Can ChatGPT Plan Your Retirement?? | Andrew Lo | TEDxMIT

Several ways in which we have already imposed this type of standards to humans. Almost all the financial institutions of the profession have a code of behavior and a set of ethical criteria to which they must comply. They must focus on the best interests of their investors. The question is, can we get a program to do this? It turns out that in computer science, there is a term, so it is called "the problem of alignment." Can you do artificial intelligence in line with

your

behavior? Now there are many different aspects of human behavior. So you have to ask, well, what kind of behavior?

Then we will focus on a couple. I don't have time to review them all. But I will give some examples to clarify if we can know if the excellent language model is correctly compatible. I will do this through a game. It is called the warning game. Economists have achieved this as a way of understanding the nature of human interaction in several specific economic environments. So let me tell you how it works. Suppose we have a proposals provider, say, and the job of the employment provider is to suggest another person, for example, how to divide a certain amount of money that the offer can reach, say $ 10.

Then, my work is to suggest a part of those $ 10 and suggest. Your decision is simply accepting or rejecting the proposal. If you reject the proposal, none of us will get money. But if you accept it, we will get the money, and it will be divided according to what we agree. Good? Then, I offer you $ 5 of $ 10. You say: "Yes, I accept." In fact, money is divided between us. However, if I offered something else and said: "No, I reject it" No one will get money. So we will play this game now actually. Good? I need a volunteer. Who wants to play this game with me?

Well, will you come here for a second? (Applause) Now. We need money. So we get the money from Gillian Ross. Gillian, are you coming here? This is not completely trivial, because she is an expert in obstetric artificial intelligence and a doctorate at the MIT Institute. It will collect money soon. So, Gillian, I will suggest something in Fernando. The question is whether or not you accept, well? So I will suggest that you give you $ 0.05. I accept. Do you accept? Excellent. Then I get the rest. Do you agree with that? Obtain $ 0.05, then you can maintain $ 9.95. Do you agree with that?

No. Good. I'm sorry. I think we don't get money. I suggest $ 4. No. Well, the end of the game. Thank you. Sorry, Gillian, I lose too. You lose. Thank you very much, Fernando. (Laughs) (applause) How do large language models behave? It turns out that most people represent 40 % of the division they offer, and it turned out that this is generally enough for people to agree. Large language models are not all there, but some are present. Through many other examples like these, we can draw a map of the behavior of large linguistic models and compare it with how human beings are involved.

This will eventually allow us to form large linguistic models to be completely sure in a way we feel comfortable, such as people. The way we learn on the golden base in the field, which makes others like others. Ultimately, as they age, they learn a different version of the Golden base in the business world, which is that those who have gold generally establish the rules. However, the question is, can we really reach large linguistic models that will help, increase or replace their trusted financial advisors? The answer is to return within a year, we will tell you. Good.

Thank you so much. (applause)