Ilya, former Chief Scientist at OpenAI: Just by predicting the next token, humans can reach AGI

Editor: Run

Ilya Sutskever, who was named one of the "Top 10 Scientific Figures of 2023", has repeatedly emphasized that as long as you can predict the next token very well, you can help humans reach AGI.

近日，Nature将前OpenAI首席科学家Ilya Sutskever评为「2023年10大科学人物」。

The "AI technology beacon", which has just been removed from the spotlight because of OpenAI's great changes, does not seem to be ready to return to the public eye.

In a long introduction to him, Nature said that "Ilya declined Nature's interview with OpenAI after its upheaval." And his own Twitter did not retweet the news that he had been named Nature Person of the Year.

However, Nature still spoke highly of Ilya's contribution to ChatGPT in the article, calling him an AI prophet.

In Ilya's view, artificial intelligence has the ability to change the way the entire human civilization exists, rather than just helping humans solve some small problems as before the emergence of OpenAI.

"As long as you can predict the next token very well, you can help humanity reach AGI. 」

This is a point he has expressed on various occasions.

In a podcast, Ilya elaborated on why he thinks tools like large language models, which essentially just predict what the next character will be, can produce intelligence that exceeds the sum of human intelligence.

Ilya explains, "A lot of people think that big models just mimic human knowledge and abilities in a way that is statistically possible, and there is no way to surpass humans. 」

"But if your underlying neural network is smart enough, you just have to ask it – what would a person with great insight and intelligence and ability do? Maybe such a person doesn't exist, but there's a good chance that the neural network will be able to deduce how such a person behaves.

The task of AGI then becomes to predict how such a person might behave.

Good enough to predict what the next character will mean? It's actually a deeper question than the question seems literal.

Predicting the next token well means that you understand the underlying realities that led to the creation of this token.

Just like statistics, in order to understand these statistics and compress them, you need to understand what is the world that created this set of statistics?

And if AGI is to predict human behavior very accurately, what determines people's behavior? Everyone has their own thoughts and feelings, and they do things in a particular way.

All of this can be inferred from the prediction of the next token.

I think that as long as the next token can be predicted well, the AI will be able to guess what a person with this great insight, intelligence, and ability will do, even if such a person does not exist. 」

How do you become a scientist like you who has made such a breakthrough in your field of study?

"I've worked really hard, I've given everything I've gotten and so far my hard work has paid off. I guess that's all there is to it. 」

How much economic value can AI generate by 2030?

"It's hard to answer that question, I think it's going to be a lot. But there is no way to give an accurate number.

But if you ask if AI doesn't have much economic value by 2030, what is the most likely reason?

How far are we from AGI?

It's a difficult question to answer, and I'm not sure if I can give a specific number.

Because researchers who are optimistic about the technology tend to underestimate the time it will take to achieve their goals.

The way I keep myself grounded is to observe the development of autonomous driving. For example, if we look at Tesla's progress in autonomous driving, we can see that he can do almost any behavior required by autonomous driving.

However, it is also clear that Tesla still has a long way to go when it comes to reliability.

Our model may be at a similar stage: it seems to be able to handle everything, but until we solve all the challenges and ensure its reliability, stability, and good performance, it's hard to say that we've reached AGI.

Do you think we still need to think about the breakthrough of Transformer before we reach AGI, or is the existing technology already capable of getting us to AGI?

Technological development can be a gradual process, and the reason why Transformer is considered a breakthrough is because it is not obvious to almost everyone.

So people will feel things. Let's consider the most fundamental advance in deep Xi: a large neural network can do a lot after being trained on backpropagation. What's new?

It's not about neural networks, it's not about backpropagation. But it is undoubtedly a huge conceptual breakthrough, because for a long time people simply did not realize it.

But now, now that everybody sees it, everybody says – of course, it's pretty obvious.

But in fact, it is also a very important breakthrough.

Now that different companies are developing their own models, will different models and technologies be independent of each other or will they come to a common ground in the future?

I expect a lot of research and work to move in a similar direction.

Subsequently, there will be some disagreements in the long run, which means that different research groups or projects will choose different paths and methods.

However, once these long-term efforts begin to bear fruit, the field will converge again, i.e., multiple research paths may converge again on similar results or theories.

The authors also mention that the current number of published articles has decreased, which may mean that it will take longer to rediscover and explore promising directions in this field.

Why did OpenAI abandon the direction of robotics?

"In the past, the difficulty in robotics was that there was too little data, which limited development.

Previously, to enter this field, it was necessary to join a specialized robotics company, and it also required a large team to build and maintain robots.

Even with hundreds of bots, it's hard to get enough data. Because advances in robotics depend largely on the combination of computing power and data, the lack of data is a major obstacle.

Now the situation is different, and the possibility of opening up new paths already exists.

But this requires a real commitment to the research and development of robotics.

This means building thousands of robots, collecting data from them, and finding a way to improve progressively so that the robots can perform some basic useful tasks.

As data accumulates, more efficient models can be trained, enabling robots to perform more complex tasks.

It's a gradual process of improvement, with more robots to build and more data to collect.

In order to achieve the development of robotics, it is necessary to be fully committed and willing to solve all the relevant physical and logistical problems.

This is completely different from pure software development. With enough effort and enthusiasm, it is possible to make significant advances in robotics, and there are already a few companies that have made efforts in this area. 」

Resources:

https://www.youtube.com/watch?v=YEUclZdj_Sc