MY AI CLONE AND WHAT WE CAN EXPECT FROM AI CLONES
Published by Ray Dalio published on his X profile.
The purpose of this note is to tell you: 1) why I’m excited about my AI clone, 2) how AI clones are significantly different from AI agents and LLMs, 3) what it takes to make AI clones so that they are as good or better than the people they’re based on, and 4) what I have built so far, why I built it, and what I imagine will be built in the future.
Why I’m Excited about My AI Clone
To me, the most exciting thing that is happening now is the merging of intelligent individuals with intelligent AIs because together they have the power to make extraordinarily effective decisions that neither the individuals without the AIs nor the AIs without the people can make on their own. More specifically, I’m most excited about AIs that can make decisions that are consistent with an individual’s values, principles, and preferences so that they make decisions as the individuals would want them made, though better. That is what I mean by an “AI clone,” which is different from AI agents or LLMs that help people do things. I am excited to share V1 of my AI clone with you because:
- It will allow me to have unlimited conversations with people I previously didn’t have the time to have exchanges with.
- AI cloning is the natural and most exciting extension of what I have been doing for 40+ years.
- Since shortly after I started Bridgewater 50 years ago, I have been building computer decision-making systems that did what I wanted them to do. The computer became a great decision-making partner because it had strengths that I didn’t (and still don’t have), most importantly greater capacity to see more, greater ability to process complexity automatically and instantaneously, and less emotions to get tripped by in my investment decision making, and I had great strengths the computer didn’t have, most importantly my imagination, my learnings and research abilities, my common sense, and my empathy to imagine what is motivating people to do what they are doing. I built my company (Bridgewater) around this approach, and as I’ll explain in more detail below, I have been doing this continuously for decades. I am excited to go further on this path of pushing the limits of merging human and artificial intelligence in the ways I will explain in this memo.
- I expect that doing this will lead my AI clone to eventually become a super-capable thought-partner for both me and others by merging it with my other decision-making systems.
- Sharing all that I’ve learned, all that I’m doing, and all that I am imagining is consistent with my current main goal in life, which is to pass along what I have that can be valuable to others.
- From my experience building my AI clone, I know what AI clones can do and what it takes to get them to do those things. I believe that I have a good idea of what they have the potential to do for both a) those who use the digital clones to access the thinking of the individuals cloned and b) the individual being cloned who wants to use the digital clone as a thought partner. Beyond building my own AI clone for others to use, I hope that by sharing my thoughts about these things, I can be helpful to others.
- I am eager for you to use it and give me feedback to help me make it great.
The Difference Between an AI Clone and AI Agents and AI LLMs
An AI clone replicates a particular person’s thinking with all his or her qualities, including that person’s particular values, perspectives, preferences, and abilities to reflect. It comes from all that makes that person who they are. In contrast, while AI agents do specified tasks, and AI/LLMs are great generic products made up of the thinking of many people and sources, they lack the values, perspectives, and preferences that are important to human decision-making. While I see an enormous amount of knowledge and processing ability in LLMs, I don’t see much philosophical wisdom, imagination, or out-of-the-box thinking. Instead, I see very strong inclinations to take what is known and process it quickly. Also, there is a lot of bad stuff in LLMs—stuff that is wrong, misguided, and lacking a philosophy of life or good principled and higher-level thinking. Maybe LLMs will eventually get these things. Certainly, they need these things to meet our needs.
If you want to hear what I think about something rather than the more generic thoughts about that something that an LLM will give you, you will have to ask me or my AI clone rather than an LLM. Because we all care about the values, perspectives, preferences, and wisdom of certain people more than the generic ones that LLMs give, it’s unlikely that the opinions and advice of LLMs will soon replace those of the most valued thinkers. And an LLM won’t make you as good a decision maker as you could be if it isn’t personified with your values, perspectives, preferences, and wisdom, along with those of other people you respect.
Think about it. Is there any LLM that you would have confidence in to make your most important decisions for you? I don’t think so, and I don’t think that will be rectified until there is the sort of personification that I’m talking about. In contrast, I believe well-developed AI clones with similar values, perspectives, and preferences can be great thought-partners to help guide you to make better decisions. And they can have unlimited conversations with you. With well-built AI clones, rather than just listening to influencers and podcast interviews, you will be able to have conversations without limitations directly with their AI clones. There are many people I admire and would love to have lengthy conversations with, but I can’t because their time is limited. I’d love it if these people all had great AI clones that they certify can converse with me in the same ways they personally would converse with me. And I’d love to have a bunch of them working together as a committee of advisors who could collectively discuss things. Who knows where this thing could go? I want to help push the limits to find out.
AI clones are now in the nascent stage of development. A lot of different people and companies are experimenting with them, and those that exist so far appear interesting because they can be made to look and sound like the people they are copying and answer a limited number of preprogrammed questions. But I haven’t yet seen one that can have in-depth conversations with users that are comparable to the quality of the conversations users would have with the people being cloned by AI. I believe that is because their creators haven't done the arduous job of training the clone on all the thinking of the person being cloned.
The main reason my AI clone can now have in-depth conversations with you that are almost indistinguishable from the conversation you would have with me personally is because of the enormous amount of training that has gone into it.
What It Takes to Make AI Clones as Good or Better Than the People Being Cloned
It’s all about the training.
Over the last roughly 40 years, I have spent a huge amount of time writing down my principles and decision rules and recording the questions I have received and my answers to them. Thanks to the habit I got into, practically every time I made a decision, I thought about and wrote down the principles/criteria I used to make it. Then I programmed these principles/criteria into a computer to make a computer decision-maker that was essentially my AI clone in its earlier stage of life.
At first, I used this approach only with my team at Bridgewater for my dealings with the markets. Together, we made decision-making systems that we could back-test so the computer could execute a well-thought-out game plan, just like a well-programmed computer could play chess. As a result, the computer made its decisions in parallel with me and my team making our own decisions in our heads. We reconciled any differences, and the computer and our criteria merged into the computer system. Those decision-making clones became Bridgewater’s main investment decision makers and the foundation of its success. They still are.
Then I started using it with people management at Bridgewater, and it eventually spread to almost all my decision-making. As I transitioned into my mentoring phase of life, I applied it to my mentoring. About 8 years ago, I developed an app that I called "Coach" that people interacted with to get advice. I also wrote books, provided advice on social media, and created an investment course with
the Wealth Management Institute in Singapore to pass along my investment principles. Then, starting in late 2022, Chat GPT and other LLMs came along, so I dumped most of that stuff into training an LLM that we customized in a number of ways to create a version of Coach that I am calling Digital Ray for now.
While I previously followed this approach to succeed in investing and in running Bridgewater, I am now doing it because I believe I can help people. For example, a few days ago, I asked people for their questions about gold. I got several hundred, which I then asked Digital Ray to answer by drawing on everything I had previously told it about gold; then I reviewed all the answers and tweaked them to make sure they were exactly what I wanted to say, so that they are now perfect for all future times. If I didn’t do that sort of curating, the answers would have been more like the generic answers that an LLM would have given than my actual answers.
As you might imagine, if you recorded everything that you know and feel (including your values, intuitions, principles, and preferences), cataloged almost all the questions that people ever asked you and almost all the answers you ever gave, and then dumped everything into training an AI clone, you would have a pretty good AI clone of yourself. That’s essentially what I did, and it’s what needs to be done by the people who want to be well-cloned in to have quality exchanges on a wide range of subjects.
What I Now Have and What I Imagine for the Future.
With the help of my great team, I now have my first version of an AI clone of myself that can have conversations with people that are roughly as good as the conversations they could have with me, but without the time limitations that I have.
I believe this is true because, for about two years, the AI clone has been tested by me and those who worked with me to see how its answers compare with answers I would have given. Several hundred independent testers have also assessed the quality of the answers. The AI clone received excellent grades on those tests.
According to those who have tested it, it’s about 95 percent as good as speaking with the real me about life and work because it has been well-trained on my life and work principles. It’s about 80 percent as good as speaking with me about markets, investing, the economy, politics, and geopolitics. That’s because it hasn’t yet been as well-trained in my thinking about these subjects. That training is now happening, so I expect that it will soon be able to converse with you like I can in these areas, too. I expect that with more training, Digital Ray will become much better educated and much better at processing many complex considerations more quickly than I can, so it will become much better than me.
What I’m doing now is another, bigger round of beta testing to get suggestions to make more improvements. I suspect that, to the extent that you’d find value in speaking with me as an advisor, you will find Digital Ray to be a very good thought-partner now and that it will learn very quickly and be greater later. I hope you will work with me to make Digital Ray an advisor to you that’s both smarter than me and constantly at your disposal. (If you’re interested, you can sign up
here.)
Digital Ray can interact with you via both text and voice, so you can now have conversations with it about life and work just like you would have with me. Based on the ratings of those who have used it, you will find both the thinking and the communications that you will get from Digital Ray in these conversations virtually indistinguishable from those you would get from me directly, and it doesn’t hallucinate. I also know that if you let Digital Ray get to know you, you can have high-quality personalized conversations with it that are as good as a conversation with me. For that reason, I suggest that you take the
PrinciplesYou personality assessment, which will tell Digital Ray a lot about what you’re like.
In summary, to get Digital Ray to be as good as it is now required a huge amount of curated training that is difficult for others to match. Matching it would require the person being cloned to be asked all the questions that might be asked (a number that is at least a few thousand) and to provide answers to them, which is essentially what we did.
As for what to expect from others' clones, I know that my ability to build this clone was easier for me than for others because I spoke to some of the most knowledgeable and well-resourced people in the world. They would like to do this for themselves, but they can't because they haven’t gone through the process of making all their decision-making criteria clear. I hope they will.
I will now turn to what I think can happen based on both my experiences and what I have learned through my research.
What I Think Can Happen
As for what AI can do relative to the person, while computers can make excellent decisions like those that are made in chess, Go, and self-driving cars, they currently fail at being adequate decision makers that people can rely on to make decisions that capture preferences and wisdom and weigh things as well as the person being cloned would. Just as you would prefer to get advice from some people rather than others because you respect some people's values, abilities, skills, preferences, wisdom, and personality more than others, it is obviously true that you would want those same qualities in the AI clone of a person whom you’re seeking advice from.
Regular LLMs don't have these things because the LLM's “thinking” is generic. For example, your AI can't tell you what wine and food it loves and recommends for you, which you would appreciate because you think it has great taste. It also can't tell you how it feels or empathize with how you feel; it can only pretend to do these things. LLMs don't yet have personalities and preferences that are simpatico with yours. Instead, they’re one-size-fits-all thinkers. But I know that they can have these qualities if they are input from the person being cloned, because that’s what I’ve done.
All of that is a lot.
I believe that the evolution of AI will lead to people having their own personalized AIs that will be called “My AIs” that will take in and keep out information to make sure that they are well curated to suit each person’s preferences. That would make them much more reliable than the generic AIs that now exist.
What AIs won’t be able to do that humans can is a lot. We know that there are limitations because artificial intelligence lacks the type of intelligence that humans have, which comes from the interactive complexity of the brain's parts, such as neurons, synapses, chemicals, and electricity. These elements produce subconsciousness, instincts, intuitions, and emotions that come from millions of years of experience and evolution. AIs can pretend to have these things, but they don’t.
So, while they can be great thought-partners that are superior to people in many ways (such as analyzing complex information thoroughly, quickly, and emotionally), they will remain inferior in other ways (such as having no EQ and intuitions). In fact, I find that because they rely so much on what has been learned, they tend not to have the type of out-of-the-box insights that are critically important to being successful. At least, I haven’t yet seen this kind of thinking in investing and economics. (Having said that, there certainly have been some pretty out-of-the-box discoveries made in some areas, such as in medicine, material sciences, and a host of other areas where AI provided what some would consider incredible insights/discoveries that were not possible before.)
In any case, I look forward to continuing to partner with my AIs to compete in the markets and in other ways against people who don’t have great AI partners (or AIs that don’t have great people partners), as I think that’s a fair test of our different approaches.