. Okay. About a million.

[00:03:25]Okay. About a million weights. So I know these things do matter, but I don't know how much they matter. So I'm trying to figure out what the waiting is. And I could try to play play by myself like say 1% 2% 5%.

[00:03:43]But as you can imagine this will take too long because there are too many possibilities. So what I do is this. I let the computer figure out it by itself. I I let the computer figure out the waiting by itself. Okay? And the way I do that is using my technique called back propagation.

[00:04:04]All right. So what? So I control the input.

[00:04:10]Okay. The input then I control the output.

[00:04:16]Yes or no?

[00:04:19]Okay. All right. So does a face match or does it not match?

[00:04:24]And what I'm trying to do is I'm trying to figure out a situation in which all faces are matched perfectly. And I do that by training the computer to constantly back propagate until it gets the waiting perfectly. Okay. So basically what I'm trying to do if if you understand um how this works is I'm trying to turn each face into a distinct mathematical model.

[00:04:58]All right.

[00:05:00]That is unique to it.

[00:05:02]Okay, does that make sense? All right, so it's pretty simple. It's not doing that much. But to make it sound really fancy, I give it really fancy names to trick people to believe that this is actually much more soph sophisticated than it is. Okay, so what names do I give it? This waiting system, I CALL IT A NEURON NETWORK, GUYS.

[00:05:28]It's a brain.

[00:05:31]It's magic.

[00:05:33]Okay. And back propagation, I don't call it back propagation. I call it deep learning.

[00:05:40]You see, and I don't call it supervised machine learning. I call it AI. AH, THERE YOU GO. MAGIC. You see, all I've done is taking a very simple process and giving it like a really really fancy names.

[00:06:04]People listen like why do I do that? And so people will say, "Oh, it's for marketing purposes. It's to get more money from investors. It's to trick people." No, no, no. The real reason is you're trying to with his names create God. Okay? It's what we call the occult.

[00:06:24]So the AI is fundamentally an oult practice and I'll show you why in a moment. Okay. Yeah. You have Yeah.

[00:06:33]Vincent, you have a question. But why do people need to create God using the way of AI? That's a great question. Okay.

[00:06:41]The answer is AI only works if it becomes God. You understand? AI it by itself does not do anything. Once it becomes God, then it becomes everything.

[00:06:51]Okay? And how God works is you imagine God.

[00:06:56]>> But why do people want to make a make a god? To control the world.

[00:07:03]>> Oh, >> to become God, right? What's the point of existence?

[00:07:08]You live, you die. You have an opportunity to become God. Why not?

[00:07:13]Right? But but but I'll I'll talk more about this later on. Okay? But are you guys clear about what's going on? Now what's really important to understand is that there's certain problems with the system.

[00:07:26]Okay, you need to create certain conditions for supervised machine learning to work and these three conditions are clean data.

[00:07:41]Okay, the data you present to the computer have to be correct. Okay, it can't be an opinion like I like computers. It has to be an image of some sort. All right, it has to be clean data that will help the computer learn.

[00:07:57]That's actually hard to do. That's why most of the data that's presented to the computer is actually from the internet.

[00:08:03]Okay, that's the first constraint.

[00:08:05]Second constraint is that you need a measurable goal.

[00:08:13]Okay, you have to ask computer does this face match the name. Okay, you cannot ask a computer what is God, what is good, what is evil. That that you it has to be a measurable goal. Okay, that's the second major constraint. The third major constraint is um define parameters.

[00:08:41]Okay, you're you're meaning in other words you need to present it with a database of some sorts. In fact, all machine learning works of database. So you look at um translations. Translations are working off databases as well. Okay.

[00:09:01]And the great danger to the system is what we call edge cases.

[00:09:08]Edge cases.

[00:09:13]Okay. Edge cases. Edge cases breaks the system down. All right. And so the classic example is self-driving cars.

[00:09:25]We have cars that drive for a long time, but and we're almost like 99.99999% there to um self-driving cars, right?

[00:09:44]The problem are edge cases.

[00:09:48]And the pro and the major edge case is how do you deal with humans who are intentionally trying to cause an accident with a self-driving car? Does that make sense?

[00:09:59]And the answer is you cannot.

[00:10:02]In this situation, there's only one solution to make this 100%.

[00:10:10]And that is to take away the right of everyone to drive to make every single car a computer and a robot. Does that make sense? Okay. If you take away the steering wheel, you can't cause an accident and then the world will be perfect. Okay. So not only is AI very limited in its um capacity and capability, but AI if it is to be effective, it demands that we fundamentally restructure human society to benefit AI to make sure AI can be effective and that means taking away the individuality, the diversity and the autonomy of human beings.

[00:10:57]Okay. Does that make sense, guys? All right.