Principles, Not Black Boxes: Ray Dalio on When to Trust AI

摘要

30 秒看懂

Ray Dalio argues that you should not rely on AI when the future may differ from the past and you lack a deep understanding of cause and effect. His alternative is to write down your decision criteria as principles, convert them piece by piece into algorithms, and let computers handle the processing while humans handle invention and judgment.

1Bridgewater has spent about 25 years encoding its thinking into algorithmsthat run as an independent decision-maker in parallel with the humans, like a second player in a chess game whose moves can be compared with your own.
2If the future can differ from the past and you lack deep understanding of cause and effect, you should not rely on AIRepeated, stable tasks like surgery are fine to automate; novel conditions are not.
3Machine learning builds its own equations from data, and those equations often cannot be explainedThat is why they can escape understanding and mislead when conditions change.
4Use the mind for inventing and the computer for processingComputers can process more information faster, more accurately and less emotionally, while goals and judgment calls should stay with people.
5Almost everything can be expressed as an algorithm if you write down your criteria at the moment you make a decisionDalio says he has been shocked at how much he could express that he thought was impossible.
6The next big move is from systems of record to intelligenceDalio expects principles converted into algorithms to turn data into guidance, including on strengths, weaknesses and who to work with.

主题企业与战略 AI Strategy Decision Making 金融 AI Philosophy 企业 AI

全文

按章节阅读完整文稿

Bridgewater runs its own thinking in algorithms alongside human decisions

Lex Fridman opens by asking what role AI will play in analysis and decision-making at Bridgewater Associates, which manages around 160 billion dollars. Dalio answers with a broader rule for AI. For the last 25 years, he says, the firm has taken its own thinking and put it into algorithms that take in data and make decisions.

Those algorithms operate as an independent decision-maker, in parallel with the humans. Dalio compares it to a chess game in which a computer plays one side of your own game: it makes a move, you make yours, and you compare the two. The firm has done a lot of this, but he wants to state a rule for when it should and should not be trusted.

Do not rely on AI when the future differs from the past and you lack deep understanding

Dalio's rule has two conditions: if the future can be different from the past, and you do not deeply understand the cause-and-effect relationships behind a bet, you should not rely on AI. He contrasts this with surgery. Filming doctors, converting that into algorithms and having robots operate would be fine, he says, because the same situation such as appendicitis recurs and is handled the same way.

The risk is that most AI takes lots of data at high sample sizes and builds its own algorithms. Dalio distinguishes two routes: expressing your own thinking in algorithms, or feeding in data and asking what the algorithm is, which is machine learning. The latter often yields equations that are hard to describe, so they can escape understanding. Used on things that repeat, that is fine; applied to a future that differs, he says you are going to get in trouble.

Let computers process and let humans decide what is worth doing

Dalio says AI and computers are particularly good at processing, while the question of what you want to do is still mostly determined by the human mind. He gives examples such as how to raise children or whom to marry, saying it will be a long time before AI's judgment is good enough for those. A computer may help, but pure machine learning on data will not find the answer.

He thinks you are much better off taking one of your own criteria for such questions and putting it into an algorithm than asking AI to decide. By and large the mind should be used for inventing and creative thinking, and the computer for processing, because it handles more information faster, more accurately and less emotionally. Fridman then probes whether the human element can really be reduced to a program, and Dalio says almost everything he thought impossible to express he has managed to express in algorithms, though not all things.

His method is to take a hard question like raising children one piece at a time: at what age, which school, and so on. In the moment of a decision, or just after, write down your criteria in words. Dalio has built an iPhone app, with an Android version expected in a few months, that helps people write down their own principles.

Written principles become algorithms, and intelligence is the next step beyond systems of record

Writing principles down forces you to articulate them clearly, Dalio says. You can access them next time the situation arises or show them to others to test whether they are the right principles. Then you ask how to express each as data: you form an equation showing the relationship between parts, identify the variables, put that piece into the computer, and repeat. Before long, he says, you have a decision-making system.

Fridman notes this seems to contradict Dalio's earlier claim that a computer cannot raise a child, since a society could arrive at principles for child-rearing and encode them. Dalio replies that the two are consistent: originality does not come from machine learning the first time, but once principles are developed they can be automated, and the computer processes them effectively.

Dalio expects this to define the future, with principles converted into algorithms that also carry the benefit of collective decision-making, since individuals now decide in ignorant ways from what sticks in their heads. He expects medicine to work like this, with a system giving better guidance than a rushed local doctor. He says the next big move is from systems of record to intelligence on questions like your strengths and weaknesses or whom you work well with, and at 70 he wants to pass his tools along.