Unaligned #5 (part one): Psychiatrist who uses AI

All right, this is going to be a special session for people who didn't see months ago. I did a therapy session with Dr. Munden, who you'll meet in a second, and she's a psychiatrist and she uses chat GPT to listen to her psychology psychiatrist therapy sessions. And she did that with me. I'm supposedly patient number one, and that really changed everything. I'm thinking about AI, and so this is really special for me to have a second look at psychiatry and AI. So I guess let's start with Dr. Munden. I usually ask, I start out with who are you? It's a thing I started back at Microsoft 15 years ago. So who are you?

Who am I? I am a psychiatrist by trade and I practice in a couple of different settings, which becomes relevant to our AI story as it's unfolded over time. But the way that I'm here today is through Vikas. I currently am the medical director for Psychiatry Outpatient Clinic in, based in California. But we're expanding and vicass has been really instrumental in helping us develop and incorporate AI in our real patient practices. So my, my role is seeing patients, but also I become a prompt engineer. I don't know how that happened, by the way.

We should have covered this. If you're not talking, if you can mute and then unmute when you're about to talk, that would be helpful. That way there's no sound coming in. Good morning. Who are you?

Good morning. Hey, I'm right here in Los Altos, very near to you, and I've been here in Silicon Valley for 25 years now.

Yeah, and Scott, who are you?

Hey, my name is Scott Solit. I'm been a police officer for 30 years. I'm currently a lieutenant at the Portman county jail, working with Dr. Munden, trying to work with our people who are seriously mentally ill.

So let's set this whole conversation up a little bit. I tried to in the intro, but I don't know if that'll work in the show. Let's try to set this up a little bit. Vikas, you built a system that listens to psychiatrists like Dr. Munden, and in almost real time writes highly detailed notes. I posted the notes on X.com that you did for my session months ago, and in 30 minutes it nailed all of my mental illnesses. It laid out what, what all sorts of things that it noticed about me because of our conversation. Is that correct? So let's start out with what is this system doing and what is the company that you're trying to build that's trying to do this. And then we'll have a fun conversation about what this means for psychiatry because it's deep.

What we're trying to do here is models like GPT, Llama 2, Bard, Anthropic, they're surprisingly good. And I think what most of us are missing in our human world today is that they think a million times faster than us. So that 30 minute conversation patient number one of mental health had with Dr. Munden, that 30 minutes was 30 minutes multiplied by a million times for GPT, you were patient number one. And after that, we have had around, I would say almost 200 patients on the system. The system continues to surprise us. The system continues to surprise us that how much it can infer from 30 minutes of conversation. And the way we are trying to understand this new invention that we have today, that we, the humankind has discovered, which Scoville, both you and I, we have talked, we think just like in humans, discovering fire one day. And then we built on top of it just that, the way the fire had million times more energy than what our human bodies could produce. This thing works a million times faster than my mind works. And when we are able to use the speed of this silicon intelligence, which is working a million times faster, and couple that with Dr. Mundan's mental health expertise, Lt. Solent's expertise in the jail system, we are able to fundamentally change how things are practiced. And that keeps me very excited. What I've been trying to do is if this system can think a million times faster than me, how do I use the system to affect the biggest change I can, Right? Because the Silicon Valley ethos is that you affect the biggest change, money, fame, and everything follows, right? The goal is affect the biggest change, money, fame, everything will follow. So after we met you four, four months ago now, there are a couple of things that happened mind blowing. First thing that happened was you were patient number one. We had a 30 minute appointment with you. Dr. Mundin then went out and treated 200 patients with the system. And with a lot of these patients, it was appointment number three. Appointment number four. Then Dr. Mundin came back and said, I want the AI to synthesize what's happened across the four appointments. So today when you use the AI EMR, what is you actually see a synthesis of the patient's past four appointments, where the new prompt that Dr. Mundin engineered and there's something to be said about that. Psychiatrists are better than computer scientists to engineer prompts. There's something that she does, which is unique. But the new prompts that she has engineered, it literally tells the large language model that, look, here is my last four appointment notes. What patterns are you noticing in this patient? And without using names, we have had the prompt come back and say that this prompts this patient's mental health deteriorates where she is with a significant other. There's a correlation between her being with a significant other and a mental deteriorating.

Can you tell us a little bit about the architecture of what's going on behind the scene before I start talking with Scott and Dr. Munden? Because I want to know what's going on when this thing is listening to me. A very common question. I, I talk about you guys with a lot of people, and a lot of people are like, oh, HIPAA rules, or oh my God, what do you mean OpenAI company is listening to your psychiatry session. That's very scary to people. But you built a whole architecture to make sure that private information doesn't go to the large language model. Tell us a little bit about the architecture that you built. It's a thousand lines of Python code, right?

Yeah, it's. And it's. We literally are building on the shoulder of giants. Like, the power that you have with thousand lines of Python code is just, is just too much. So the architecture that we have is we get your audio on our servers. And these servers are owned by the medical company. So it's HIPAA protected servers. These are not. And. And the audio is collected on the medical company servers. Then the medical company servers do the transcribing. So the transcript is collected on the medical company service. And once you have the transcript, then there. And again, there's a lot of open source code out there, AI code that we use to remove pii. And the way. And one of the things that happens when you're a pioneer in industry is that you come up with new conceptualizations, right? Which you talk to the world and then the world catches on with the conceptualization. So the conceptualization that our team came up, and I think it was led by Dr. Mundan, the idea being that there is PII. There are a lot of shades of gray. There's a black, there's a white. Okay, this information, we cannot recognize cobol. With this information, we can recognize cobol and there's a bunch of gray. So with this information, the people we have defined circles of trust. So circle one of trust, which is Kobel's wife, Kobel's children, they can recognize him with very little pii. Second level of Second circle, little less third circle. Hey, we tell you're a reporter in Silicon Valley, the first guy to do podcast, that's enough to recognize who we are most likely talking about. So we have these ideas of gray areas of pii, and we send it to the machine learning model and we tell it remove not only the blank areas of PI, which is very obvious, phone number, Social Security numbers, any medication you're taking, and so on, but also try and remove these gray areas of pii. And our models are continuously learning, and I think there's a lot of value being created and added over there.

For me, it's a much different problem because when they built this large language model that's running OpenAI, running ChatGPT, they read all my blogs for 10 years. Right. So it knows my pattern of language, it knows who I am, it knows things I've done, it knows a lot of people, it knows a lot about. So it, yes, you could start seeing, oh, it wouldn't take much data for it to go, oh, Dr. Munden's talking to Robert Scoville.

Right.

But I'm a public figure, so that's a little different than the normal everyday person who walks in the street. But even with me, it doesn't recognize me. Right, Right. It didn't put together. This is. Robert Scoville just said, this is patient one, Right?

Yeah. So once we run it through the AI model, we actually then take the anonymized transcript and send it to GPT and say, hey, can you recognize this person? We send it to Bart, we send it to ChatGPT, and so it's not, it's not able to recognize. So this is the pipeline we built. See? So again, the consideration we have is that the raw data collection has to happen on your own service. You can't use some other service to collect the raw data. For example, if I was using the ChatGPT tool on my phone to do the recording with, say, Dr. Mundin and the patient, OpenAI gets the data right there. So the raw data has to be collected on the medical company service, then you have to transcribe it, because still you transcribe it, I can't anonymize it. So you have to transcribe it on your own service. Once you've transcribed it, then you anonymize it. Don't anonymize only the black, but also anonymize the gray. Build a model to anonymize the gray, and then when you get anonymized transcript, then you send it out to a large language model. Now another fundamental thing happened in the last four months. The speed of progress is. I've never seen anything like this. And I'm.

Me neither. I, I have a list of 3,600 separate companies that are using AI, building AI. I've never seen this kind of wave.

We've never seen anything. Yes. And what happened with llama 2? Like the analogy that, that seems to work like for old faults like me been around for just so many years. We've seen this happen with the microprocessor industry, right? Intel was making Intel AMD were making the microprocessor, then there was the operating system and on top of that with the applications. So similarly in this we have the GPU vendors and then the operating system might be the LLM vendors and then we have the application there, right? So guys like us who are trying to innovate in the application layer, it's important that we don't start writing our own operating system because you have small teams, you have limited budgets, you don't have unlimited time, unlimited money. But what's happening in the operating system is that after Windows came out, it took many years for Linux to be similarly good. But llama 2 is catching up very fast. I literally woke up yesterday and I started downloading something from the torrent because a friend started release their model and I'm like, what's that like? Thank you for the gift on a Friday morning. Things are moving so fast that the state of the art commercial models are still six to nine months ahead of the open source models. But I do see as we're collecting more data, we have 200 patients now after you and thank you so much for doing what you did. If someday history books are written, 10 years from now will all mental health is being done using AI when it would be medical malpractice to do mental health treatment without AI? I hope the history books recognize you as the person who showed on Twitter what was possible, because before you we spoke to many patients internally saying hey, can we use you as a research subject to even show the world was possible. And nobody was comfortable with that. So I hope that the our historians recognize what you did talking about such personal subjects online. But I think the way the world is headed, if I was to look into the future five years down the line, I do expect me, us or other application players like us to be using models like Llama 2 using this French model, because we've also collected now 200 patients and I have something what, what the commercial LLM gave and I have Dr. Munden final edited note. And I can now use that as a training set. That look, LLM gave this. But the doctor, the. The biological intelligence wanted this, learn from this. And that becomes, as a way for me to RLHF, say, the French model.

Wait, you just used a few terms there. AI Insider terms. You called a human a biological neural network at the end. And that gets a little crazy. And that's how we work. We are a machine. Right. We're a very weird and unusual machine. But we're getting into brain computer interfaces where there's computing hooking up to our brain and hooking up to our machine, our human machine, directly. Right?

Yes.

We might talk about that later.

Yes.

But Let me introduce Dr. Munden and explain how you two met and what. How you work together to build this system. Because it's not. It's not just Silicon Valley entrepreneur Vic Vickris building this. She built most of this, of the prompts. And that's the key part of how the system works. So introduce her.

Yeah. Before I go in there, I. I was like, your feed is a gem. I literally was looking at your feed and I say, and I got Metaverse 2.0. Do you believe me now? I'm rephrasing it. But something like that, and that had Zuckerberg and Lex Friedman talking in 3D world. This world is getting weird very fast. This world is getting weird very fast.

That's why we're doing a show.

Yeah.

We're trying to cover the weirdness of this world. And so many people are afraid of AI. We were doing some interviews in stores yesterday. Somebody was walking around the store just going up to random people going, what do you think about AI? Everybody's. It's scary. It's gonna. This is something we don't wanna use. Right. Normal, everyday people don't know what's going on in AI and how it can help human beings. So why don't you introduce Dr. Munden and we'll have a conversation with Dr. Munden and then we have a special guest who is using the system in his work.

So I was looking for someone who will.

And look at the camera, if you can, by the way.

Sure. I was looking for someone who will bridge the gap between what technology can do and the real world. And I must have spoken with, around, or interacted with around 20 psychiatrists in my life as I'm trying to bridge this gap between what technically we can do and what the frontline responders in Dr. Mundane with someone she gets bored with. The with the mundane like she gets bored with if she has to do the same thing again and again. And she's someone who can literally be online with you at 10:00pm at night. And I literally had to ask her is your partner going to be okay? You're in Texas, there's two hours ahead of us. I don't want John to start hating me. She has that, she has that. That she's not lost at innocence. I think many of the psychiatrists that I meet, they're cynical. It's not going to work. That's their first. They start with the cynicism and Dr. Mundon starts with it will work. I know it will work. We'll just have to figure it out how it works. And since one out of 20 I've got to know in my life I've interacted with and what I've tried to do is just build a meta system using which she can express her creativity and she just continues to. She actually reminds me of a high school person who still has that optimism that yes, it has to work. Yes. We cannot have people languishing in jails because we don't know if they're competent to stand trial or not. We cannot. I cannot be treating people in 30 minutes giving them medications to alter their minds where I've seen them for 10 times over the last one and I don't remember. I am literally asking me to fake you. This is actually totally weak. I said half of the time if people knew what I did it should be considered malpractice because you're asking me to do things that are not possible for a human being. I'm just lucky to have one of my team. She is. She still has optimism. She's still willing to be foolish. She's willing to use a better software which has so many times stopped working in the middle. And that's my relationship with her. She is the frontline person who's taking it out. And probably this gets recognized too. I hope this gets recognized. I tell it. There are many guys like me. Guys like me are dime a dozen in Silicon Valley. Go to Starbucks right now nearest to my home there'll be five guys like me, well hacking on their laptops. But guys like Scoble to be able to talk about the sexual abuse on a camera. Guys like Dr. Mundan willing to use alpha level software six months ago, beta level software three months ago and probably release 0.01 today. It takes a lot of courage and I'm really glad that I have on my team.

Thanks Dr. Mundan, tell us what this system does just to start out with, when you're meeting with a patient and having a conversation, what's going on your screen and how does that help you do what you do?

So most of, actually all of my work with my AI partner is currently being done through telehealth platforms. So in the same way, if I was your psychiatrist, you're coming to see me, you would see me on the screen. I can see you on the screen. And then we would have a third window where our AI notetaker would be sitting on the screen as well. And so from the patient's perspective, and from my perspective, there's nothing different to this appointment. We are just talking to each other as usual in a psychiatric appointment. But what my AI scribe, or my AI partner is doing is collecting the data from that appointment and in under three minutes, by the conclusion of the appointment, generating an incredibly detailed consultation letter that captures the patient's history. It captures observations and exam findings of this patient's mental condition during the appointment. It can propose diagnoses and treatment plans and propose how best to bill the insurance company for this patient. So in the span of three minutes, all of that is done for me, that would probably take me 15 to 20 minutes to do, traditionally either typing on a computer or even worse, handwriting, which is, I'm old enough now to remember the days where we hand wrote notes. And so because I've got my AI scribe sitting with me in all of my appointments, I can capture data and think about theories of this person in a much more detailed way than I can do just by myself with my neural network.

When I met with you, we just did a 30 minute conversation. It was a conversation just like this, right? We just talked about life and things that are going on in my life and things that have happened or problems I'm having with my life. And it wrote, oh, he has PTSD on this topic. And he has, right? He, it went in and did very scientific, detailed psychological notes on me out of that, out of just one 30 minute conversation, which is very mind blowing to me. And by the way, your conversation with me and this tool showed me that this technology is a very powerful listening tool. And I use the listening tool now in the iPhone app all the time. Right? I'm talking to it all the time. And in fact, by the end of next year, we're going to have an assistant that we talk to all the time, an AI assistant.

Right.

Because it's so powerful at listening. How has this changed your life Though as a psycho psychiatrist.

It'S in terms of change. It's funny, when I look back at that original interview that you did with us, I cringe because when I clicked generate that note, I didn't know if it was going to work. I didn't know how well the system would perform for us. And you can hear my gasp when I see the note that's generated for you. And I said, it worked, it's there. And it was, what did you call it? Alpha? Not even beta testing. It was alpha testing. And so just in the span of three or four months since we last spoke, my life has been changed in that I believe a my AI scribe partner is always going to work for me. It's just always there for me. And so I can just sit and talk to somebody now. I can just sit and make eye contact with them. They can make eye contact with me. I'm not down on my computer screen typing away. I can really give this person all of my attention in those moments that I have with them because I know that my AI scribe is catching the data that has to be captured because lawyers want it captured, insurance companies want it captured. It's all important that it has to be captured. But I don't have to think about that anymore. I have to think about what a psychiatrist talking to a distressed person needs to be capturing and thinking about. So life changing in the past where I might have a 30 minute slot to talk to a patient, I might give them 10 or 15 of actual FaceTime and then spend the remainder of that 30 minutes writing a note. That's crazy. Now I can spend 28 minutes with my patients, two minutes generating my AI note and then I just do a quick overview of the note, approve it, correct it if I need to, and I'm on to the next person. Major, major life changing tool for me.

There's some problems still with these large language models. First of all, they have AI guardrails, right? They don't like to talk about some topics. And they come up, I call it the church lady. Some people call it a woke AI or something like that, right? But it's a AI guardrail that comes up and goes, hey, we don't want to talk about sexual abuse and we're OpenAI corporation. We don't like to do that. Sometimes that guardrail gets involved in the system and stops the system working. Or has that been fixed? Or tell us what's happening with the AI guardrails and does that. Yeah, tell us about what's going on with you. And AI guardrails. Yeah.

So our solution to the guardrail is a really cool result of just, I think, the relationship that Vikas and I have and we have with his team. Because I remember we were up at 10 o' clock at night just trying to figure out what the heck is going on with these guardrails. And it was a big issue because probably 60% of my notes at the time, before we figured this Guardrail problem out, 60% of my notes were not being generated. So I would get 40% of my notes, but not the other 60. That's not feasible. I can't see patients and run my practice that way. And we have come up with a solution, a bit of a secret sauce. And ever since we've done that, I now have a hundred percent note generation. We met the church lady head on and we now have an agreement with her.

I, I haven't met the, the guest we have on before Scott. And could you introduce him and what are you working on with him and how is your system changing his life?

I work in a couple of different settings and where I use AI exclusively at the moment is in the private psychiatric outpatient telehealth practice. And a big part of my week also involves seeing patients in jails across the country, including the jail where Lieutenant Soland is based. And I'm sure I've been talking to him about the AI experience I have in the rest of my professional life. And I know I've said, why aren't we doing this in jails? Why aren't we using this in jails? What's important to know about jails in the United States is that they really aren't some of the biggest psychiatric hospitals in the country. Many people detained in jails have severe mental health problems. And that's part of the reason that they brushed up with the law and they're now detained. But the issue is that, constitutionally speaking, nobody can go to court and have their case adjudicated unless they are clear thinking and they can understand the process. And when someone is severely mentally ill, like what we see in our jails in the United States, they're not capable of going to court and having a fair process. And so what I've been doing is talking a lot with Lt. Soland about how we can truly incorporate AI charting and AI tools into the jail setting. That's going to be a big lift and I think we can do it. There's my optimism, but that's the stage we're at now. Just in the past three to Four months we've moved past using AI charting in patient care to let's incorporate it in body cameras, let's incorporate it in all the GL systems that need to collect important data for patients. So I'm hoping that Lt. Soland is going to be our important partner in moving this forward.

And I've been in a jail getting a tour. I wasn't in because I was a prisoner, but I got a tour of San Quentin, which was quite life changing for me. You work with what groups of prisoners? Because I know I, I've seen AAA meetings in prison and other things like that where you're working on your mental health with a group of people. Is that what you're doing? You're going in and meeting with a group of people like that?

So yes and no. Mental health is found in every corner of any jail. And so we will meet patients where they're at. Sometimes that means talking with them in a veteran tank where there is a group of veterans doing group therapy. Sometimes that means meeting them in a competency restoration program. So those are probably the sickest of the sick psychiatric patients in facilities where it's a very small group and they get very intensive services to try and help them recover and get back to court. And then we'll meet people one on one in big dorm settings on TV there. They live in a dormitory. It still is jail, but it's both a one on one and a group experience with them. So every moment of every day in a jail is a different experience. Where we need the constant is, in my opinion, the AI partner to, to just keep that, that experience with patients as high quality as possible so they get the care that they need.

Where I was going is the computer can listen to three or four people talk. In fact, it's listening right now to us talking. Right. Descript is using an AI system to listen to us. That's similar to what you use, but different for video purposes. Right. The computer is really good about listening to five people talking and keeping them all separate and being able to track them over weeks. Right. So if you keep having a meeting every Monday with the same four people, it can track them and track what they're talking about over weeks and see problems that maybe you even missed. I think you even told me about a couple of things that you would miss that your AI assistant caught.

Agreed. One of the most important lessons that we learn in psychiatry when we're training to be psychiatrists is that we cannot base our opinion on a snapshot of the person we need to Base our opinion on the movie of their life. And what AI does for me now is it gives me that movie in 30 seconds. It. It finds patterns, it finds trends. It tells me this person's story to date based on everything that it's learned about this person through prior appointments and in a setting of jail. My prediction is that we're going to be able to use AI to create a very vivid movie about our patients, because I might input some data through AI about this person. Lt. Soland will have his input on the person. Our social workers, our nurse practitioners, there are many people in jail settings who are taking care of really sick people. And if we can use AI to consolidate everything that we all know as a team about this person, then we can provide efficient, cost effective, and high quality care for them.

And Scott, how is this system that these two people have built, how has that changed what you do in prisons?

Let's start with this first. I work in a jail, which is, it's where people are housed that are pretrial detainees. They haven't been convicted of their crimes. So one of the issues that Dr. Munden brought up was that they cannot have their case adjudicated until they are competent to stand trial. So in my particular facility, we house about 800 inmates a day. About 40, 45% of those inmates are labeled for some reason or in some way, they've been identified as having a mental illness. About 1 to 2% of them are seriously mentally ill and have to stay in jail until they're competent. And we're not talking about three months, we're talking about years, and they haven't been convicted of their crime. What the earlier talk about our jails are becoming the mental health institutions. That's the truth. And it's just thought it's not right, it's not constitutionally sound to keep someone in jail because they're not well enough to go to court. Working with Dr. Munn and realizing how AI can expedite the treatment and actually give us a better picture of the patient, not just for their immediate treatment and their competency restoration, but to be able to provide them with treatment after they leave, after their case is adjudicated, to keep them from coming back. Because we have a really high recidivism rate with these folks that are seriously mentally ill. And they, they just keep coming back time and time again. We have, when we're, we have a staffing meeting once a week and we look at the, at the patients, and the first thing that we talk about is how many times They've been in our jail. So when we're looking at 30, 35 times, 40 times that this person has been through our system, we're looking at over a 20 year period, they've been in our jail 30 times. That's a lot. Why is this happening? What are, where are we failing? Either inside or outside. So we've. I've been working not just with the sheriff's office, but also statewide. The vice president of Texas CIT Association. I work with, with the police officers and first mental health, first responders across the state, helping police officers identify mental illness and make better decisions on what to do with these folks. Instead of incarceration, getting them to a treatment facility, a hospital or mental health hospital. So we've. I've been working for the last 12 years trying to initiate that system of keeping them out of the criminal justice system and into treatment. But we still have those that wind up here because some of the crimes they are alleged to have committed are fairly serious. Being able to expedite this court process and get them moved through the system, getting them stable on their medications and competent to stand trial is huge. Constitutionally and personally, I don't like to see people suffer while they're in our jail. Ability to provide quality treatment, it's a tough one.

You're in, you're in Texas.

Yes, sir.

Getting a tour of San Quentin. Everybody in San Quentin murdered somebody, right? So it's, it's a very harsh place. And, and the people you meet, there's one building they didn't even let me get close to because the people are so dangerous in that building to, to interact with. Even the prisoners don't go to that building. They're like, no, we don't go. Those people are like, whoa, that's another level of mental illness. So there's a whole level of mental illness. How does this system that takes notes on people, how does it help you get to getting them through the system better or help them get their care that they need? Because I bet this helps a lot.

It's. It surely sped things up, right? But just being able to incorporate all the different parts we have, when we have people come into our facility, we are constitutionally obligated to provide them with medical, mental health and dental care. We have several different people in those fields that are talking with this patient and has an opinion or an observation on what's going on with this patient. Now when we think about some of the realities, Mental health is diagnosed and treated based on observed and reported behaviors. Not, not a blood Draw it's based on what people are observing. So when we can combine those observations and have, and have the, the note taker look at their past and piece all these things together is huge. Because that's what we need. You can imagine we have people that are seriously milio and then we have people that, that have really committed serious crimes that want to feign mental illness. So in particular settings, like when they visit with a psychiatrist, they're going to put on an act and role play to try to feign that mental illness and hopefully convince someone down the road that, that they shouldn't be convicted. When we have all these, all these other avenues of information come in, it's really easy to identify those and identify those people, we call it malingering, that are trying to feign the mental illness to get relieved of their charge. So it serves two purposes. Helps us weed those folks out and then it helps us better treat the folks that are really ill and need help.

In my session with Dr. Munden, it wrote all sorts of notes. What I was wearing, Right. What my appearance was, what my demeanor was. Right. Because if we talked about that, oh, it's, if Dr. Munden said, oh, you seem a little anxious today or angry today, it would put that in the notes.

Right.

Tell me, when you're meeting with a patient, what's it doing on the screen for you? What, how does it help you do your job?

I don't get to see the screen. That's Dr. Munden's side. I work on the security side of the house. I make sure I ask questions or I, or I speak what I'm seeing. So it can be determined. Now, AI can't see. So when someone is having some obvious rocking back and forth, wringing their hands, it's coming. But we don't have that just yet. But we have to call those things out and ask those questions. And it picks up on it.

Yeah, the kinds of. I, I have a special needs son and he does stimming. Right. So he's always running around the house clapping. That would be something we, we talk about. Hey, how's your stimming today? But in prison, I saw people yelling and doing all sorts of crazy stuff that kind of behavior would be put into, into the conversation. Hey, you were really causing a lot of ruck that gets put in the notes. Right?

Correct. Yeah. And we, believe it or not, we have a lot of folks in our system that are on the autism spectrum and they do stim. And it's being able to talk about that and ask Them we have some very serious stimming of banging their head on the wall, self harming, stuff like that. So being able to be able to speak to that and recall all the information. One of the other issues that Dr. Munden had brought up is the incredible amount of time it takes to take that encounter and reduce it to paper. And another reason we're talking about getting this into body cameras and using it on patrol, because the incredible amount of time it takes to reduce that to writing. We could really save money on personnel staffing. We could make things happen a lot faster, have it have better and more accurate documentation of what happened. We see a lot on TV about how. And it doesn't matter if it's a law enforcement officer working the street or someone in corrections or in a prison where they're abusing their authority. Well, with the help of AI, someone else translate or transcribing the incident that occurred, it would take a lot of that guesswork out or he said, she said so video cameras are great. The videos can't see everything. But having note taker in there, that's just transcribing. An accurate description of what happened during that incident would be huge.

If you're a rich person in the world, you get a, a weekly session with Dr. Munden and you have her on call and you can pay her fee. These people don't have that kind of control and probably don't have the same consistency over time that a, a rich person does with a doctor. These kinds of notes can notice things happening over time with various different doctors who are coming in and meeting with them. Right. For instance, I, I think Dr. Munden told me that she had one of her patients that the system caught that he was commenting a lot about his psoriasis. And while psoriasis is a indicator for depression, it causes depression in some people. So over time, as he was talking over weeks, he was talking about his psoriasis, it picked that up. Are you noticing the system that these two people have built is catching things like that over time with these patients?

Let's just be clear. I'm, I'm a police officer. So being able to understand or interpret the medical influence there, the psychiatry. Dr. Mundan has to explain a lot of things to me about what that means. But we're seeing that it's picking up a lot of, a lot of things that, that we don't and having that transcription there, doing it real time for you, like Dr. Munden was bringing up, those are things that we forget. If you spend 30 minutes having a conversation. I am not going to remember every word that was said. But now I have it transcribed for me and AI is going in and comparing those to past encounters and being able to pick up on those traits.

Yeah, I kind of remember what I talked about with my wife yesterday. When you're looking at a patient or a human over time, you can't remember what you talked about six weeks ago with that person. Particularly when you're like, Doctor. Dr. Munden. Dr. Munden, how many people do you meet with a. A week?

Yeah, I was, I was just going to highlight that. So in a typical office setting that looks like this, I might see 10 people in a day. But in a jail setting, for example, the Fort Bend jail, there are two towers, six and eight stories tall, I think. And there is also the booking department. And all of these locations house patients, hundreds of patients. And so when we walk into a unit that looks like we on TV with the jail cells and the clanging doors and all of that, there could be, I don't know, 30 to 50 people in one of those tanks, and each one of them will have a mental health need that they want to highlight to us as we walk past. So in a typical experience in a jail tank, I could encounter 30, 40, 50 people. And it's impossible to give those people the time that they need. And for me to also handwrite or type the data that I need to remember about this person to then go and communicate to the team about this person, this is where AI is going to come in and just sit there on my shoulder like my, my little scribe and do all that work for me so I can just have situational awareness in what's sometimes a dangerous setting, but also just really give this person the attention they need in a very distressing experience for them.

One of the things that people bring up when I talk about this is, oh, large language models. They hallucinate, they confabulate, they. They make wrong answers. I was in an Austin restaurant and I asked chat, GPT, oh, what should I eat at this restaurant? And it gave me six things. Right. Four of which were great. That highly recommended two had never been on the menu. It just made them up. Right. What, do you see that happening in your notes? It. It did. In my note. It made up my age. Right. How did that happen, by the way? You guys actually went back and looked at why it figured it thought I was 25 years old instead of 58. Right, Vic?

Yeah, it's cool. But we actually, our initial hypothesis was that was A hallucination. But then when we went through our debug logs, what we saw was that because you were a demo patient, one of admin assistants for Dr. Mullin had created you as a patient in the system and they had entered your date of birth, just a random date of birth. And that actually positioned you as a 25 year old.

It's LLM is where I was going with this. They're more accurate at listening to things and figuring things out than trying to tell you how and what to eat at a restaurant or something. Something like that. Right.

The way we conceptualize internally as LLMs are that they are pleasing assistants. They really want to please you. So if you say, hey, what can I eat at the restaurant? It wants to be like a really good, pleasing assistant and tell you things, even if it has to make them up. But at least, and I'll have Dr. Mundan take the lead and say that, but with the 200 patients we are seeing is that the hallucination problem is much less, at least in this context, we were not asking it to please us. The way Dr. Mundin has been able to craft the prompt is that he's able to put the right framework to the LLM that I have not noticed that Dr. Mundin, maybe you can talk more about it.

And this prompt engineering is really interesting. I want to talk to Dr. Mundin about the prompts that she wrote. I had dinner on Sunday with a computer vision pioneer and he told me about one of the prompts. He found that that makes the system hallucinate less. He actually puts in his prompts, if you don't know, if you don't know, tell me you don't know. Right. In other words, it forces the LLM to rate its own answers internally before it gives em to you and says, oh, I don't really know that answer, so I can't answer that question. Dr. Munden, tell us about the prompts that you've written to try to get around some of these problems and make sure that the notes are highly accurate.

The way that I talk to the LLM through prompting is similar to the way I would talk to a medical student or a psychiatry resident who's in training to become a physician or psychiatrist. So if a student went off and talked to a patient and collected data, synthesized an opinion and came back to me, I might say to the student, well, did you think about this person? And depending on that student's level of expertise and confidence, they could give me a very precise, appropriate answer about what they thought about this person, or it could just be chaos and they could just fall apart because they don't have an internal scaffold on which to hang their answers and their comments about this person. And so what I think we've learned about prompting an AI system is we need to remind it what it knows. And within the prompts, when I ask somebody, sort of when I ask AI, do you think this patient's competent for court, I will get a certain quality of answer that's quite different to if I remind AI what it knows about statutes and laws related to competency. If I remind it to think about that structure and those important details, I get a different quality of answer. So I think a good way to reduce hallucinations is to just know what I want from the system and then let the system know what I need it to do for me, rather than giving it the chance just to please me without really knowing why or how.

And this is an important, interesting segue, Scoville. We tried to get research patient number two who we could demo the competency restoration program, but unfortunately, you still are the only person who would volunteer. So we weren't able to get anyone else to volunteer. We spoke to over 30 people.

It's. It's a tough one. We argued about it before we did the therapy session, because when you're in a mental health situation, what you. What you talk about, what, what can hurt yourself or hurt others. Right. It can remind people of things. You can remind your. Yourself, you can have some real problems. And most people are not willing to go, go through that process in a public way. Right.

Dr. Munton actually had a partner simulate a patient, and she recorded a transcript, like a call, like a appointment. And then we had the system AI with her prompt. And this time we have made internal rules. We're not showing prompt on your show anymore. But she wrote that prompt and the AI gave her an answer. And she said that because we were doing it yesterday. And Dr. Mundan said to me that I feel the same how I felt back in May. Like, the kind of answer that AI came back with was just phenomenal.

To end this. Tell me a little bit about the company that you're building. What's the name you're calling, calling this the Palo Alto Clinic, right?

Yes. Now we are calling it the simplenotetaker.com. so what we realized, and again, Scoville, I hope they. A bunch of startup guys just like me will watch this. Right? And a bunch of us are facing the same problem today. Doc could you maybe stop the screen and a bunch of us are facing the same problem today and how this interview becomes a way in which the conversation starts on how to solve this problem. The problem that everyone is facing is that a jail system already has an EMR product. Mr. Soland, Lt. Solent already uses an EMR product to take care of his 800 people in the GL, even though four months ago the hypothesis was that we'll have a parallel to Clinic AI EMR and we'll be able to say, hey, use this, don't use your Jeff. Lucian Suland, though he knows that this is a better product, he's not willing to give up his existing product because his people are trained on it. It works. It clicks, but it works. If I go to the shiny new system, I don't know if it's going to work or not. How am I going to take 10 years of data from my old system, put it into new. My new system and tech guys like me will always come up with solutions. Oh yeah, I'll write a script and I'll do that. But lieutenant Solen lives in the real world where he, he can't just believe me a hundred percent, things will go wrong, things won't work. So there's too much of things. What I'm seeing is what I'm seeing in this startup environment, and I hope you're seeing it with your 3600 companies. That AI has given us new capabilities, but the entrenched players and the world is not willing or ready to remove these entrance players because what these entrenched players have done is that they have started to sell vaporware, right? Because they're saying, oh, we'll have this feature. We literally had vendor tell us we have a secret team working on our AI features. We can't tell you what that team is working on, but we're going to have all these features. So now being in this industry for 23 years now, I told Dr. Mundan that. Dr. Mundan, I'll have a thousand dollars bet with you. And I think we had this conversation a couple of days ago that the demo that we gave on with our patient number one, Scoble, in May of 2023, May of 2025, there will not be a similar demo available on the Internet. Everybody's already made this product on their website or all the big companies. We have AI, we have integrated, it works. But none of them will take it from a vaporware to a real working product. So I think this problem is faced by most startups on your list, the 3,600 companies on your list, we invented FHIR. We can do great things with it. But others, they don't want to give up how they were living because it's scary for them. It's scary for Lift and Solenoid to throw away his old DMR and to bet on a startup. So what are the two things we're doing? And I hope the two solutions that me and Dr. Mundan talking with Lieutenant Solent have come up with will be something, a pattern or like a playbook that will be applicable to other startups. And the playbook is this. It's really cool. The idea is that you draw a boundary among the bit around the other players. The way you draw the boundary around the other players is lieutenant Solent has an EMR in his jail today. He has a phone system transcription company that does the phone calls and then transcribe the phone call in his jail today. But he doesn't have a 360 product. He doesn't have a product that can take data from all these disparate systems and help a human make sense out of this data. So I think startups on your list will try and figure out that most industries have this problem. They use three or four different systems to get to do what they do. But AI can help us develop a 360 product. And the second thing is we have to think of in ways of how we can plug into the existing system. So we have actually abandoned palo alto clinic AI and now we just have simplenotetaker.com go to simplenotetaker.com click on start a visit, record the quant, generate the note and you get the note now if you take it. And now you can take this note to your own EMR system. And we are now going to try and replace your EMR because that was too much of an uphill battle that we were fighting. So now just call simple note taker.

Entrepreneurs are going to have this problem because there's so much change that we're trying to push through human beings and they don't want to change.

So connect.

Right? You're right on. This has been a remarkable hour for me. Thank you so much to the three of you for spending more than an hour with us. It's a real important thing to pay attention to, not just for psychiatrists, for everybody really. Because if you understand that the LLM that's at the heart of these new systems can listen and write notes and understand things, it's very powerful and it's going to help a lot of people. Right. I'm honored to be patient number one with you all. Is there thank you Sco any last thank you we need to cover we covered where to find it how do you charge for this you charge per patient use do you charge per doctor? Doctor pays 20 bucks an hour a month or something or.

Right now it's free. You just go to simplenotetaker.com and you start using it and we haven't figured out how to but the goal is to build something fundamental. The goal is to affect the biggest possible change and we don't know we don't have the answers to that yet. And I just want to say one thing. The world has invented fire. And what you're trying to do with your show is to tell the 99% of the world that look these 1% of people here in the valley have invented fire. Thank you for that. I think that's so important. Thank you.

Thank you so much.