Source: Ray Dalio
Translation: The Wall Street Journal
This article aims to explain to you: 1) Why I am excited about my own AI clone; 2) The significant differences between AI clones, AI agents, and large language models; 3) How to build an AI clone that performs as well as or even better than the prototype individual it is based on; and 4) What I have built so far, why I built it, and my vision for future development.
Why I Am Excited About My AI Clone
For me, the most exciting current development is the fusion of intelligent individuals and intelligent AI, because the decisions made by combining the two are far more effective than what an individual without AI assistance or an AI without human involvement could achieve. More specifically, what excites me most is that AI can make decisions consistent with an individual's values, principles, and preferences, thus making decisions according to the individual's wishes (or even better). This is what I call an "AI clone," which is different from AI agents or large language models that help people accomplish tasks. I am excited to share with you the V1 version of my AI clone for the following reasons:
1. It will allow me to have unlimited conversations with people I previously could not engage with deeply due to time constraints.
2. AI clone technology is a natural extension of the work I have done over the past 40 years, and the most exciting extension. Shortly after I founded Bridgewater about 50 years ago, I began building computer decision systems that could execute my intentions. Computers became excellent decision partners because they have advantages I do not possess (and still do not), most importantly: stronger insight, greater ability to automatically and instantly handle complex problems, and less emotional interference in investment decisions. On the other hand, I have huge advantages that computers lack, most importantly: my imagination, my ability to learn and research, my common sense, and my empathy in understanding others' motivations. I built my company (Bridgewater) around this approach, and as I will explain in detail below, I have been doing this continuously for decades. I am excited to continue advancing and exploring the limits on this path of human-AI fusion, and I will explain the specific ways in this article.
3. I expect that by integrating my AI clone with my other decision systems, it will eventually become a super-capable intellectual partner for myself and others.
4. Sharing everything I have learned, am doing, and envision aligns with my current main life goal: to pass on what I have that is valuable to others. From the experience of building my own AI clone, I deeply understand the capabilities of AI clones and the conditions required to achieve these capabilities. I believe I have a clear understanding of its potential, whether for a) users who want to access the thoughts of the cloned individual through a digital clone, or b) individuals who wish to use a digital clone as an intellectual partner. In addition to building my own AI clone for others, I hope to help others by sharing my thoughts on these matters.
5. I eagerly look forward to your use and feedback to help me improve it.
The Difference Between AI Clones, AI Agents, and Large Language Models
An AI clone replicates the thinking style and all the characteristics of a specific individual, including that individual's specific values, perspectives, preferences, and reflective abilities. It is derived from all the elements that shape this unique individual. In contrast, AI agents perform specific tasks, while AI/large language models are excellent general-purpose products, aggregating the thoughts of many people and sources, but they lack the values, perspectives, and preferences that are crucial to human decision-making. While I see a vast amount of knowledge and processing power in large language models, I do not see much philosophical wisdom, imagination, or out-of-the-box thinking. Instead, I see a very strong tendency to quickly process and output known information. In addition, there is a lot of poor content in large language models—errors, misleading information, lack of life philosophy, or lack of good principles and high-level thinking. Perhaps large language models will eventually acquire these abilities. Of course, they need these abilities to meet our needs.
If you want to know my view on something, rather than the generic ideas a large language model would give, you must ask me or my AI clone, not a large language model. Because all of us value the values, perspectives, preferences, and wisdom of specific individuals more than the generic content provided by large language models, the views and advice of large language models are unlikely to replace those of the most respected thinkers in the short term. Moreover, if a large language model does not incorporate your values, perspectives, preferences, and wisdom, as well as those of others you respect, it cannot make you the best decision-maker you could possibly be.
Think about it: is there any large language model you would trust to make your most important decisions for you? I don't think so, and I believe this will not change unless the kind of "personalization" I am talking about is achieved. In contrast, I believe that well-developed AI clones with similar values, perspectives, and preferences can be excellent intellectual partners to help guide you to better decisions. And they can have unlimited conversations with you. With well-built AI clones, you will be able to have unlimited conversations directly with their AI clones, not just listen to influencers' speeches or podcast interviews. There are many people I admire whom I would love to have long conversations with, but I can't because their time is limited. If these people all had excellent AI clones, certified by them to converse with me as if they were themselves, I would be very happy. Moreover, I hope to have a group of such clones work together as an advisory board to collectively discuss issues. Who knows where this will go? I want to help push the limits and explore the answers.
AI clones are currently in their infancy. Many different individuals and companies are experimenting, and the clones that exist now look interesting because they can be made to look and sound like the person being copied and can answer a limited number of preset questions. But I have not yet seen a clone that can have deep conversations with users, with a quality comparable to a conversation between the user and the person being cloned. I think this is because their creators have not yet completed the arduous task of training the clone on all the thoughts of the person being cloned.
The reason my AI clone can now have deep conversations with you, with content almost indistinguishable from a conversation with me personally, is mainly because it has received massive training.
How to Make an AI Clone as Good as or Better Than the Cloned Individual
The key is training.
For the past 40 years or so, I have spent a lot of time writing down my principles and decision rules, and recording the questions I have received and my answers. Thanks to the habit I developed, almost every time I make a decision, I think about and write down the principles/criteria I use. Then, I programmed these principles/criteria into computers, creating a computer decision-maker, which is essentially the early form of my AI clone.
At first, I only used this approach with my team at Bridgewater for my market trading. Together, we created decision systems that could be back-tested, so the computer could execute a well-thought-out game plan, just as a well-programmed computer can play chess. As a result, the computer and my team and I made decisions in sync in our minds. We reconciled any differences, and the standards of the computer and ours ultimately merged into the computer system. Those decision clones became Bridgewater's main investment decision-makers and the foundation of its success. This is still the case today.
Later, I began using it in personnel management at Bridgewater and eventually expanded it to almost all my decisions. As I transitioned to the mentoring stage of my life, I applied it to my mentoring work. About eight years ago, I developed an app called "Coach," through which people could interact with it for advice. I also wrote books, gave advice on social media, and worked with the Singapore Wealth Management Institute to create an investment course to pass on my investment principles. Then, starting at the end of 2022, with the advent of ChatGPT and other large language models, I input most of these materials into a large language model that we customized in various ways, thus creating the current version of Coach, which I now call "Digital Ray."
While I previously followed this approach to succeed in investing and managing Bridgewater, I now do it because I believe I can help people. For example, a few days ago, I solicited questions about gold from everyone. I received hundreds of questions, then I had Digital Ray answer these questions based on all the knowledge about gold I had previously told it; then I reviewed all the answers and made adjustments to ensure they were exactly what I wanted to say, so they would be perfect for all future occasions. If I did not do this curation, these answers would be more like the generic answers a large language model would give, rather than my real answers.
As you might imagine, if you record everything you know and feel (including your values, intuition, principles, and preferences), catalog almost every question people have asked you and every answer you have given, and then input all this into the training of your AI clone, you will get a pretty good AI clone of yourself. This is basically what I have done, and what those who want to be high-quality clones, so they can have high-quality communication on a wide range of topics, need to do.
My Current Achievements and Vision for the Future
With the help of my excellent team, I now have the first version of my AI clone, which can converse with people at a quality roughly equivalent to their conversations with me personally, but without my time constraints.
I believe this is true because, for about two years, my colleagues and I have been testing this AI clone to see how its answers compare to those I might give. Hundreds of independent testers have also evaluated the quality of its answers. The AI clone has performed excellently in these tests.
According to those who have tested it, in terms of life and work principles, it is about 95% similar to talking to the real me, because it has been well trained in this area. On topics such as markets, investing, economics, politics, and geopolitics, it is about 80% similar to talking to the real me. This is because it has not yet been as fully trained in these topics as in the former. Training in these areas is now underway, so I expect it will soon be able to converse with you in these fields just like I do. I expect that with more training, Digital Ray will become more knowledgeable and faster at handling many complex considerations, so it will become much better than me.
I am now conducting another larger round of testing to collect suggestions for further improvements. I guess that if you find it valuable to talk to me personally as an advisor, you will now find Digital Ray to be an excellent intellectual partner, and it will learn very quickly and become even stronger in the future. I hope you will work with me to make Digital Ray your advisor, smarter than me and always available to you. (If you are interested, you can register here.)
Digital Ray can interact with you via text and voice, so you can now talk to it about life and work just as you would with me personally. According to those who have used it, you will find that the thinking and communication you get from Digital Ray in these conversations is almost indistinguishable from what you would get directly from me, and it does not hallucinate. I also know that if you let Digital Ray get to know you, you can have high-quality, personalized conversations with it, just like talking to me. Therefore, I suggest you take the "PrinciplesYou" personality assessment, which will allow Digital Ray to deeply understand your characteristics.
In short, to bring Digital Ray to its current level requires a large amount of carefully curated training data, which is difficult for others to match. To reach the same level, you need to ask the cloned individual all possible questions (at least several thousand) and get their answers, which is basically what we have done.
As for what to expect from others' clones, I know that my ability to build this clone is easier for me than for others because I have had the opportunity to interact with some of the most knowledgeable and resourceful people in the world. They also want to do this for themselves, but they can't because they haven't gone through the process of clarifying all their decision criteria. I hope they will do so in the future.
Now, based on my experience and what I have learned through research, I will talk about what I think might happen.
What I Think Might Happen
Regarding the capabilities of AI relative to humans, although computers can make excellent decisions, as in chess, Go, and self-driving cars, they are currently still lacking as fully capable decision-makers that people can rely on to capture preferences and wisdom and weigh things as the cloned individual would. Just as you prefer to get advice from certain people because you respect their values, abilities, skills, preferences, wisdom, and personalities, you will obviously want to see the same qualities in the AI clone you seek advice from.
Conventional large language models do not have these qualities because their "thinking" is generic. For example, your AI cannot tell you which wines and foods it likes and recommends to you, and if you think it has good taste, you would appreciate that. It also cannot tell you how it feels or empathize with your feelings; it can only pretend to do so. Large language models do not yet have personalities and preferences that resonate with you. Instead, they are "one-size-fits-all" thinkers. But I know that if these traits are input from the cloned individual, they can have these qualities, because that's what I have achieved.
All of this is already a lot.
I believe the evolution of AI will lead people to have their own personalized AI, which will be called "my AI." These AIs will receive and filter information to ensure that its content is carefully curated and suitable for each person's preferences. This will make them much more reliable than the generic AI that exists today.
There are still many things that AI cannot do but humans can. We know there are limitations because artificial intelligence lacks the kind of intelligence humans have, which comes from the complexity of interactions among different parts of the brain (such as neurons, synapses, chemicals, and electrical signals). These elements produce subconsciousness, instincts, intuition, and emotions that come from millions of years of experience and evolutionary accumulation. AI can pretend to have these things, but it actually does not.
Therefore, while they can be excellent intellectual partners that surpass humans in many ways (for example, thoroughly, quickly, and unemotionally analyzing complex information), in other ways they will remain inferior (for example, lacking emotional intelligence and intuition). In fact, I find that because they rely too much on what they have learned, they often lack the kind of out-of-the-box insight that is crucial to success. At least, I have not seen this kind of thinking in investing and economics. (That said, in some fields, such as medicine, materials science, and many others, AI does provide some insights/discoveries that are considered incredible and unprecedented.)
In any case, I look forward to continuing to work with my AI, competing in the markets and elsewhere with those who do not have excellent AI partners (or those whose AI does not have excellent human partners), because I think this is a fair test of our different approaches.
