The birth of ChatGPT showed people the amazing ability shown by the big language model based on artificial intelligence, probably for the first time ever, through the interaction of natural language, people really feel the relationship between human and machine, which produces a subtle understanding, and a new round of artificial intelligence boom followed.

ChatGPT's amazing "comprehension" is a true awakening of machine consciousness? What are the remaining obstacles on the road to generalized AI? Is the threat of AI overstated? What is the future of artificial intelligence development? With these questions, Tencent News "subliminal" recently had an exclusive conversation with Ruslan Salakhutdinov, professor of artificial intelligence at Carnegie Mellon University and former director of AI at Apple.

Salakhutdinov studied under Geoffrey Hinton, the Turing Award winner and "father of deep learning," and is the author of famous learning methods such as pruning and deep coding.In 2016, Salakhutdinov joined Carnegie Mellon University, and in the same year, he was awarded the Intel NVIDIA Artificial Intelligence Pioneer Award, and joined Apple as its first Director of AI Research.

As a senior scholar in the field of artificial intelligence, Salakhutdinov expressed optimism in the conversation. He expressed his excitement about the capabilities demonstrated by this round of AI and did not believe that the current development of AI would pose a threat to human survival; he also said that big language modeling, while not necessarily the only path to general AI, is an important foundation, and that progress in this area provides more inspiration and confidence for academics to explore general AI. As a former student of Hinton, he also expressed his understanding of Hinton's concerns, but still insisted on his more optimistic judgment of AI.

The following is a transcript of this conversation, which is partially abridged:

Big models will revolutionize education and other fields

Tencent News' The Dive: Prof. Salakhutdinov, thank you for the interview with us today. We're in a very complex world right now, and a lot of things are changing rapidly at a daily, hourly rate, so we feel very fortunate to have the opportunity to be here to really discuss some of the very important issues that we care about. Let's start by talking about the hottest topic that's come up in a while, ChatGPT. what was your first reaction to the initial release of ChatGPT late last year?

Russ Salakhutdinov: Before that there was actually GPT-2, GPT-3, and we had expectations that those models would get stronger and stronger, but when ChatGPT went live in November of last year, a lot of people were very surprised to see that it worked amazingly well.

Building these large language models on large amounts of data doesn't just allow you to ask and answer specific questions, it also summarizes things for you. A lot of my friends are using ChatGPT for programming. If I had been asked a few years ago, could I have this technology today? I probably would have answered no, so that's why I think it's an amazing technology that also opens up whole new opportunities in a lot of different areas.

Tencent News' Dive: Speaking of programming, it's really great, I didn't have any prior programming experience, but I used ChatGPT and used natural language to describe to it some of the goals that I wanted to achieve, and asked it to help generate the appropriate code, and it literally provided me with program code that I could execute.

Russ Salakhutdinov: Yeah, this is one of the areas where I see a lot of people actually using these big language models to help with programming because behind these big models, Microsoft and OpenAI basically have access to the entire GitHub repository of program code, so the language models perform very well when users are asking code-related questions. In the future I think almost every programming tool will have a large language model that can help you, and this is an area that will see a lot of applications.

Tencent News' Dive: While we can directly relate to the amazing capabilities that ChatGPT has demonstrated, for many people they still don't fully understand what's happening within the AI industry, so what does ChatGPT, or any other large language model, really mean to the average person?

Russ Salakhutdinov: First of all these models are based on deep learning techniques that people have been working on them for the last 20 years, and the model that provides the underpinnings for ChatGPT or Bard comes from Google, it's the Transformer architecture.

But what this means for the average person, I think we're going to start seeing more and more people you know, interacting with data through natural language.

For example, I think there's going to be a lot of opportunities in the education space where in the future you're going to be using models like ChatGPT or Bard, which can help you with your homework, and writing essays is one of those things, and ChatGPT can probably write better essays than you can, so I think we're going to see a transformation in education.

For example, if your child is in 9th grade, 10th grade, we're going to have AI tutors who are going to personalize their tutoring for each individual and you can ask it questions and it can teach you new things, explain to you how to write code, or explain to you some mathematical terminology, and I think that area is going to grow rapidly in the near future.

Also, the way we interact with computers today is that we usually use Google or Baidu to search for information, and I think that's going to change as well because in the future the interaction is going to be done through these chatbots or language models, where I can ask it a question, and it will find the right information for me and give me the information that I want so that I don't have to go to a website and click and look for the right information.

Tencent News' Dive: So that's why Google is feeling very anxious right now?

Russ Salakhutdinov: Yes, but these large language models like ChatGPT aren't perfect at the moment because they create "illusions", can we solve this problem in the future? Some people don't think we will be able to solve this problem, and maybe it will take us longer, but in any case, I think the whole way we interact with computers will change.

"I don't think we're at a stage where we need to worry about the risks of AI."

Tencent News, "The Dive": now, we also see that there are a lot of debates around the development of AI and a lot of concerns about the development of artificial intelligence, and we are very curious to know the position you are in now. Are you more optimistic or more pessimistic from the perspective of the future development of humanity?

Russ Salakhutdinov: That's a great question. I think right now in the field of AI research, it's basically being split into two camps, one camp believes that AI will do us a lot of harm, and even eventually threaten the survival of the human race, and become the end of the human race in the future.

The other camp believes that AI brings so many opportunities and new things, for example, the healthcare system will be transformed, AI will help you design new drugs, it will also help you access new information, the potential for the use of everything related to AI is so great that it will be a bright future.

I'm in the optimistic camp, I don't think we're currently at great risk from AI, even though my former mentor, Jeffery Hinton, really believed in the threat of AI. He's very, very smart, and I don't think I've met anyone smarter than him, but I don't think we're at a stage where we need to worry.

I think one of the biggest concerns at the moment is the spread of misinformation, for example today I can create an image and using hints I can easily fake someone's voice, and in the near future we'll be able to generate videos like that on a mass scale at a very low cost, and that's going to make it very difficult for most people to tell the difference between information that's true and not true, just like Photoshop.

It used to take a long time for this to be done, but now with this technology, maybe someone can make a video of me in my voice, and you might think you're talking to me, but it might just be an AI talking to you, so a lot of people are a little bit worried about this area. Now the question is whether there should be regulations and what we can do to prevent that from happening. And then of course there are some concerns raised about information bias, where a lot of models are trained on a lot of data, and if the data itself is very bad data, the models will learn that bad data, but those models know nothing about it.

Tencent News' Insider: Yes, we noticed that you retweeted a New Yorker cartoon on your personal Twitter feed a couple of days ago, which is a very interesting image depicting robots enslaving humans, and you commented that "this image basically depicts the situation we're in", so do you think that's the future we're facing?

Russ Salakhutdinov: It's a joke, but if you look at ChatGPT, these are models that are essentially trained to predict or autocomplete sentences, to come in from human feedback to do some fine-tuning of these models, but at the end of the day, these are what's known as autoregressive models, they predict one word at a time. While these models do well in natural language understanding, these models have never seen a video, I mean, right now we're looking at images and language, but they still lack some of the so-called models of the world, which is how the world works, how physics works, how do you know that we're human. We understand the world and these models are lacking because they can only look at text correctly, so that's an upper limit on the capabilities that these models might reach, which of course doesn't mean that they're not useful, they're going to be very useful in a lot of different domains, but at least I think it's going to be a long way off to be smarter than humans which is still a long way off.

Tencent News' Dive: As a former student of Prof. Hinton's, can you understand why he's so worried?

Russ Salakhutdinov: Yes, in a way I can understand his concern. Because from a human perspective, we have to keep learning, and it takes about twenty years before you start to become very smart in your field; but with digital intelligence, because of this massively distributed training, it can read everything on the entire Internet in a matter of days.

Right now we have tens of thousands of GPUs that are constantly acquiring data, so he's saying that if this continues, the likelihood that we'll have superintelligence is quite high, and if you have superintelligence, then there are risks associated with it that exist, and that's one of the concerns that he's raising, and it's a reasonable argument. I don't refute it, but I just don't think these models find patterns in the text and form decisions from that against humans, and as said before, they lack perception of the physical world.