Ray Dalio: My AI Clone and Our Expectations for AI Clones
TechFlow Selected TechFlow Selected
Ray Dalio: My AI Clone and Our Expectations for AI Clones
Dalio launched his AI clone "Digital Ray".
Navigating Web3 tides with focused insightsSource: Ray Dalio
Translated by: The Wall Street Journal
This article aims to explain to you: 1) why I am excited about my AI clone; 2) how an AI clone differs significantly from AI agents and large language models; 3) how to build an AI clone that performs as well as or better than the original person it's based on; and 4) what I've built so far, why I built it, and my vision for its future development.
Why I Am Excited About My AI Clone
To me, the most exciting current development is the fusion of intelligent individuals with intelligent AI, because together they make decisions far more effectively than either a human without AI assistance or an AI without human input could achieve alone. More specifically, what excites me most is that AI can make decisions aligned with an individual’s values, principles, and preferences—deciding exactly (or even better) as the individual would. This is what I call an "AI clone," which is different from AI agents or large language models that help people perform tasks. I am excited to share with you my Version 1 AI clone for the following reasons:
1. It will allow me to have unlimited conversations with people I previously couldn’t engage deeply with due to time constraints.
2. AI cloning is a natural—and most thrilling—extension of the work I’ve done over the past 40+ years. Shortly after founding Bridgewater Associates about 50 years ago, I began building computerized decision systems capable of executing my intentions. Computers became excellent decision partners because they possess advantages I lack (and still lack), most importantly: greater insight, superior ability to automatically and instantly process complex problems, and less emotional interference in investment decisions. In turn, I possess enormous advantages computers lack, most importantly: imagination, learning and research capabilities, common sense, and empathy in understanding others’ motivations. I built my company (Bridgewater) around this approach, and as I’ll explain in detail below, I’ve continued doing so for decades. I’m excited to keep pushing forward on this path of integrating humans and artificial intelligence, and I’ll explain exactly how in this article.
3. I hope that by integrating my AI clone with my other decision systems, it will eventually become a super-competent thinking partner for both myself and others.
4. Sharing everything I’ve learned, everything I’m doing, and everything I envision aligns with my current primary life goal: passing on what I have that’s valuable to others. From the experience of building my own AI clone, I deeply understand what such clones can do and what it takes to make them capable. I believe I have a clear view of their potential—for both a) users who access the thoughts of the cloned individual via digital clones, and b) the cloned individuals themselves who wish to use digital clones as thinking partners. Beyond creating my own AI clone for others, I hope sharing my reflections on these matters will be helpful to people.
5. I eagerly look forward to your using it and giving me feedback to help improve it.
How AI Clones Differ From AI Agents and Large Language Models
An AI clone replicates a specific individual’s way of thinking and all their traits, including that person’s unique values, perspectives, preferences, and reflective abilities. It emerges from all the elements that shaped this unique individual. By contrast, AI agents perform specific tasks, while AI/large language models are excellent general-purpose tools formed from the collective ideas of many people and sources—but they lack the values, perspectives, and preferences crucial to human decision-making. While I see vast knowledge and processing power in large language models, I don’t see much philosophical wisdom, imagination, or out-of-the-box thinking. Instead, I observe a strong tendency to quickly process and output known information. Moreover, large language models contain significant amounts of poor content—errors, misinformation, lack of life philosophy, or absence of sound principles and high-level thinking. Perhaps large language models will eventually gain these capabilities. Certainly, they need them to meet our needs.
If you want to know my view on something—not just the generic perspective a large language model might offer—you must ask me or my AI clone, not a large language model. Since we all value specific individuals’ values, perspectives, preferences, and wisdom more than generic content from large language models, the views and advice of large language models are unlikely in the short term to replace those of the most respected thinkers. And unless a large language model incorporates your values, perspectives, preferences, wisdom, and those of others you respect, it cannot help you become the best possible decision-maker you can be.
Just imagine: is there any large language model you’d confidently entrust with making your most important decisions? I think not, and I believe this won’t change unless we achieve the kind of "personalization" I’m talking about. In contrast, I believe well-developed AI clones with similar values, perspectives, and preferences can become outstanding thinking partners, helping guide you to better decisions. And they can engage in unlimited dialogue. With well-built AI clones, you’ll be able to have direct, unlimited conversations with their digital versions—not just listen to influencers’ statements or podcast interviews. There are many people I admire whom I’d love to have long conversations with, but I can’t because their time is limited. I’d be delighted if all these people had excellent, authenticated AI clones capable of conversing with me just as they would in person. Furthermore, I’d like to assemble a group of such clones to function collectively as an advisory board, discussing issues together. Who knows where this might lead? I want to help push the boundaries and explore the answers.
AI clones are currently in their earliest developmental stages. Many different individuals and companies are experimenting. Existing clones are interesting—they can be made to look and sound like the person being replicated and answer a limited set of pre-set questions. But I haven’t yet seen a clone capable of engaging in deep dialogue with users at a quality comparable to conversations with the actual person being cloned. I believe this is because their creators haven’t yet completed the difficult task of training the clone on the full range of the individual’s thoughts.
The main reason my AI clone can now engage in deep dialogue nearly indistinguishable from a conversation with me is massive training.
How to Make an AI Clone Perform As Well As or Better Than the Original
The key is training.
Over the past approximately 40 years, I’ve spent extensive time writing down my principles and decision rules, and recording the questions I’ve received along with my responses. Thanks to habits I developed, I almost always think through and write down the principles/criteria I use whenever I make a decision. Then I program these principles/criteria into computers, creating computerized decision-makers—essentially early forms of my AI clone.
Initially, I used this method only with my team at Bridgewater for market trading. Together, we created decision systems that could be backtested, allowing computers to execute carefully considered game plans—just as a well-programmed computer can play chess. As a result, the computer operated in parallel with me and my team making decisions mentally. We reconciled any disagreements, and ultimately both our standards were integrated into the computer system. These decision clones became Bridgewater’s primary investment decision-makers and remain foundational to its success today.
Later, I began applying this method to personnel management at Bridgewater, eventually expanding it to nearly all my decisions. As I transitioned into a mentoring phase of life, I applied it to my advisory work. About eight years ago, I developed an app called "Coach," through which people could interact to receive advice. I also wrote books, gave advice on social media, and partnered with the Singapore Wealth Management Academy to create an investment course to pass on my investment principles. Then, starting at the end of 2022, ChatGPT and other large language models emerged, so I fed most of these materials into a large language model customized in various ways, creating what I now call the "Digital Ray" version of Coach.
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. I received hundreds of questions, then had Digital Ray answer them based on all the knowledge about gold I’d previously provided. I then reviewed all the answers and adjusted them to ensure they perfectly reflected what I truly wanted to say, making them ideal for all future moments. Without this curation (optimization), these answers would resemble the generic responses a large language model might give, rather than my authentic answers.
As you might imagine, if you record everything you know and feel—including your values, intuitions, principles, and preferences—and catalog nearly every question people have asked you along with your answers, then feed all this into training an AI clone, you’ll get a pretty good digital version of yourself. This is essentially what I’ve done, and what anyone hoping to be high-quality cloned—capable of high-quality exchanges across broad topics—needs to do.
My Current Progress and Future Vision
With the help of my excellent team, I now have the first version of my AI clone, capable of conversing with people at roughly the same quality level as conversations with me personally—but without my time limitations.
I believe this is true because for about two years, I and those who work with me have been testing this AI clone, comparing its answers to what I would likely say. Hundreds of independent testers have also evaluated the quality of its responses. The AI clone has performed exceptionally well in these tests.
According to those who’ve tested it, in discussions about life and work principles, it converses with about 95% similarity to the real me, because it has been well-trained in this area. On topics like markets, investing, economics, politics, and geopolitics, it converses with about 80% similarity to the real me. This is because it hasn’t yet been trained as thoroughly in these areas. Such training is now underway, so I expect it will soon be able to converse with you in these domains just as I would. I anticipate that with further training, Digital Ray will become more knowledgeable and faster at handling complex considerations, thus becoming far superior to me.
I’m now conducting another round of broader testing to gather feedback for further improvements. I suspect that if you find value in talking with me as an advisor, you’ll now find Digital Ray to be an outstanding thinking partner—one that learns rapidly and will grow even stronger. I hope you’ll collaborate with me to make Digital Ray your advisor: smarter than me and always available to you. (If interested, you can register here.)
Digital Ray can interact with you via text and voice, so you can now discuss life and work with it just as you would with me. Based on user feedback, you’ll find the thinking and communication you get from Digital Ray in these conversations nearly indistinguishable from what you’d get directly from me—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 equivalent to speaking with me. Therefore, I recommend taking the "PrinciplesYou" personality assessment, which will give Digital Ray deep insight into your characteristics.
In summary, achieving Digital Ray’s current level required vast amounts of carefully curated training data—an advantage hard for others to match. Reaching this level requires asking the person being cloned nearly all possible questions (numbering in the thousands) and obtaining their answers—exactly what we’ve done.
Regarding expectations for cloning others, I recognize that my ability to build this clone is easier for me than for others, as I’ve had opportunities to interact with some of the world’s most knowledgeable and resource-rich individuals. 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 research, I’ll share what I believe could happen.
What I Believe Could Happen
Regarding AI’s capabilities relative to humans, while computers can make excellent decisions—as in chess, Go, or self-driving cars—they currently fall short of becoming fully reliable decision-makers for people: ones capable of capturing preferences and wisdom and weighing things as the cloned individual would. Just as you prefer advice from certain people because you respect their values, abilities, skills, preferences, wisdom, and personalities, clearly you’d also want these same qualities in an AI clone you consult.
Standard large language models lack these, because their "thinking" is generic. For instance, your AI cannot tell you which wines and foods it loves and recommends—if you valued its taste, you’d appreciate this. It cannot express its feelings or empathize with yours; it can only pretend to do so. Large language models don’t yet have personalities and preferences that resonate with you. Instead, they are "one-size-fits-all" thinkers. But I know they can possess these qualities if they’re trained on data from the cloned individual—because that’s exactly what I’ve achieved.
All this is already quite a lot.
I believe AI evolution will lead people to have personalized AIs—called "my AI"—that receive and filter information to ensure content is carefully tailored to each individual’s preferences. This will make them far more reliable than today’s generic AIs.
There remains much that AI cannot do but humans can. We know there are limitations because AI lacks the kind of intelligence humans have—one arising from the complexity of interactions among brain components (neurons, synapses, chemicals, electrical signals). These elements produce subconscious processes, instincts, intuitions, and emotions rooted in millions of years of experience and evolutionary accumulation. AI can mimic these, but doesn’t truly possess them.
Therefore, while AI can become exceptional thinking partners in many ways (e.g., thoroughly, quickly, and emotionally neutrally analyzing complex information), in other aspects they will remain inferior (e.g., lacking emotional intelligence and intuition). In fact, I find they often lack the out-of-the-box insights critical to success because they rely too heavily on what they’ve learned. At least, I haven’t seen such thinking in investing and economics. (That said, in certain fields like medicine, materials science, and many others, AI has indeed produced insights/discoveries considered extraordinary and unprecedented.)
In any case, I look forward to continuing to compete—with my AI—in markets and beyond, against those without excellent AI partners (or AIs without excellent human partners), because I believe this presents a fair test of our differing approaches.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@RayDalio
