AI in Healthcare, Grok’s Deepfake Scandal, and LeCun Leaves Meta

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: All right, welcome everyone. I'm Rob May, co-founder and CEO of Neurometric.

Ryan Eppley: And I'm Ryan Eppley, co-founder and CEO of RootAccess.

Anna Kirk: I'm Anna Kirk. I'm sales lead at Thread AI.

Rob May: And this is the AI in NYC show. And tonight we have a very special guest. We have Carrie Hodge from Dimer Health joining us, CEO. Carrie, do you want to give yourself a quick introduction?

Carrie Hodge: Sure. I'm Carrie Hodge, CEO and co-founder of Dimer Health. And we are a service enabled tech company where we are bridging that gap between the discharge and the follow-up with really personalized proactive AI-driven care.

Rob May: Cool. And how h how does the you have an app or you have a widget or you make like how how do how do people use your product?

Carrie Hodge: Yeah. So we have an app but we also have a medical practice. It's a virtual medical practice. So our patients can reach us and interact with us really anyway. But our AI is we call her Amy and she's through our app and tied into the patient's medical record. So, when a patient has a question, they can engage with Amy and we'll get personalized recommendations and feedback based off of their actual medical record.

Rob May: And so, was AI a factor in you starting this? Like, did you see these trends and say like this is something we have to go do because AI is a possibility now or this was a separate idea and you went down that path and like oh, and we should use AI because that's what everybody's doing.

Carrie Hodge: You want like the story story?

Rob May: Yeah, real talk.

Carrie Hodge: The story story. So my background is I'm a clinician. So I spent most of my career in emergency medicine and clinical leadership on a national level. And during COVID I had this like profound moment where we were triaging ventilators to get to hospitals across the country through the company I worked with and laying off or furlowing our emergency providers and people were staying because our volumes were down because people were staying home instead of going to the hospital. And then dying at home of heart attacks and strokes. And it blew my mind that we could get something from Amazon in two days and we couldn't get basic healthcare to people where they were at when they needed it.

And I took a leap into the startup world at that time and did a lot of work across the country where I was looking at ED diversion and readmission avoidance and operationalizing these projects where we got really great results. And it became so abundantly clear that we could absolutely change healthcare by filling the gap here and the only way to get it to the people that need it the most was scaling it with technology and putting it into their hands.

One of my co-founders at the time he had a tech background exited and retired back in 2013 and went ended up on a pathway where he got his EMT and paramedic and then went to PA school and was working in one of the ERs I was overseeing and he got his doctorate in using AI and technology to produce ED recidivism. So we reconnected and we're like with my skills your skills we are going to solve this problem. So AI was always a part of it and technology to really bring care to people where they were at when and where they wanted it was always at the forefront.

Anna Kirk: Carrie, talk to us a little bit about kind of the trust barrier that you might have to overcome with some of your customers kind of when they're getting to know how to use your product when they're getting comfortable with Amy versus, you know, talking in person to a health care professional, like what that might look like.

Carrie Hodge: Yeah, I think it's a really interesting dynamic with information and healthcare and trust because in healthcare is it's extremely vulnerable. It's extremely personal and I think in general people people want to be treated as in a four-dimensional way like they're not they're not a an algorithm of yes or no. They're they're more the health the status of health is way more complicated than that. And but I will say h we haven't had an issue with people not wanting to sign a consent or not want to utilize Amy because of the AI. They mostly love it and they say stuff back to us like I didn't know healthcare could be like this. Yeah, it's really amazing. I think the I actually find the issue with information about patient information a bigger a bigger concern for like health systems not necessarily individual patients.

Ryan Eppley: So, whenever I get to meet founders that name their AI, like walk us through how do you land on Amy? Why was there an art? Was there a science? Was it a gut check? Like, oh, I just think Amy is the right name for us. Like, how did you come up with the name Amy for your your AI?

Carrie Hodge: Well, I will not take credit for it. Our national medical director, Dr. Zang, came up with it. So, I want to give him full credit. But he we it's the name just worked because it's AI me and we're tying the information that Amy gives to people back to their personal most vulnerable information.

Ryan Eppley: Got it. Okay.

Carrie Hodge: So, I liked it. Okay. So, we did it.

Rob May: So, you've been you've been working on AI company for a little while now in the New York ecosystem. Have you got plugged in at all? How do you feel about the AI ecosystem in New York? Are you off on your own and not aware of much going on or do you feel like this is a good place to do a thing like this? Has it been easy to find talent? Like how how do you feel about the ecosystem?

Carrie Hodge: So I will say I am on the other side of the river, so maybe I don't I'm not I'm not like the ultimate best person to ask about this, but I have I haven't been super plugged into the community. I have gone to like a few different events and met some people. I would love I would love for that to be like more accessible because I think sometimes the journey of a founder is a lonely one and it's hard.

Rob May: We'll have to shoot a show from Jersey sometime.

Carrie Hodge: Yeah. Come over. You get the whole view of the New York skyline.

Ryan Eppley: That's a whole pitch, right? Yeah. Yeah. Go more into like your technology as a differentiation from I've even used ChatGPT to ask random things like hey like I'm feeling this way, right? I'm sure you've gotten the benchmarking or or the feedback from customers like how have you really gone about building a better health and wellness AI agent app however the product is kind of framed for you versus the traditional out of the box LLM like how do you really feel the differentiation is for your company?

Carrie Hodge: I mean, I think ultimately, well, so was it GPT Health just came out, which is, I think, amazing. And I love that people's information and the conversations and information that's important for them to have is closer to their fingertips. And that I think there's an we've been using this word a lot actually in the company. There's an inevitability to all of this. Like the AI is becoming so good, it's evolving so fast that it's going to be relevant to people. I think the thing that we do differently is that we are there's there we train our models to escalate to a clinician when appropriate and we review all of the models. So there's checks and balances and also we're catching those subtleties that might exist if maybe the AI gets it wrong or it's too conservative and you know it doesn't it's not able to see nuance. So I think that's a huge differentiator for us.

Anna Kirk: Maybe on that note I'd love to kind of hear in terms of how a typical user might interact with your platform, how they might interact with Amy. Like I'm a patient who was just discharged. What would my kind of journey look like?

Carrie Hodge: So, I can I'll give you like a typical journey with us and then what a conversation might look like with Amy. How's that? So, we you get discharged. We receive a ping that you got discharged and all of your records and your the information that we need for you is at our fingertips. We reach out to you almost immediately and you're on a video visit with a physician, an MP or a PA within usually of your discharge. And so at that point, that's like the starting point where from there we risk stratify you in terms of your likelihood to have complications and that's getting better and better with the data analytics that we do. And then we create a journey for you and part of that journey is downloading our app and having Amy where you can ask questions and multiple check-ins and subsequent appointments and you're were available for text and questions 24/7.

If you were, let's say you were discharged for something and you have a headache and you might say to Amy like, "Hey Amy, or hey doctor, I have a headache and I want to take Advil for it." Amy will say, "I'm looking at your medical record. It looks like you take Eliquis, which is a blood thinner, and Advil's contraindicated in that. Let's get some more information on this headache." And then you might give some more information. And at that point, the AI will say, "Let's let's hook you in immediately with a provider to, you know, have a live visit with a with a doctor or if it's basic common medical information, it'll just tell you what to do."

Anna Kirk: So Amy is mainly kind of the frontline like level one triage and then it's escalated to a healthcare provider.

Carrie Hodge: Yeah. Recommendation. Although I look at it like and I maybe this is a terrible analogy. I don't know. But I look at it like you know like in a have you ever seen bull fights? They've got like the multiple gates that the bowls come through. Like I look at AI is like the bowl and it's ready to go and the gates are like all the rules and regs that are still out there. But like when it's time I do think that AI will be doing the clinical visit as well. And I think that's coming faster than we think.

Rob May: Wow. And so how do you feel about, you know, going back to sort of like the ChatGPT health thing, which just launched. And there have been a lot of different perspectives on it. Some people have said like, oh, this is great. It democratizes access to medical stuff. Other people would say it gives hypochondriacs another thing to be worried about or it's going to put too many health resources and personal information in the hands of one company. Do you worry about any of these other criticisms?

Carrie Hodge: I mean, maybe this is a blasphemous thing to say and I hope they don't kick us out of network, but I think much of our health information that's out there is already in the hands of one company. So, it's I think like that fear is probably misplaced right now. So yeah. So for that for that I think that maybe it does give hypochondriacs something else to worry about or maybe it'll like assuage some of their concerns. Though I will say like one of our actually our head of AI, she her daughter was sick and you know with a stomach bug and she was like playing between the different platforms and she had said at one point you know she looks a little bit yellow to me and it was like a very alarmist reaction from the AI. So I think like there it's still it still needs training. It needs to get better and I think subtleties and nuance are really hard right now for these things to catch. So, we'll see. I mean, but I think it it is the step in the right direction and I think these things are learning very quickly. Would you guys use it? Would you put your information in? Are you going to use ChatGPT health? Anybody not going to use it?

Ryan Eppley: I guess I'm not as like guarded with it. I don't know, maybe I should be after hearing kind of like opinions on it, but I feel like I'd never ask anything too sensitive. I say like, you know, how do I get rid of this cold a little little faster other than saltwater gargles? What else I got up my sleeve kind of thing. I'd Google that though, so it's not like I would put anything different into Google that I would ChatGPT right now.

Anna Kirk: Yeah, exact. I was just about to say that like my my status quo today is, oh, I'm going to look this up on WebMD, which is arguably much more trustworthy than just whatever ChatGPT spits out, but I think where I do you know, I'm pretty like blasze with my information. Maybe shouldn't advertise that, but I think where I have questions is around like, you know, concept drift and like LLMs are notoriously bad about maintaining memory at a high fidelity over time. Doctors who can look at a physical chart, maybe they are too and it's just a little bit more buried. But I'm so curious kind of to hear your perspective carry on. Like do you think if an LLM is dealing with patient history that dates back, you know, 6 10 12 years, what that kind of journey looks like versus you're pinging ChatGPT health and you're like, "Hey, I have a question about the common cold." Like those feel like very different kind of use cases to me.

Carrie Hodge: Yeah. And I think that's actually a great point because even in regular medicine, like you could go into an ER and urgent care to get like a quick one and done thing taken care of. But if you have a complex medical history like most people eventually do, you want to have all of that all of that accessible for the decisions that will be made about your health. So I just one more note on the ChatGPT health. I think it's going to be really interesting when it ties into the wearables and all of that data and how that data is going to be able to recognize trends and early intervention to things. And then in terms of making decisions, I'll tell you as a clinician, like if I have a patient that has a really, really complicated medical history and they've had 10 hospitalizations in the last few years and they've got a diagnosis list of 20 different diagnoses and a med list that like goes from the floor to the ceiling.

To prep for that visit and then do a good visit on that patient where you're having a meaningful interaction and able to like access object objective information and then take all of that like cognitively process through it and come out with a good decision. You have five clinicians doing that, you're going to get five different opinions on how that outcome should be, you know, probably all in the same ballpark. So I think like the way the using these using the AI to condense all of that into something where you can get the important information in front of the you know clinician or in front of the AI clinician. I think it actually will help with mistakes.

Ryan Eppley: It almost like in my mind brings up this this thing that happens of like malpractice, right? And lawsuits with that. Is there like a an insurance policy that like AI is taking when it comes to giving out information about like remedies or what to do? Because I'm sure all practices have some form of medical malpractice liability insurance, right? Like walk us through how you think about that as a technology company.

Carrie Hodge: Well, we also have medical malpractice because we're also a medical practice. So, we do have that.

Ryan Eppley: Is that under like the same umbrella? Is the insurance company like aware it's under the same umbrella? Like, yeah. How does that work? That's actually very fascinating to me.

Carrie Hodge: So, we don't our AI doesn't make clinical decisions. And that's what the laws say now. AI cannot make clinical decisions. Although, Oh, what state was it in? That was Utah. Utah. Yeah. They just they just came out with an AI that refills medications, which is making a clinical decision because you know if you well I could go off on a tangent about this, but so right now the laws are pretty clear that AI cannot make clinical decisions. So I look at like AI right now as just like another person like another member of the team like sitting at the table with us. I think once AI starts to make clinical decisions, it'll be really interesting to see what happens with the malpractice and the insurance. I mean, I'm sure it won't be perfect, but if I think the way to really assess how good it is to look at complication rates with actual humans versus the AI. So, yeah, I don't think it exists yet, though.

Ryan Eppley: Yeah, I have to imagine that's like a market being made though, right? Like how do how do people ensure against like wrong information from AI agents? I have heard about that in other industries but not health yet.

Rob May: Yeah, we there was a company here they I think they pivoted but there was a company in Brooklyn that did did that for robotics and they basically initially just outsourced their like there was another insurance provider and they were front end for it and gathered the data and they were trying to build the data set on you know if you if I deploy a humanoid robot in your house and it you know knocked over something set your house on fire or what you know are you going to see the robot company or what are you going to do and so but yeah I guess they pivoted that might have been a tough market could just early.

Typically for those insurance businesses, you have to have so my my first company was a cloud backup business, right? We did backup for Gmail and Salesforce and stuff like that. And people would say like, well, I'm moving to the cloud, so I don't have to do that. Doesn't Google take care of that? And you're like, well, they do, but if Marian Accounting deletes your Google doc, like you can't get it back, right? Google's going to delete it from their servers and everything else. And so we would have to rely like we didn't take off until there were like some big incidences of cloud data loss. And there has there have been some big instances like OpenAI has had a pretty public horrible Yeah. Yeah. With people that discuss suicide with it. Yeah.

Anna Kirk: And that's another question too like the inherent nature of ChatGPT being people pleasing. Like what does that do when you need to give very candid direction that is not what the patient wants to hear?

Ryan Eppley: Yeah. How did you system prompt Amy's personality? Like who is Amy? Like what what do they feel like? I'm sure whether you're making that decision or your team's coming up with that decision like how did you give personality to your your AI agent?

Carrie Hodge: So I so Amy right now doesn't have a ton of personality.

Anna Kirk: So like I'll say that I kind of almost want that for my like healthare provider though provider. Yeah.

Rob May: You know I've met people like Amy. Yeah.

Carrie Hodge: Though I do think there's something there like the best doctor is going to be the one that you listen to, right? Like so if you can deliver care to people in a way they can hear it best. I mean I think that just goes with any sort of leadership role too. But yeah I think if we if Amy can morph which she doesn't do yet. If she can morph and speak language and inference and have cultural sensitivity and language level and all of that that's right for the patient that factors in their social determinance and their medical history and all of the information we can put in on someone I like that is really exciting to me I'm and I have no doubt that we'll be there yeah but right now Amy is pretty boring she's kind of new

Rob May: how do you guys any of you guys how do you how do you think AI and medicine are going to start to intertwine as let's say let's say medicine is fragmenting in terms of because because part of what you got from like social media and people able to do their own research and all that kind of stuff is you have you have a bunch of antivaxers now and then you have like the highest I think death rate from measles that we've had in 40 or 50 years or something like that and so what advice does Amy or like would Amy get like how do you treat that that you might have to give advice like your oh your kid needs the measles vaccine and people are like no I just that category of questions like how how are those going to be handled on in AI these AI related things

Carrie Hodge: I mean I think there is standard of care and that's what we are defaulting to with Amy so there's like the medical standards that people are trained on that you're taught on so you know that kind of cuts out all of that noise although I do think that like I think a great example of this is some of the studies that were done on women in hormone replacement therapy and menopause like they're finding now that those really weren't representative studies and actually they've they led the practice of medicine somewhat arai for many years to us poor women and I think like AI will be able to really help get help with the data analytics that come in and then give science scientifically sound like insights.

Anna Kirk: I'm so curious and I hate stop me if I'm going on too much of a tangent here, but your point Rob made me think a lot about like there's medical care and how it's delivered, but then there's also how people receive it. And I know from personal experience, I know from kind of talking to some of the medical professionals I know like you have to in some ways kind of angle your advice towards who you are talking to. If you are talking to a 25-year-old who doesn't go to the doctor very much may or may not kind of follow your advice and maybe you need to catastrophize a little bit more to make them take it seriously. I'm I'm curious if you know that's something that that you foresee Amy doing. Is that something that you think is necessary to deliver some of this medical guidance?

Carrie Hodge: Oh, that's like a whole psychology. We're kind of misanthropic on patients as a Right?

Rob May: Like well humanity is messy.

Anna Kirk: So well and it's also like you're if you are conducting any kind of science and you are assuming that your subjects will fall adhere to that 100% of the time, you're probably not being pragmatic about how you're delivering that.

Carrie Hodge: Yeah. I think it's an interesting point because like medicine is a partnership between like the provider and the patient and so that information has to flow back and forth in a way that's meaningful and I mean not every clinician is great either right and they don't always listen and you know I think one and I think actually in my like view of what's going to happen in medicine is and if you guys are not in healthcare you probably haven't heard this But those that have like we talk a lot about working at the top of license and utilizing resources under the most expensive resource optimally so that we can decrease cost without compromising quality.

And I think that you know you might like I have this wonderful endocrinologist and I had this virtual visit with her and it was like 3 minutes long and you know cuz she's like going going and the workload on doctors is so burdensome to chart appropriately is a lot and this like will take off a lot of that not valuable work that a like AI or a less expensive resource could take care of and hopefully elevate people to top of license so that you've got better higher expectations for the clinicians to deliver really better care, oversee some of the AI that's going to be emerging and have more capacity to give more meaningful time with people.

Ryan Eppley: It's on like the other side of medicine if if you will. But it's something that I see almost every week. A new AI startup is about drug discovery, right? Have you interacted with some of these founders or do you have a take on what's going on with like drug discovery and AI? Because I feel like there's got to be at least like four or five in just New York alone that have raised pretty sizable rounds and their AI for new drug discovery.

Carrie Hodge: Yeah, that's definitely not my expertise at all. But I do think that you know we changing how we treat disease and healthcare has traditionally been a very long process and the research could even come out and the resources come out and it takes 20 years to actually have that adopted into actual regular practice. And I think this is going to really accelerate who is going to be right for which drugs and who's going to have the most optimal outcomes. And I think like some of the stuff that is like way be like people way smarter than me are discovering is like using AI algorithms to direct mitochondria to you know do X Y and Z which and one of there's a founder that I met who is brilliant guy he he figured out a way to use AI to tell cells to regenerate muscle for musculardrophe and like it could cure the disease which is amazing like So, I that's about all I know, but it's I think really exciting.

Rob May: I mean, according to my Instagram feed, you don't need things like that because red light therapy fixes everything. Everything. So, it's like I see it for like every part of your body is there's a red light thing and it'll fix whatever problem you have, which seems a little crazy to me.

Carrie Hodge: Yeah. You have another prompt for your AI generated image.

Ryan Eppley: Yeah. Speaking of which, you want to get into that story?

Rob May: Yeah. So, one of the pieces of news we want to talk about is, you know, Grok launched a new Grok Grok with a K. So, we for those of you, let's just clarify, there's two Groks in the AI world. Grok with a Q is a piece of hardware. Nvidia bought that company recently. So, you're going to hear about that in the news. Grok with a K is the AI that's behind X.AI, Elon Musk, and all that kind of stuff. So, Grok with K launched a new image tool that apparently was generating so many inappropriate unapproved nudes of people per minute. And it it spawned a law in Wisconsin to prevent AI deep fakes of people and it spawned a bunch of, you know, backlash against Grok and then a bunch of other people sort of don't care. So, I mean, you know, how do how do how do you guys feel as like average citizens that somebody could generate a nude of you and put it out everywhere?

Anna Kirk: And I mean, that's a good question. Like, my my immediate reaction is probably not the best one, which is like, well, it's not actually real. So, like, but how do they know that, you know, like what what kind of and how is this law in Wisconsin going to be enforced?

Ryan Eppley: There's definitely laws for like name and likeness, though. So there is some type of like precedence for how the law and well there's revenge porn laws, right, which are which are also kind of like related. So you these probably all tie together somehow. And I know they're both at the state federal. I don't even know if they're at local as well, but I think there is like policies in place to handle this. I just think the amount and how easy it is has just changed the game. And now it's like, okay, well, if it's not a full nude, is that still like against it? Right? We were joking before the show about, you know, generating you in a bikini, right? Like that's what people are doing on the timeline now. That's what people want.

Carrie Hodge: Yeah. And the red light mask, right? Like someone's going to do that.

Ryan Eppley: I know. I know our followers will. Send us photos, please. We'll see that on LinkedIn tomorrow. But you know, I feel like there is policies in place for name and likeness to prevent this. It's just, you know, what's actually going to be done, right? Because people can create an account, do that really quickly, and then I think your point that the scale is the problem, right? It's like there that when you when you enable this kind of scale, the tools can't keep up. When it's small scale, you can do enforcement one off and have paperwork and whatever that process is, but otherwise it's hard. Yeah.

Anna Kirk: And it puts the burden on the victim or whoever kind of was was targeted of making sure that's enforced and when you're doing that at a massive scale, like how do you quantify that damage?

Rob May: Yeah, I think I think they're going to need legislation at a high level. It's the only way to deal with it.

Ryan Eppley: I thought I thought it was really interesting Elon Musk's reaction to it. I was like, well, it's not the technology that's the problem. It's the people prompting the technology that's the problem. And I almost think of it like behavior is really strange. Like when you're driving, you know, people have this it's even, you know, you're I think I saw an Instagram thing where someone was like if you if you treated walking down the sidewalk like people treat driving in their car and they're like slow down and like giving the person the finger and it's almost like there's this barrier of like no one sees me so I can like say whatever I want and I agree with you that there's going to have to be some sort of consequences is to Elon's point like it it isn't the technology but there is you know I think at some point some moral barri

Rob May: I think I think we've talked about this on the show before but if you're so I was an electrical engineering major in college and in the engineering department if you're a civil engineering major and you're building bridges and buildings and whatever you're required to take ethics courses because they don't want you to like let's skip on the concrete and save some money because the people die. We never had to take an ethics course.

Computer scientists didn't have taken an ethics course. And so, you know, we used to joke in the early 2010s that like all the best and brightest computer scientists in the world were at Google and Facebook trying to get you to click on more ads. Is that really how they should be spending their time? And like no, you know, there's been a lot of arguments if social media is good or bad. The computer the people that built those things never really asked the question, right, or thought much about it? And I there's not an easy answer to a lot of these questions morally and ethically, but we didn't have those discussions. And so I think that's a problem. I think it needs to start at the education side, which is like, hey, if you're going to go out and build these things, you do need to have a perspective. And I also I think Elon's perspective is a little bit naive. I mean, I had that when I was younger in tech. I was like, well, tech's neutral and it's how you use it.

But some tech lends itself better to bad usage and tech is overly obsessed with lowering the friction to do everything. And sometimes the friction is good, right? Right. When like when you used to have to write a letter to the editor instead of just tweet, a lot of people gave up, right, or weren't going to do it. And now that you can tweet, you can just say every dumb thing that comes to your mind. Mhm. And that that's probably not good. They're they're filters for a reason. Friction can be a good thing. And I So yeah, I'm not on Elon's side.

Anna Kirk: So we think we'll see more legislation pop up pretty quickly about this.

Rob May: I don't know pretty quickly. Tech has become a huge lobbyist against legislation and the current people that are sort of in charge of the eye side, you know, in the in the Trump White House are sort of very anti- any legislation. You probably see it out of Europe first. I think Europe's Europe's led the way on most data privacy and AI related things. So, that's where I would expect to see it and then America will follow suit at some point. They'll have to. Even like the revenge porn stuff. A lot of that happened when I was in college and you had the first time that you could post, you know, digital nudes of old partners online and stuff like that. And so that started to be an issue and that took a long time to fix. I probably like a decade. I knew a guy when I was in Louisville, the guy who started realitydefender.com. That was actually sort of the original use case was to be able to go send out automate and send out like takedown notices of things like that.

Yeah. Other news we've got Yann LeCun from Meta who's their sort of AI scientist and also one of the fathers of you know neural networks he a lot of people don't know but he worked on neural networks all through the 80s and 90s when people thought it was a stupid technology and was not going to go anywhere and he was eventually vindicated and we made a lot of progress in the last 10 years because of neural networks 15 years and he has now been saying for a couple years like it's capping out and we need to move something to something else. So he's left to start a company. I forget what it's called, but it's building world models. And the thing about world models is they use what's called Jeepa, which is like joint embeddings for something. I don't know, but embeddings in AI for the viewers are like you take something and you convert it to a space that computers can understand. So they're converting like physics and things about the world to a space computers can understand. So you can put that in these models as well and they won't do things or give answers that might violate common sense or physical principles or anything like that.

I actually saw the deck for that deal. And they're raising $500 million just to get started which is pretty pretty crazy. But I mean yeah he's he's kind of the godfather of all this kind of stuff. And this is going to be good for New York too because like you know he splits his time between Paris and New York because he was here for the Meta office.

Anna Kirk: I'm so curious and maybe you know more about this than I do Rob like what are the what are the logistics of actually training a world model like I think that so Gary Marcus' opinion article that he posted this week which is why it's relevant this week was around kind of the story behind Yann LeCun leaving but then also kind of what makes sense in the context of like should we have like models that are grounded in a physical and logical representation or we should should we have models like we do currently that are a little bit more symbolic and look to words and images and symbols. And he likened it to basically like either you're if you're a baby and you're learning how to crawl, you need to like touch your physical environment and figure out what gravity feels like in order to be able to do that. You need to fall a couple times. Whereas like we're not built to follow these statistical representations that are basically just like a corpus of different words like what what goes into training something like that. How like it seems like it would be expensive I guess is my point.

Rob May: So this this this this goes back to like the you know decades old like Noam Chomsky versus everybody else debate about how the human brain works right which is like do we are we born through through the years have we evolved some set of expectations about how the world works that we're born with and are we not and some things they know you know they they they test things like like if you if you drop a bunch of things right in front of a baby and then you do some kind of trick where you drop something and it suspends you see if the baby's surprised And that tells you if they understand dropping, right? Because they can't speak. So they do this at children. They they run weird tests like this of children of different ages to try to understand what they know and at what age. My favorite one they do, just as a side note, is they they'll put a dot on the kid's head and put them in front of a mirror. And they'll see like like it's like dirt, right? Ashes. And they'll try to see if the kid wipes it off the other kid or if they realize it's on them.

And that sort of tells them when they get a sense of self that like I'm the one in the mirror, it's on me. So they do a bunch of things like that. So I don't know the answer to that. My my guess is what what it sounds like from what I know about Jeepa is it's probably going to be trained with back propagation in sort of like similar ways. Maybe there's more reinforcement learning somehow. But I you know you could you could hardcode some of this stuff in you could use these RL gyms I think they call them sometimes gm right which is like a place where you can train re you know use reinforcement learning to train models to do different things. They've had a lot of success in video games, but it's hard to define objective functions outside of video games in the real world. You know, particularly because h humans respond to things on such different time scales. There's an immediate like I touch something that's hot. There's you know, you know, intermediate times. There's things that we, you know, we have long-term goals that are a year or multi-year. And so, how do you know, how you get feedback is very different. So, I don't know.

Ryan Eppley: It's interesting you brought up the gaming thing cuz you know growing up I played a lot of video games and those worlds that they like create takes artists you know three four five years right even like that's why the joke online is like we'll we'll get back to the moon before the next Grand Theft Auto comes out cuz creating the entire Grand Theft Auto you know world is a lot of artwork and details about the physics of how a character runs into the wall and they'll stop how a car drives around that's all physics- based you know game development. Just last week though on on X I saw someone live prompt an entire world within one prompt and it was generated within one hour and they could run a character around in like a game developer environment. So we're we're at the spot where like LLMs are being trained good enough to build games that have physics based reality. To your point, you know, how does it translate into like the physical world? That's why Nvidia spent all that money on Omniverse. You know, they're going all in on this. So if if Jensen, you know, one of the leaders in the space believes in it, I don't know if it's a year out or five years out, but it's definitely coming.

Rob May: Yeah. I you know, I mean, they're raising a lot of money probably to do the training. So we'll see. But I didn't see at least in the deck they didn't have any even clear plan what they're going to use them for.

Anna Kirk: It's like that's my question. Like as someone who's not as well versed in this particular area like what what is the application for this and what like words can only get you so far or so I hear but then what what comes after that

Ryan Eppley: got to be robotics that's you know you know physical intelligence it's how do you how do you give open-ended non-deterministic prompts to a robot and it actually performs accurately in the real world because right now there's all this like oh wow look at what this humanoid robot did but it's teleoperated almost every time they do a live demo. The reason why is because it's really hard to pick up even like a bottle of water without having all the coordinates down to like the you know parabola of like hey this is like the y-axis xaxis z-axis that your arm moves like doing that with an actuator in a robot is really hard without doing it step by step.

Rob May: C can we get one of the humanoid is there a humanoid robot company in New York? Can we get them to give us a humanoid for one of these shows who sits here and we ask it to do stuff while we're

Ryan Eppley: There was one in here, but I don't know if they're still in around. All right. I don't know. We'll see. That would be fun. That'll be our task for the week.

Rob May: We have a couple more articles to talk about, but Carrie, is there anything that worries you about AI? Like just you know, you're kind of working on the space and you're so you see some of it and obviously you seem pretty pro the healthcare things, but you know, are there are there use cases in you know, self-driving cars or other things where you're like that worries me? It worries me for society or like I would never use it or are you are you pretty excited?

Carrie Hodge: I think overall I'm pretty excited just like nonhealthcare related I think you know I think without checks and balances and news and image creation and social media like to me that's a little worrisome you know just the spread of information and what's correct and what's not and the difficulty telling the difference between the two but overall like you know what we're building I think you know the information in is so important to get good information coming out and like what we're building, we're really thinking about like making sure that we're collecting clean data as we're, you know, talking to patients and you know, putting it into their charts and now that's going to be part of what's forever for them. And so making sure that that part's good so that when we have information that comes out, there's recommendations, the real recommendations that are relevant.

Even like there's a really cool company that the barriers really are the FDA laws, but they take vital signs just by doing the face scan, the same face scan that unlocks your phone and it gets your blood pressure, your pulse, your pulse, like really good objective data. You don't even like need to be in person with someone anymore unless you need a procedure done. Like it's so I think that I'm really excited about that.

Ryan Eppley: How is that even possible? Like are they doing that through like image? It just vibes your pulse right now is Yeah. I don't I don't know. There's got to be some underlying technology that's not really it can take your your your heart rate just by looking at you.

Carrie Hodge: Yeah. Yeah. I thought the blood pressure was maybe not 100% accurate, but the pulse socks and the heart rate were really good.

Rob May: Yeah. Yeah. A lot of it was down to the quality. I mean, there's been a virtuous cycle, too, where the sensors have gotten a lot better over the last decade because as people, you know, as you as you have these devices like Fitbit and Whoop and all the other things out there, you know, they their suppliers, they go back to it and they say, "We need something now with a higher level of tolerance and they're, you know, then that expands the market and now it makes more sense for the supplier to invest research and well, how do we make a better I however they measure these things, you know, on your blood vessels with lasers, red light, I mean, just put red light back to red light robin blood as goes through. I don't know.

Yeah. What else do we have here? Maybe last thing to talk about. Did anybody do a review of some of the stuff that's going on at CES? Because the if you didn't, the m the most interesting thing I mean I watch CNBC every day. It's like my favorite channel and I particularly in the morning and they they were talking a lot about for a lot of years it seemed like CES was dead as a conference. Like it was there but like people didn't go. There was nothing novel. Consumer electronics were sort of okay, new phones and all that. And then AI's revitalized it with robotics and everything else, but there was a lot of talk this year about enterprise AI and CES was hot again. And that, you know, I think tying back to the New York thing, I mean, enterprise AI, I think, is going to be great for New York because So, and enterprise is here, right? Yeah. I mean, this is where so many of the Fortune 500 are headquartered and this is like we've always said sort of the applied AI capital of the world. So, I think it's the place you want to be to do some of that. So, I think CES was a good, you know, bell weather for where this might go.

Ryan Eppley: You have take any like notes on what's going on right now because the only things I've taken from is robotics based on CES?

Anna Kirk: Mhm. A lot of talk of world models also like that was very prevalent especially with Yann LeCun leaving that and then enterprise AI. I mean, I think the question is people know that enterprises need to adopt some type of AI technology. I think now we're just getting to a point where everyone is experiencing so much overwhelm around tool sprawl and like where do I start and how do I not lose sight of like the tangible business outcomes that have to come from this technology and because I have a mandate around like oh I'm like head of AI that means I need to deploy this technology. No, but like you also need to use it to accomplish real tangible business outcomes and how do you not lose sight of that?

Rob May: I mean, I could tell you, you know, we so at Neurometric, we work with a lot of companies because we work at their inference layer, right? So, we talk to people that are like, here's what I'm using AI for. Here's what I'm trying to I have this data. Here's what I'm trying to do with it. I'm running it through these models. You know, and then is it is it is it too slow? Is it too expensive? Is it not accurate enough or whatever.

I think one of the things that's interesting about the state of AI per your comment about the problems people are having this time last year I think the markets were really excited about AI and everything that was coming out but the companies were struggling right so we've talked about the MIT paper that 95% of these project AI projects fail well I mean that paper came out last you know in the fall that data was collected in the summer from projects from the spring whatever so like this time last year lots of stuff was failing and I think there was this big disconnect where the markets were like AI, AI, like this is awesome and companies are like we can't get it to work. Now I feel like it's actually the reverse where people are like man I don't know is AI a bubble or whatever but like what I'm seeing inside companies is people having real use cases that over 2024 and 2025 they really figured out

Ryan Eppley: well I think they found where it could be profitable even if it wasn't great. That's why there's so many call center AI startups like customer service ones because it costs a lot of money to hire a bunch of people to sit on phones all day. But if AI can do like a decent enough job, it's a huge cost center that a lot of these places are trying to figure out how to reduce costs in. So I think I think enterprises started figuring out like where can even a okay model work to save my company money. I think that's found out through all these 95% failure rates. But then they said okay like this 5% we're going to dump a money bunch of money in. I think that's what's happened.

Anna Kirk: Yeah, makes sense. I think the question is going to be like what is a dreaded edge case and like how do we control for that and put guard rails around that in an area where like we can't afford to get this wrong and that's something we talk a lot about at at thread is like AI is an important part of a workflow. It is not the entire workflow. So how do you craft a technology around your business processes if you need to pause for human in the loop intervention like how do you make sure that AI can accommodate that versus managing all these different point solutions like across customer service across like you know you want to use co-pilot but you also want to use a glean like how do all those systems work together and how do you make them speak together in a way that people in your organization can understand because it's also like it's a process problem. It's a behavioral momentum problem.

Rob May: So, we didn't cover this because it was it happened after we shot our last show of 2025, but Ethan Mollick who's at Wharton published a blog post called The Shape of AI, Jagginess, Bottlenecks, and Salience, and he talks a lot about where it's getting applied and why and where it's not. And so, yeah, so note to our producer Madi to like link to that in the in the show notes. It's a good You catching up with CES at all?

Carrie Hodge: So, I don't know much about CES, but I do think like in terms of how AI is being applied, I think you're spot on. And just in healthcare, it's scribes, though there were a ton of scribe companies.

Rob May: Totally.

Carrie Hodge: Yeah. But like I remember, you know, 10 years ago seeing a patient in the ER and thinking, okay, I got to go back to the computer and then I got to go chart this and it's going to take me forever. And then, you know, at that time it was like all the emojis were coming out. I'm like aging myself big time here. Sorry to like the world. I had a name, you know, and I was like, people aren't even writing words anymore and it's like all copy and paste and like getting the relevant information. Like why are we still writing charts? And I mean, if only I had thought of that, like you know, company back then. But yeah, so now like what we're looking to outsource is charting and call centers. So both of those things.

Rob May: Yeah. AI. Yeah. Makes sense. Makes sense. Pretty fault tolerant.

Ryan Eppley: Oh, what I took from CES for sure is there's like a clip about obviously Boston Dynamics has been a leader in robotics for years now, right? They're probably the first one when you thought about robotics 10 years ago. Like, oh, they look at that demo video, but they were kind of quiet for the last like couple. Yeah. They haven't mentioned too much publicly until CES about their humanoid and they reemphasized that they've been doing robotics for over a decade now. And I think it was like kind of a side eye towards like figure who's raised at what now a $ 39 billion valuation and hasn't shipped one commercial product. They'll tell you they have, but they really haven't if you ask a lot of the people insiders and Boston Dynamics came out and said, "No, we're like the real leader in robotics that have actually done this. We have all the expertise and now we just made our first humanoid because it's ready." I just felt like there was like a little tea being kind of, you know, sipped in the scenes of CES in that kind of space. That's what I picked up at least. I don't know if you saw that on CNBC or not.

Rob May: I didn't see that particular thing, but I have some friends at Boston Dynamics and I know their position on humanoids for a long time has been hands aren't good enough, battery life is not good enough. All the demos are staged. So, yeah, they're they're pretty pragmatic about the robotics they build. Well, Carrie, thanks for coming on today. If people want to learn more about Dimer Health, what's the what's your web address or what should they do?

Carrie Hodge: Yeah, they can just go to dimerhealth.com. Dime h and we you know you can email us or call us. Everything's right there.

Rob May: All right. And we always wrap up with a question about New York. So we forgot to send you this in advance, but maybe the question today for everybody is what's your favorite New York dive bar? And I'll start I'll say the place I go sometimes down by NYU is Red Lion because Red Lion they so Red Lion had and part of the reason I go there I mean it's pretty dingy but because it's a college bar right but they have live music every night of the year 365 days a year so you know when it's too late and you should go home and you're like where can we go there's live music and you're like you know it's relatively easy to get to from the parts of the city I frequent and so yeah we end up at Red Line.

Anna Kirk: I have to say I don't I don't think this qualifies as a dive bar. It's too good. But Smithfield Hall on 25th between 6th and 7th. I used to live right next to it. The owner of that bar has coached me through breakups. Like I would go for like every soccer game. They have the best burger in New York also. So I think this is actually I've used this answer twice now because I also think I used it for best New York burger, but it would have to be Smithfield. Yeah. You know, of all the places because there's got to be about 10,000 plus bars in New York, maybe more.

Ryan Eppley: That's also my favorite spot because I'm a huge soccer fan. I knew that because you're wearing that's where I watch the, you know, Euros. That's where I'll be watching the World Cup. And it is a blast. Even at 10:00 a.m., they will have it packed out for a Premier League game.

Rob May: I don't know if it's called a dive bar. It's, you know, it's that that's more like an upscale sports bar, maybe.

Anna Kirk: It's sticky on the floor, though. By New York standards, it's a dive bar. Maybe, right?

Carrie Hodge: Oh my god, this is like the worst question for me. I one, I'm from New Jersey and I'm from like rural New Jersey, so yeah, we like Anyway, and I'm also kind of boring. I'm a drinker, so I would definitely take recommendations though for totally great mocktail places.

Anna Kirk: Carrie, maybe we can swap that question out for another special one I have for you, which is you mentioned you at the beginning you had a lot of fun stories about the ER, about the ER. Do you have any that are appropriate standouts or well appropriate to We can always edit it out if it turns out not.

Rob May: Yeah, it's very hard for you to be too inappropriate on this show. Marketing tagline.

Carrie Hodge: One of the things I loved about the ER was that like people come in. I mean, I think in general I love that there's like this intense vulnerability and trust that's like developed right away and you get to help that person just in general. But I also really love doing procedures where a patient came in with a problem and then like they left with it fixed like really quickly. And so, you know, I'm not one to shy away from like gross things like plus and big wounds and stuff. I have seen a lot of the best and the worst of people and really some, you know, as soon as you say like nothing is going to surprise me ever, something new comes in. And but I feel like what story to tell? I think that none of us a sign of a good story. Well, yeah. I don't want to

Anna Kirk: We'll tell it at Smithfield Pub.

Rob May: Yeah. There you go. Great. Well, Carrie, thanks for coming on. For those of you listening or or watching on YouTube, give us a rating, give us a thumbs up, give us a follow. If you know guests that should be on, let us know. Watch the old shows and we'll see you next week. I'm stupid.