How does Bridgewater invest in AI?

How does the world's largest hedge fund view AI?

On Monday, July 3rd, Greg Jensen, Co-CIO of Bridgewater Associates, discussed Bridgewater's views on AI technology in an interview, sharing his perspectives on how Bridgewater invests in AI, how it utilizes AI for investments, and its outlook on AI technology.

Bridgewater's Approach to Investing in AI

Jensen stated:

As part of the restructuring of Bridgewater, we have done something we haven't done before, which is to have some people research and invest in things that may not be immediately profitable but are long-term projects for us.

Therefore, we established this AI project with a team of 17 people, led by me. While I am still actively involved in the core work of Bridgewater, the other 16 individuals are fully dedicated to reshaping Bridgewater through machine learning.

We are setting up a fund specifically operated by machine learning technology, which is the work we are currently doing in the lab, pushing the boundaries of artificial intelligence and machine learning capabilities.

However, there are significant challenges in establishing such a fund. If we use large language models, they have two issues. First, these models have received more training in language structure, so they tend to generate responses that appear structurally and grammatically correct but are not always accurate. This is a problem. Second, they create illusions and fabricate things because they focus more on the structure of the upcoming words or concepts rather than the accuracy of those concepts.

Therefore, Jensen believes that AI can help conceptualize and theorize observed phenomena, but there is still a long way to go before AI can be effectively used for stock selection. Thus, Bridgewater's real focus is:

However, there are other ways to combine it with statistical models and other types of AI. This is what we are truly focused on, combining large language models with statistical models that are good at accurately describing the past but perform poorly in predicting the future.

By combining these, we are building an ecosystem that I believe can achieve what Bridgewater analysts are doing.

If this ecosystem is built, we essentially have millions of investment partners who are at an above-average level. If we can control the illusions and errors of AI through statistical data, we can accomplish a significant amount of work quickly. This is what we are doing in the lab and proving that this process is feasible.

Bridgewater's Approach to Investing through AI

If a system is built that incorporates AI and other technologies, how will Bridgewater utilize this system for investments?

Jensen believes that statistical AI and large-scale language models can complement each other and play the role of Bridgewater's "left and right hands" in investments:

Statistical AI can adopt theories, retrospectively examine whether these theories were at least correct in the past, identify their flaws, and provide recommendations on how to do things differently. Then, we can have a conversation with it. One advantage of large-scale language models is that they adopt a complex statistical model and discuss what it is doing. There are several methods to train a language model to do this. The way we simulate this scenario is that the language model can propose potential theories. This is not the most creative thing in the world, but it is a scalable theory, that's for sure. Again, large-scale language models are very good, but we must somehow adjust the language model, and we can use statistical data to control it.

Then, we can use the language model again to obtain results from the statistical engine and discuss them with humans or other AI, and report the discovered content, context, and types of theories. If the conclusions drawn are contrary to people's understanding, then more testing is conducted.

This is the exciting loop I'm talking about. So far, statistical AI has been limited because it focuses on market data. The benefit for language models is that they can better understand things that statistical models do not have.

For example, market statistical models do not have the concept of greed, but large-scale language models can almost understand the concept of greed - these models have read all the articles about greed and fear. Therefore, combining the two can produce thinking patterns similar to humans.

What does AI mean for human employees?

Over time, computers are capable of doing more and more things. Jensen believes:

What I mean is that today, humans have become accustomed to only fulfilling roles related to intuition and creativity, while we use computers to remember and accurately execute these rules. This is only halfway through the transition, and now we are once again experiencing a leap forward.

There is no doubt that AI will change the role played by investment assistants. Specifically, in the foreseeable future, we still need people to work around these things, and we still need time to build ecosystems for these machine learning agents, and so on.

Utilizing AI will become part of future work, and I think it is difficult to not utilize these technologies in any knowledge industry.

In terms of computer program coding, we are seeing significant breakthroughs in coding. Now, with the help of AI, people only need to know what they want to code, rather than how to code, which is a huge breakthrough. Therefore, a group of people who have not been well-trained or skilled in C++, Python, or other aspects can suddenly obtain what they want faster.

So suddenly, the combination of skills needed in the workplace is changing, and the way they change is surprising to many people, as it actually involves many knowledge-based work, such as content creation, etc., which people once thought would be replaced by machines in the distant future, but it is actually happening right before our eyes.

So the most important thing now is that there are too many changes, and it is necessary to have flexibility in the workplace and be able to utilize any tools. On the market, there are now various AI investment management tools available, and people are wondering whether in the future, humans will only need to entrust their investments to AI.

Jensen believes:

I think this will both lead to accidents and excite me greatly. Obviously, I am excited about the power of AI, and I think there are ways to make good use of it. However, AI will also make many mistakes.

Some foundations use GPT to select stocks, but these fund managers do not really have a deep understanding of AI and its potential weaknesses.

There is an example in the real estate market. The real estate platform Zillow uses AI technology to predict and evaluate house prices, and enters the market to buy houses that AI considers undervalued. However, Zillow has several problems.

First, although they have a large amount of housing data, these data are generated over a relatively short period of time. Therefore, despite having seemingly large data points, there is still a macro cycle that affects their evaluations.

Secondly, when it is actually a competitive market, they underestimate the disconnect between theory and practice.

Therefore, this is obviously a huge problem for Zillow. They have a significant impact on the real estate market and then encounter huge failures.

Returning to the stock market, very short-term trading can be said to be more suitable for machine learning because there is a large amount of data, and AI can learn faster from this data.

But on the other hand, in the longer term, the role of AI may not be fully utilized. Data is often like a person's lifelong heart rate data. You may think, wow, my heartbeat has been going on for 49 years, which seems like a lot of data, but when you have a heart attack, these data become completely irrelevant. Therefore, even with a large amount of data, it may still be misleading, and these issues will lead to significant problems with these technologies.

Therefore, people must understand these tools, what they are good at and what they are not good at, and combine them in a way that can leverage their strengths and avoid their weaknesses.

There is still a lot of work to be done on large language models, and of course, we can train them through reinforcement learning to ensure that they do not make known mistakes.

Is the market still dominated by optimism?

Jensen believes that the market is still dominated by optimism. He said:

The actions that the Federal Reserve seems to be taking appear to be more realistic than the market. When you look at the market's reaction, you will find that it is very optimistic.

But we have to be aware that historically, the market often tends to be overly optimistic.