← The Practical Planner
Loading summary
Transcript29 lines
- [00:00]
Sung Woo
Foreign.
- [00:12]
Thomas Coleman
To another episode of the Practical Planner podcast. I'm your host, Thomas Coleman. Here with me are a lot of people on the wealth team. We got Dave, we got Ann, and we have Sung Woo sun, who is basically help leading all the efforts with wealth and our AI. And so, Sung Woo, maybe best place to start is just a little bit about you and who you are before we dive into the whole world of AI.
- [00:36]
Sung Woo
Yeah. So I'm currently VP of Applied AI here@wealth.com and I lead all of the AI and machine learning efforts here. And previously I was a senior applied scientist manager at Microsoft AI, where I led AI machine learning investments such as Genai NLP search and ranking recommendation systems across products like Bing.com, edge Browser and the Windows operating system.
- [01:05]
Thomas Coleman
Okay, perfect. You definitely check the boxes on education and knowledge base here, for sure. So I think maybe best place to start this conversation is just like, what is the state of AI today? Because I feel like when we talk to advisors, you have the people who are like, I've never thought about using AI for anything. And then we have the people who are like, okay, chatgpt is pretty cool. I can type in some questions to get some answers. And then we have people who. It's like a full integrated part of their company already and how they do things.
- [01:30]
Sung Woo
Yeah. So I would say with AI, we are currently in a new industrial revolution. So in the last industrial revolution, water comes into machine. You light the water on fire, you turn it into steam, it turns into electrons, atoms come in, electrons go out. So in this new industrial revolution, electrons come in and floating point numbers come out. So floating point numbers is a type of data that computers and artificial intelligence communicates in. And Jensen Huang, the CEO of Nvidia, famously said last year, for the first time in history, we're producing floating point numbers at high volume. So here's the thing. These floating point numbers have value. That is because it's intelligence, artificial intelligence. So you take these numbers and you reformulate it in such a way that it turns into English, French images, videos, and here at Wealth.com, beautiful estate visualization reports and flowcharts.
- [02:31]
Thomas Coleman
I mean, it's just such an interesting thing. I was listening to a podcast actually yesterday on a walk, and it was just talking about the future and where AI is going, how it's going to replace jobs. And I think, you know, I'm very curious to your point here, like, how do you feel about the future moving forward? Like, does this mean advisors get replaced? Estate planning attorneys get replaced? Is This a tool that makes us more efficient, so we spend less time figuring things out and more time talking to people or, you know, where do you think we're going with this?
- [02:59]
Sung Woo
Yeah. So I think one of the common misconceptions is that AI can fully replace human experts. And so in reality AI can assist, but it lacks human intuition, ethical reasoning and nuanced judgment. So AI cannot replace trust building and financial advisor as an example, because clients value personalized human relationships as we do all and in law AI can draft documents, but lawyers are needed for negotiations, court cases and ethical reasoning. And another misconception is that AI is always accurate and unbiased. But in reality AI models are do inherit biases from the data that they're trained on and can make mistakes. So that's, that's another misconception. But I do think that you know where I see AI going in the next three to five years. Right. I do think that AI will shift from just simple assistance to autonomous decision making, but with human oversight. Right. So AI powered financial advisors, legal assistants and estate planning tools will move beyond basic automation into predictive, proactive decision making. For instance, AI will execute legal filings, manage portfolios and optimize tax strategies automatically with human approval. And like Hyper personalization will happen where AI will tailor financial legal strategies in real time. They will analyze behavioral financial patterns, client preferences and life changes to customize investment and estate and legal strategy in real time. And advisors will use AI generated financial and legal blueprints with automated scenario modeling. And all of which here at Wealth.com, we are investing heavily and for the long term as well.
- [04:52]
Thomas Coleman
So where do you see the advisor and how does the advisor still keep really the relationship on them? Because I think you're right. I mean like Google's existed forever. Pretty much everything that a financial advisor helps their clients with you could have figured out with enough time and research. Now AI basically just speeds that up, right? Like we can be more specific, you can get questions answered, maybe it's right, there's going to be certain times there's wrong. But like we talk about like how we're going to use AI but what does that actually look like? Is it really like, hey, we're over here typing in things or like how is, how is AI going to change and be better integrated in how we do things?
- [05:27]
Sung Woo
Yeah, I guess I could take an example of how we leverage AI here@wealth.com so imagine you have an ultra high net worth client that drafted a trust multiple decades ago that is scanned and is Hundreds of pages long. So it would take someone many hours to go through the document to understand what the disposition details are and what's potentially missing from the document. For optimal text strategy, imagine uploading that PDF doc and within minutes, AI summarizes key details and spits out a beautiful visualization report of how the wealth flows through the initial and successor trustees. And so that's what we have@wealth.com and we are working on so much more.
- [06:13]
Dave
Yeah. And I'd say, you know, it's so powerful and incredible just coming from that perspective when you're talking about the summaries. You know, when I was in advanced planning for financial advisors supporting, you know, a couple thousand financial advisors, they'd submit trusts and wills every day. And we didn't have an AI tool, and we'd have to manually review those and come up with summaries. And it would take a lot of time. And it wasn't a matter of comprehension. We knew what we were looking for and we knew how to summarize it. It's where to find the information and to be able to quickly and efficiently, at scale, be able to provide those summaries for advisors. And so that was difficult to do. The turnaround times were long. So, you know, from the estate planning perspective, just piggybacking what he's saying about, you know, the summaries, that's just been a game changer.
- [07:12]
Ann
And I would say, actually, you know, it's so interesting, Dave, that you talk about your experience in your prior role, you know, having to do these manual reviews. I think too often with AI, we think of, because it's, I guess, the limitation of the human imagination, at least on my part, is like, who does this replace? Right? Like, what role does this fulfill? And it's, in my world, coming from big law firms, it's like, oh, maybe that's like a paralegal or a legal assistant or even honestly an associate, right? And you're always mapping or trying to map AI's capabilities on top of what a human can do. But actually, that I think, is the wrong way to think about it, which is that AI can be so, so good at doing very kind of specific kinds of tasks better than humans can do. And it's about, like this revolution is about rejiggering what it is that humans do and where they contribute value versus where, honestly, AI is going to replace a function, a part of what an army of humans might do. Right? And so it's kind of reimagining what this world might look like from My perspective, it's so interesting because, you know, Sung Woo, I know we work, you know, together internally on all sorts of, you know, making sure that we have an ethical approach, a very studied approach, a thorough approach towards the AI models that we produce here and that you produce for wealth.com. but it's this idea of human oversight that keeps coming up over and over again. Can you tell us a little bit more about what that means? Once AI is able to synthesize, you know, so many thousands, if not millions of data points into what it summarizes, how is it possible for a human to kind of to bring some oversight into this? And I asked this question because, for example, I have a friend who works in marketing in the beauty industry, and she tells me she uses chatgpt all the time because it is just such a powerful way for her to be able to understand what the market looks like for her. But because she has been so long in this industry, you know, she has an intuition for what is right and what is wrong. So that that's the human oversight that she brings to this. I think an experienced lawyer would say the same thing. I have an intuition for what is right and what is wrong. Let's leave aside the guy who got into trouble in New York State a couple of years ago and who clearly did not know the case law he was pulling from and quoting was wrong. But you know, overall, like, if you practice in this industry, like, I know what is right and what is wrong, generally speaking, for trust and estates and tax laws, right. How do you marry? You know, there's that tension of like, we need to bring oversight, but then the younger generation coming through. How do you grapple with that issue, you know, of people not getting the chance to build that over that, that intuition of what is right and what is wrong?
- [10:17]
Sung Woo
Yeah, that's a really great question. And it's a really great question because AI models are only going to get better from here, right? But at the end of the day, what we need to realize, understand, is that AI models are probabilistic models based on probability of predicting what's the next best word for given the all the preceding words. So that's how basically the large language models work today. And so based on that premise, we need to understand that AI can hallucinate, quote, unquote, and what that means is generating false or misleading information. And so, you know, you talked about the famous case where somebody drafted a brief filled with fake case citations, which was actually done by ChatGPT. And knowing that it can invent case law statues or precedents that do not exist. Right, so take what we're doing at here, wealth.com, one pager executive summaries out of hundreds of page long document. What we do is we also give out citations for each of the bullet points where it is verifiable references that the end user can take. Advisors can take a look and actually go once clicked on to that very page, to that very paragraph where that particular information appears. And so that way it is very important. That is an example of building AI responsibly in a way that it is easy for the end user to go and verify, fact check whatever the AI has claimed. And then another thing to realize is that AI cannot provide legally binding advice at all because AI is not licensed to practice law and cannot establish an attorney client relationship. So we and any AI application should clearly disclose to the end user that it is leveraging AI model. And similarly the financial advisors who are using it should disclose to the clients when AI assistance user being used. And so AI should support research and drafting. But again, the final legal advice must always come from a human attorney. So I would say also when it comes to evaluating how well the AI is doing right, we always need to involve human in the loop. And what I mean by that is nowadays we leverage what's called the LLM as a judge type of methodology, which is leveraging AI to figure out whether or not the claims made by the AI is accurate or not. But then what the best industry standards practices is, is take a sample of that and then actually involve human in the loop to evaluate what the LLM has made the judgments on to verify that those are accurate. So that would be another way of bringing human in the loop to make sure that AI does not hallucinate.
- [13:19]
Dave
And so yeah, so you know, I like everything, I always bring everything back to basketball. And so whenever I'm thinking about AI, sometimes I think about, first of all, we had a conversation before about who your favorite NBA player of all time was. Is probably my favorite trivia fact about you. You want to say who that is.
- [13:41]
Sung Woo
Allen Iverson, AI, AI, his favorite.
- [13:45]
Dave
How perfect is that? But when I'm thinking about artificial intelligence, I, I think about it like end of the game, Michael Jordan's at the free throw line, games on the line, he's going to be taking the free throw. You feel very good that he's going to hit that free throw. You could basically walk away, but you can't. You got to keep watching the tv. You can't just Shut it off because there is that small percentage that he might miss. So you know, where, where would you say that fits as far as like what should we be expecting from AI as far as the probability that something like that is going to happen?
- [14:24]
Sung Woo
Yeah, I would say that's a really great question because to answer that question I would have to involve AI agent workflow. So what that means is nowadays what folks are doing in the, in the industry are building AI agents. So what, what it is is an agent that can do like a web research or summarize a document or go and find, retrieve like a client profile, do all of that in parallel. And so it really depends on the type of questions that you ask to the AI and how complex that question is. So it's really hard to put a number to, you know, the accuracy of an AI claim. So I would say always the best practice is to always verify, fact check the all the claims that are being made by the AI. So that is the obviously the safest way. But obviously the, an ideal AI application would do the fact checking beforehand before it even gets to the end user. Right. So one of the, one of the, one of an example is let's say an AI is explaining how there is a precedent case. We could build an API that checks the west law and check whether or not that exact precedent does exist. So that would be an example of doing a fact checking before it even gets to the end user. Human in the loop.
- [16:03]
Thomas Coleman
Makes sense. Earlier you were talking about advisors and how advisors, you know, can use this in their business. We need disclaimers, etc. But like where should advisors be looking to leverage AI in their, in their business? Like what is it good at that we can feel confident about using it for and where areas that we should maybe stay clear for now.
- [16:24]
Sung Woo
I would say yeah, like AI driven estate planning tools such as, you know, trust and will creation, trust management, tax optimization are all, you know, areas that we know that we are working on and other, you know, companies in the industry are working on, you know, automated compliance and regulatory monitoring is another good one where we would monitor, you know, changing tax laws, SEC regulations and estate laws, you know, like sentiment analysis to help with wealth managers anticipate market shifts or like AI powered agents which are like personal assistants to help financial advisors to improve client engagement. So all of those are areas that we can immediately do today to supercharge the wealth management space.
- [17:16]
Thomas Coleman
Yeah, and I think before. Ian, I know you have another question. I'll just kind of let people know how I'm Using AI in my business. So like there's a few key areas for me. One is obviously wealth, like reading estate documents is probably the one of the most impactful in my business. One because that takes the longest time. Like most other things I have to read or do are not 100 to 200 pages long. So reading, summarizing, making sure everything is correct is really important. And when my clients go through wealth to build theirs, it's also just really helpful to review the plan. Right? Did you pick all the right people? Blah, blah. I use it somewhat in marketing. I'm still working on learning how to train a model to speak like me because I don't really want like chatgpt, like written stuff. I want it to sound like me or not use it. I use it for note taking. So I use Zox, which is a really cool tool where it basically listens over. It'll put like five 15 minute segments. I can create specific agendas for after the call and it'll fit in the key things that I want. And then over time like I could search a client, I could look up like Thomas Koppelman and it would search all my past meetings, my CRM, everything to find answers to questions. We use it to like list a plan to read tax documents and then we'll make changes from there. And I would say that's probably all the uses I've found for AI so far that work for us.
- [18:33]
Ann
Yeah, I wanted to mention a couple of things so you know, there is so much promise from AI. One of the questions that we get a lot, as you know, you are you, our listeners, financial advisors and wealth managers are weighing, you know, what tools are out there, is whether or not our document creation at wealth is powered by AI. Currently it is not. And I wanted to make that clear because we see a ton of promise in that area. But right now they are very, very complex decision trees and that was very intentionally done. So that's something that I wanted to mention just in case it was confusing to our listeners. And the second thing is just as I in my capacity as chief legal officer have to do a lot of learning as to what tools AI enabled and large language model tools that we use at wealth, how our customers data is being used, how your client data that you upload on to the wealth platform is being used. We are incredibly careful about preserving privacy, confidentiality, not having those providers somehow train their own models using your data. And so those are the things that are really top of mind for me and they are probably also the concerns that you have about how wealth uses your data. And so sort of, generally speaking, Sang Woo, do you have, you know, kind of best practices for what our audience might need to look for or, you know, pay attention to as they use an AI tool?
- [20:20]
Sung Woo
Yeah. So I would say as a developer, like you said it, you must be really careful not to train your AI model with personally identifiable information. That is really dangerous. And the reason being is that there is a risk of the AI model spitting out the same like PII for another client. Right. And that is a really dangerous scenario. But as a user, I think one of the things that you should look for as a red flag if you're using an AI application is that, you know, if the AI does not give citations or verifiable references, that's a red flag because it does not give you the easy ability to go and check to that very page, that paragraph where you can fact check for yourself. But then if it gives a fake citation, that's even a bigger red flag. Another thing is that if the AI accepts your input as a text input, say it's a conversational AI, and if the AI can be jailbreaked, that's a huge red flag. So jailbreak. And we can probably. This deserves another podcast session. But by jailbreaking we mean you being able to inject certain prompts to the LLM in a way where you can force the LLM to behave in a harmful way, such as being able to retrieve all the code that's being used in the backend, or being able to retrieve all the prompts that is being used in the backend, or leveraging the large language model to create training data for them to use elsewhere for other purposes. So those are some of the sort of risks associated with leveraging AI applications. And if you can do those things, it is not following the standard responsible AI protocols.
- [22:24]
Thomas Coleman
Cool. What have we not talked about today that we should have asked you on AI? I know it's a very open ended question, but I always like to kind of wrap up and make sure we didn't miss something that you're like, everybody's got to know this.
- [22:38]
Sung Woo
Yeah, I would say maybe we could talk about the recent news about Deep Seq. A little more than A week ago, DeepSeek, which is a company based in China, released DeepSeek Dash R1, which is a reasoning model which performance on academic benchmarks is on par with OpenAI's latest O model. But the reason why there was such a buzz is because they have found an algorithmic breakthrough to train and run their foundation model at a fraction of cost on cheaper compute GPUs and completely open source methodologies, unlike OpenAI or Anthropic, which are closed source models. So what that really means is the foundation model layer, like the Gemini models, like the OpenAI models, are going to even get more hyper competitive. So Deep seq, OpenAI, Gemini, Anthropic, Mistral, Meta Alibaba are some of the notable players in this layer. And it's going to get extremely cheap to run state of the art AI models. And so OpenAI's 01 cost about $60 per million output tokens, which is a measure of how many characters or words that are being output from DLR's language model. And the deep seqs R1 costs $2.19 which is almost a 30 times difference. And so it's only going to get cheaper and cheaper every year. But the commoditization of the foundation model layer is actually great for application layer where the wealth.com is in. And so with the way that we are building our AI applications, we built it in a model agnostic way where we can easily swap in and swap out with the latest model foundation model. So our AI product Esther, will naturally become smarter and faster and better over time. And, and there's what's called Jevons Paradox that I want to mention, which is Satya Nadella, the CEO of Microsoft, recently called out, which is what happens when technological advancements make a resource more efficient to use. And as the cost of using the resource drops, overall demand increases, which causes the total resource consumption to rise. And in this case lower cost to run AI models in the long run will cause demand for adoption of AI to skyrocket across all industries like the wealth management space or other industries. Which means for financial advisors or home offices, AI will become an essential part of how we do estate planning in the future. And to accelerate the adoption of AI, Wealth.com focuses on developing AI systems that are trustworthy, fair and transparent and interpretable.
- [25:41]
Thomas Coleman
Super cool. Yeah, I think that was one of the things that have been like in the news. And like people like me are like, okay, that sounds cool, I don't have any way to verify this. And I was like, see on Twitter people we talk like, oh, are the numbers correct? Did they maybe mess up some of these numbers on how cheap it is? But then I saw like when they if it's corrected, it's still significantly cheaper than really anything else. So it's okay if they kind of messed up the numbers, but seems like it's going to be, hopefully an overall good advancement for, you know, us and everybody else still here in the US how do you view it in that way? Like, is there. Is that a risk to us as a different country or is it good for us?
- [26:16]
Sung Woo
No, it's only gonna get. It's. It's only good for us because we are only. We are. We are only leveraging what's the best model out there. And so it doesn't really matter if it's open source or closed source. And the foundation, the folks that are building the foundation models, the competition is only going to get tougher and tougher, which is good for us because for us, we are only leveraging the best in the class. Models that are faster, that are going to be faster and much cheaper to run, and smaller models too. Which means in the future, we could have AI on device, you know, for instance, that. That will be possible in the future. So AI will become much more accessible and much more cheaper and much better smarter over time. Okay.
- [27:06]
Dave
You know, I read A.I. if you're an estate planner, financial advisor, A.I. is not going to replace you. But someone who uses AI may replace someone who doesn't use AI So you got to at least incorporate it in some way into your practice.
- [27:26]
Thomas Coleman
Totally. Okay, well, cool. Sung. Well, we really appreciate you coming on and sharing all this with us. You're obviously the expert. We got a lot to learn. I know all the advisors here have been asking and trying to figure out, like, how do we better leverage AI, but how do we do it in a safe way? And I think you answered that question really well, so appreciate the time. And everybody, thank you again for listening. If you guys have any topics you want us to talk about, we're in the middle of drafting the rest of. Of this year, so feel free to reach out. But thanks for listening. We'll see you back here in a couple weeks.
- [27:54]
Ann
Thanks, everyone.
- [27:55]
Sung Woo
Thank you.