AI Agents and Models — Today, and Tomorrow | LSI USA '25

Vishal Gulati  0:05  
Okay.

Amrit Panjabi  0:06  
Okay, good morning, everyone. My name is Amrit. I'm a digital health and healthcare analyst at CB Insights, where we cover everything in digital health, healthcare and Medtech. And with me, I have Vishal.

Vishal Gulati  0:18  
I'm Vishal bulati. I'm managing partner of Recode ventures. We are one of the few venture capital firms that is focused exclusively on AI, and probably the only firm that is focused on exclusively on AI and healthcare

Amrit Panjabi  0:30  
Absolutely. So today, we're going to be talking a lot about AI in general, AI agents, and what the med tech industry can learn from this industry. So maybe to start off, Vishal, how would you define the different versions or applications of AI from generative to agents? Where does that all fall in? Yeah,

Vishal Gulati  0:47  
I think first of all, the topic of today's talk, it should specify that we are talking about AI agents and AI models. I don't want people who've flown in from around the world to LA and expect different types of models and agents, so I think that the way we track this. So we started from looking at what is AI doing to healthcare and how healthcare is evolving, but very soon we realized that that was the wrong question, because we also had to start looking at how AI was evolving, because the speed at which AI is evolving is so fast that whatever projections you may have on what version of AI you were using last year, where healthcare will go, and if you don't take into account where AI is moving to you could find yourself in a very different place. So we then started to track AI much more closely to see, see what's going on. Just to give you an example, within a few weeks, if you go from chat GPT to deep seek r1 the price difference for API calls is about 130 times. So if you can, you can get you know API calls for $55 per million calls in on chat, GPT, but it's only sort of few cents. So, so what is really important for for us to understand in whatever way we are using AI is that it's it's moving and it's moving really fast, so that what you did before last year with AI is probably not even relevant today. Now that's really problematic in healthcare, because you you build products over a period of time, it can take you a couple of years to build a product, and then you also have to worry about regulation, or what your product is going to be, what what you're doing with it. So it's very, you know, most people find it really challenging. So if you just look at the historical perspective of artificial intelligence, it's, it's not near, right? It's, it's actually very old. So the the foundations of artificial intelligence, the theoretical foundations of artificial intelligence, were late in the 50s, and the people, you know, it was essentially the stuff that came out of the war. You know, when we were trying to crack the Enigma code, it was impossible for humans to do it. So they said, Okay, why don't we bring machines to try and solve that? And these are mechanical computers. Alan Turing and others were working on that. And then around 1956 there was the famous Dartmouth conference where people like Marvin Minsky and others, they were the these, essentially the founders of the theoretical basis of AI. So the reason I bring that up is because every time I would talk about AI, there would be someone in the audience saying, Oh, I was doing it in 1972 and nothing happened. And that's true, nothing did happen. So for so there were these winters that happened, and then sort of in the 1980s Geoff Hinton and others were quite instrumental in coming up with a new form, sort of using neural net networks and back propagation. And that was kind of the first thing, which is not now, it doesn't feel very unique, because back propagation is just trying to tell the computer when it gives you an answer whether it is right or wrong or how wrong it is. And then, just like you teach, teach a child to throw a basketball, you throw a basketball and sits too high or too low, and then they learn. That's how computers learn. And that that was sort of quite, quite impressive at that for its time. And it did give us some very, very valuable products. So if you go back and think about some of the early devices that were able to read ECGs, the US, post code, zip code readers, you know when you put a letter, handwriting recognition of your zip code and sending the letter to the right place. These were all products of neural networks using back propagation, CAD, for example, the earliest sort of forms of Radiology, AI, if you think. And then there was this huge sort of, what. Called the winter, long winter in AI, is that people just thought that this will continue to get better, but it stopped, and that's kind of where everyone just said, well, AI doesn't work. So we've had this phase several times of AI doesn't work, and so and then in when that happened, one of the problems was that there wasn't enough data when there weren't enough computers, and somehow everyone was no one was paying attention, and suddenly we had a lot of data, and we had really fast computers. So some kids started to say, Oh, hey, we have a lot of data. Let's try and try and use it. So that created deep learning. And this is sort of 2006 sort of sort of time horizon, and that sort of, I think that that was a watershed moment where people started to realize that this will do something more. And that gave us amazing products. So the quality of products that we got out of that, you know, there were similar applications, like ECG, like images, CT scans, but those were those machines could read these much better. And, and when, that's when we started investing. So we started investing in AI around 2016 2015, and that time we thought, Oh, my God, this is magic. And we thought we were so ahead of the curve. This is so cool and and that's sort of where a number of companies got started in, you know, from our experience, from our fun, we invested in companies that were doing mammograms, companies that were doing histopathology, companies that were doing drug discovery, all of those things. And then we thought, This is great. This is our great business. We're just going to find deep learning platforms that are going to do healthcare, and we're not paying attention to what comes after deep learning, and that was a mistake. So then the next generation is where we are now. So you know, deep learning cannot produce agent. Produce agents because deep learning can do only two things. It can do classification and it can do prediction. Those are the only two things that can do. What that means is that the best thing you have is a point solution. So you can train a computer to decide which is a famous image net thing. You've showed lots of cats, lots of dogs, and then all you can do is put a picture in front of it, saying, is this dog or is this cat? The computer can only answer dog or cat. Now she couldn't put a banana picture in front of the computer. Computer will probably call it a cat or a dog, depending on which is that banana is closer to a cat or closer to a dog or so that that is basically what the limitations of those things are, which means, if you're going to get it to do mammograms, you can only show it mammograms. And not only that, there are limitations, like all the images that you will use to train them will have to be of similar resolution, similar size, because you have to make that whole process very, very structured in itself, which also means that when you actually show it an image, when it's a product which is different, it doesn't know What to do with it, all right, so that's but what we are now, the generation we are in when we have these generative models, these models are much better, and these models have different properties. So the way all this try and describe it is when you're a resident and you somebody gives you an ECG, and when you are a professor, having worked in in cardiology for for 30 years, and someone gives you an ECG, you look at them differently. When you're a resident, you look at where the waves are disease, this elevated, is it not elevated? And you or if you're a radiologist, and someone gives you a chest X ray, you have a checklist. Is the trachea in the middle is, how's the lungs, how's the diaphragm? But when you are a senior professional, you don't do that. You look at it in one instance, and your brain is able to absorb everything that's in the image. And you know that this person has this and new forms of AI can do that so they can assimilate a large amount of data, and that's called their context window. And these new models have very large context windows, and in those context windows, they can look at many, many parameters. And that is really exciting one, because we could have more general purpose machines where you can look at whether this person has cancer or not. Cancer is what we did before. Now you can say, what does this person have that that's different, and that's sort of how where we are now. So maybe that was a very long introduction. I apologize. 

Amrit Panjabi  9:28  
Yeah, no. Appreciate the history lesson from from back in the 70s to now. So I think we've gotten a really comprehensive view with every other company, kind of having some AI element in their product. How do you Recode ventures evaluate what's truly meaningful and helpful in healthcare versus just something, just on top.

Vishal Gulati  9:47  
Sure, so. So one thing that is that is really important. First of all, is that you always want to think of it as what problem they're solving. So it is really easy. You can see you do get excited about tech. Technology, right? So I think that what so you would not want to just get a company or investment company, because they using the latest model. The question is, what problem are they solving? Is it a meaningful problem? And I think as venture capitalists, one of the things we always care about is what problem they are solving, is there a gap in the market, but for that problem to be solved, and if there is, how big is that problem? Is it really worth us investing in that company? So you always look at it from that lens. So when you look at it from that lens, you start to imagine what new things might come out of new platforms. So then you start to say, Okay, so these are the big problems that are not currently addressed by existing technologies, existing form of AI, because that's very, very limited. And then so what might be next? And so that's how you start to think about it. So it's always about the problem. It's always about how addressable that problem is. And then you kind of also look at how easy it is to get to market, and then you look at the technology. You can't invest just technology for technology's sake, absolutely.

Amrit Panjabi  11:10  
So what are those areas of interest that you're noticing now? Where are those gaps existing that AI can plug in and solve?

Vishal Gulati  11:16  
So I think that if you look at the more newer models, the transformer based models, which are agentic. A lot of the early applications have been in pharmaceuticals. And so if you think about the deep learning side of side of AI, that was very much friendlier to imaging, because imaging is actually very good use case for deep learning, and that's why it fitted very much in hospital settings, very much in in medical device industry. So there are about 900 FDA approvals now for radiology, and of which there is one company that that has 200 of them. So this is kind of very, very much concentrated within the medical device industry. And but when you look at the the the transformer models, most of the early traction has come from pharmaceuticals. And in fact, I think the medical device industry may be at the risk of being left behind. And if you look at partnerships that NVIDIA has done, they have done loads of partnerships with pharmaceutical companies. And when you and I spoke three days ago, we were saying that no, no medical device company has done any partnership with Nvidia. And I said that to someone yesterday, and they changed, they corrected me. And so thanks for that. So there is now one Partnership, which is, I hope, that start of, start of many more partnerships to come. So I do believe that there is so in general, what we have seen. So we invest both in companies. So we want these type of AI technologies to get to the patient, and we are agnostic about what path they choose. So whether they choose the path of of going via medical device industry, or they choose the path of going through pharmaceutical industry, we have companies in our portfolio that are using light field technology to do to image guidance, surgical guidance. But we also have companies that are developing new forms of ways, new ways of getting DNA into the right cells for gene therapy. So we do both. So we've seen a lot more traction in the generative side in the pharmaceutical industry. And one of the reasons is that lot of these generative technologies can actually generate designs and sequences, because remember, deep learning cannot generate anything. So deep learning can only answer a question, whether this is a cat or a dog, which is what it can do. So if you showed a deep learning platform 1000 molecules and said, which one was better, it might be able to tell you, but what it what you can't do to a deep learning model, say, design for me a molecule that will fit here. So you can't do that, but with generative you can. So you can say, Hang on, tell me what might be the best sequence for doing that. So that has received a lot of traction in the pharmaceutical industry, because if you're a company that has access to models, and you can design new drug molecules, you can patent those drug molecules, and you can create value for your company, and pharmaceutical industry is interested in those molecules. So that business model makes a lot of sense, but I think that the opportunity we are missing in medical device industry is medical device industry is not, still not looking at how to apply these models in their businesses. So it seems that device industry is waiting for people to come up with a product, and then they will see who else wants to buy it and where it fits in our existing franchises. So I think that's the that is the mindset. I'm being provocative. So please correct me if I'm wrong, and if you think that there is a different, different strategy,

Amrit Panjabi  14:51  
Absolutely. So what do you think the medical device can learn from how Pharma is operating with with AI? How can they do this to either supercharge their products or. To build a new dimension of products for the for that industry,

Vishal Gulati  15:04  
I think that there has to be more openness to disruptive innovation. I think there has to be some understanding of that the future products may not look very much like the current products. This, this ability, and drug industry has been better at it. So, you know, most of drug industry about 20 years ago was small molecules. Now, majority of drugs that are approved are antibodies or gene therapies and others. So there's been a a an acceptance by the pharmaceutical industry that the future will look very different from from what the past was. So, you know, the way the manufacturing plants you had that make small molecules will no longer we just have to chunk them and create new plants that make antibodies. And that was accepted by the pharmaceutical industry. And I think that there is a lot of install base in inside medical devices which is installed, which uses older technologies. And the tendency is to just pile more things into that, rather than and now there is some sense that I'm getting, where from Philips and others is that they want to come to the next generation platforms. So Medtronic has gone through that process, through their digital surgery acquisition and others. And I think that there, there is some sense that there will be new platform so that is really exciting if, if we go into a new generation of platforms for medical devices, then there's opportunity for founders who are developing new products to be integrated into those

Amrit Panjabi  16:27  
Absolutely, I think we're seeing a on the CV insights and a rapid explosion, especially in the clinical documentation, at least in the late of last year and then early this year, with some of the biggest rages in a bridge so that looks at more of The generative side of AI and more of a co pilot that augments the physicians journey to in speaking with their patients. As you see, all these agents start to build up, especially when we look at the hospital and we look at the agents kind of taking over in the back end, and maybe the CO pilots and the generative in the front end. How do you see that interaction playing out with medical devices such as endoscopy, surgery. Where do you see that overlay as these ecosystems start to grow?

Vishal Gulati  17:06  
So in theory, there is, there are a lot of opportunities. So if you think about a an operation, let's say spine surgery, to correct kind of scoliosis in a child. Now you could one way to think about that procedure is that you have a surgeon. The surgeon looks at the patient, they do their imaging, CT scans, they cut them open, and then they correct the spine. That's kind of one way that you can think about it. But when an experienced surgeon is doing that surgery, there are a lot of things they're thinking about. They have to think about how tall the parents of this child are to try and imagine how tall this person will be, because some of the corrections they do will be dependent, will depend on how tall this person might be. Will this will this person grow up and play sport and all of the those things? So what really happens is that we are not fully accepting the fact that experienced physicians take into account a lot more data that a device maker thinks that they need, and that data is implicit and we just accept, and that's why good surgeons are very good surgeons, and some surgeons are not very good is because of that, of their ability to assimilate and synthesize all that information, try and project into the future and see what the needs of the patient might be. Now, imagine if we could do this at scale and not be dependent on the small number of great physicians that we have who are able to do do this. And that is sort of how I think about it, that this data, a lot of this data, is available, and lot of these new platforms are really good. They have very large context windows, to the extent that you could put everyone's cradle to grave data into the system, and that system could could not only tell you what surgery to do. And I think the most challenging thing is that it might sometimes tell you to not do surgery, which is what device companies will not want to hear, but if they want to do surgery, what type of surgery to do, what type of implant to use, and what and to project into the future to see what kind of outcome this patient might have. So I think that all of those things are possible in the technology platforms. I think we just have to widen our horizon to see how much of this we can take into inside and the battle will be between the hospitals that have that data and the device companies that want to use that data. So there will be some some kind of interaction in the future,

Amrit Panjabi  19:32  
Absolutely. So we see this now with treatment selection in cancer and in oncology to help an oncologist decide if a patient would benefit from maybe adjuvant therapy or not. So you're kind of saying that this is going to extend towards the surgical platform where all this data can be

Vishal Gulati  19:47  
absolutely I can. There is no reason why it should not.

Amrit Panjabi  19:49  
Sounds great. It sounds like an incredible future. And I think AI has this potential within these gray areas to solve that and just provide as much data for the physician to make the right decision. 

Vishal Gulati  20:00  
Yeah, I think. But there is this just a word of caution that that the way AI is perceived in the media and what we hear about it, it doesn't mean it's easy. It may look easy that, oh my God, these, all these amazing things are happening. Most of things, those are happening. There's a very low bar, because most people are currently using it to cheat on their essays, right? So that's kind of how it is being used. And the bar there is relatively low. But when you start to build products and programs and then try and apply into systems that have hundreds of individuals involved, there's incentives, there's cost, all of those things. So it's not going to be easy, but what I'm here to say is that AI is not some some magic dust that you sprinkle and you get amazing products. If that was the case, I would be out of business, but, but what I'm saying is that there is an exciting future of of AI, and I think that we are it is all incumbent upon all of us to be part of it, and to try and build great products which will help patients

Amrit Panjabi  21:03  
Absolutely. so you bring up a great question about that data. So every time, for example, the generative AI, you're creating a very valuable data source between the patient and the provider, what are some of those red flags we should be aware of, so that we're not jumps and jump, jumping into a problem with the data getting weak, the data getting used for training, it towards payers and providers. What are some of those cautions that you would put as we track towards AI future? 

Vishal Gulati  21:29  
So I don't think that the didn't there are a couple of differences between how we access data for deep learning versus how we are accessing it for, for for the generative type of AI, because all those ethical standards, all those governance standards are, are are there? There are laws, there are regulations. They're different in different parts of the world, but those regulations have to be adhered to. I think that the phase of deep learning has done something really important, is that taught us all, all stakeholders, how to manage high quality data. It's hard to do, but it's it's something that now people know how to do. I think the difference is that what still needs to be to be thought through is who owns the product. Because these these machines are able to build new things. So, you know, it's these are not just reading Shakespeare. They're writing Shakespeare. So what happens if an, if a generative AI reads all of Shakespeare's work and then writes another version which is, which is different, and what, what that means? And I think, I don't think we have confronted those, those issues yet, but I'm confident that we'll be able to deal with them great.

Amrit Panjabi  22:47  
Well, we have about 15 minutes. Are there any questions or directions that we could. Thank you. 

Audience Question  22:53  
Thank you very much. So when we talk about AI agents, so I'm reflecting on the other side of monetization within Medtech. How do you see, let's say, let me give an example, a particular startup is working on building an AI chain for one specific use case. How do you see the trajectory the growth of the company towards the point of an exit Co Op going to be the potential, you know, acquisition partners? Are these going to be the companies who are owning products without AI or is it how? Is how these AI agents going to be valued at some point? How is the future going to look like for such a product that's being built like an AI agent, when you talk about a...

Vishal Gulati  23:44  
Sure. So I think that the we haven't seen a lot of AI agents being implemented yet, and the few that I have seen, I've seen in two categories. One category is where, where these agents are analyzing patients data and updating what the possible scenarios for this patient might be at at in the future, and as new data gets added to their EHR, they're able able to to do that. Another company, which is in our portfolio has an agentic system which is able to talk to a parent of a child with rare diseases and is able to guide them through question a question and answer session, and they they believe that within 20 Questions, they're able to diagnose rare diseases, which currently takes several years for a physician to diagnose. And the reason why these agents can do that better than physicians is because these rare diseases physicians see very rarely as as as would be the case, and so they don't have the expertise in being able to ask the right questions to get to the right answer. But using these large language models, they are a. Able to build that. So for those companies, again, the exit for those companies will be who benefits from that. So if that information currently, their customers are companies that have rare disease products, so they're able to help patients get diagnosed, because without diagnosis, there is no treatment. So to imagine a future for them would be to see whether which companies benefit from giving these companies access to patients. So that's how I would think about 

Amrit Panjabi  25:27  
Yeah, that's it.

Audience Question 2  25:29  
Thank you. One of the areas in bed tech that's had a lot of innovation with AI is the whole diagnostics fields and tech bias and using a lot of data, mostly on deep tech, but predictive models. However, the real world challenges have been regulatory and FDA clearance and then also reimbursement and payers, just in terms of, even though the information is valuable, being able to iterate, get algorithms that are updated. What have you seen? You know, one, do you agree? But two, what's the path forward?

Vishal Gulati  26:04  
So let's start with the regulatory thank you for the question. So just start with the the regulatory landscape. So the regulatory landscape for imaging and deep learning type algorithms, is, is is very well trotted. There are, there is there, is there. You know, FDA has been very, very upfront about it. They've been on the front foot to get this done, and they have done a fantastic job. In most cases, if you do it correctly, you are able to modify your algorithm without having to get a new approval. You just do a letter to file, and most of our companies do that regularly on reimbursement. I think that there has been there. I think it's been been pretty good. If you look at retinal imaging companies that are diagnosing diabetic retinopathy, for example, all have, most of them have, have reimbursement. One of our portfolio company, caption, health, which is in the space of taking ultrasound images of the heart, they were able to get get reimbursement. So there is, there is a path for for those things. But the challenge has been that a lot of these have been point solutions, and some of them do not solve a big enough problem for there to be a full reimbursement pathway for them. And I think that that limits what places where we can go. So you will so any system, any process which is a long process with where you are trying to build a product which is as tight, takes a tiny piece of that, those are always hard, which is why you find that a lot of the imaging products have been in cancer screening, because in cancer screening Imaging, imaging is the 90% of the pathway, because we look at the image, if you're doing imaging for a pre operated process, that's a very small part of the whole pathway. So we I look for companies where they are capturing the large part of your pathway, and that's how we think we can get

Audience Question 2  28:02  
I was thinking more proteomics and genomics. I agree, right. Do you want imaging? And those have been the one, 

Vishal Gulati  28:06  
Yes, that's right. Yeah, right. 

Audience Question 3  28:12  
Great panel, and I'm glad to see the progress happening. Sumuplement, I'm a cardiologist from Canada, and we've been deploying our own models in AI and demanding. On the panomic side, we are the only company globally to pick up a whole genome sequencing analysis and clinical report from vcfr in five minutes. 

Vishal Gulati  28:31  
Yes...

Audience Question 3  28:31  
You can have 1000 sample. You can have a million sample. 

Vishal Gulati  28:33  
Yeah... 

Audience Question 3  28:33  
The time is a constant. How you update the models, how to make the changes is very important. What matters the data getting in the real time to people hand me as a cardiologist, report in front of me is more important, from a genomic party, when is front of me? 

Vishal Gulati  28:48  
Yeah... 

Audience Question 3  28:48  
Advice after coming back after one month is a useless Yes. And we are deploying AI agents. Also developing, we are developing a longevity based virtual hospital. We have a clinic in Dubai, which just launched recently, and just planning to launch this virtual clinic hospital soon, and AI agents where we are looking to develop and deploy it, which the trainings are going on in the front end side, 

Vishal Gulati  29:11  
Yeah... 

Audience Question 3  29:11  
the receptionist work the day to day, interaction with the patients, low risk environment, 

Vishal Gulati  29:16  
Yes... 

Audience Question 3  29:17  
and I do see a growth happening in the Middle Eastern region. The demand is there. So do you see the similar way that demand is going to be happening here? 

Vishal Gulati  29:25  
So I think that there is a lot of processes within hospitals, within healthcare systems, that are unregulated, processes currently done by humans. And lot of these will get automated the one area. And those will not find great, sexy headlines, and many of those companies will not give venture type returns, but they will have a huge impact on day to day lives of physicians and patients. So one example, which is currently sort of in the gray zone of regulation, which is digital scribes. You know these note taking. Applications which are taking notes of consultations between physicians and patients are completely unregulated and and so and there are challenges with that, because there is, there is a lot of evidence which shows that the the accuracy levels are not very high. 

Audience Question 3  30:17  
Well, we have implemented that, 

Vishal Gulati  30:17  
Yes... 

Audience Question 3  30:17  
we use something called bioelr. We end up in the winning circle in him a couple of weeks ago in Las Vegas. So now Catherine permanente wants to be a pilot with the technology. 

Vishal Gulati  30:21  
Yes...

Audience Question 3  30:21  
So essentially, it's a next generation Health Record system. We call it ELR, electronic longitudinal record system with available data, environment data, we take in ingest and then AI monitors every real time, so at the same time. But the biggest challenge I faced my patient was that I'm a try become a transcriber, note writing, dictation writing. So what I did the modeling was, when I talk to my patients, it listens, then it writes a notes on my behalf, 

Vishal Gulati  30:52  
Yeah...

Audience Question 3  30:52  
and then I reviewed in the end, and they were command, okay, send it to fax to this doctor, to that patient, this and that. See facts, these are using them cutting edge AI, 

Vishal Gulati  31:02  
Yes...

Audience Question 3  31:02  
but then it has to go via fax, exactly. So that's only system. That's the only drawback. The system is you see if you have an AI, but the problem is to the facts. 

Vishal Gulati  31:12  
Yeah, that's that's a valid point. I think that what's really interesting in that sort of thing is that, because these models have become so easily available, and this is a we're living through the golden age of models. These models cost billions of dollars to make, and you can rent them for virtually nothing. So I think that model is not the product. So transcription is not the product. The product is someone verifying that that transcription to say it is accurate, that is the product. So I I'm not going to back a company that can trans you, that can transcribe, because that's a race to the bottom, but I want to back a company that is going to guarantee the accuracy of this and have it regulated. So that's so the where the product is has also shifted with the new models. 

Audience Question 3  31:57  
It is true and from genomic side, or, I guess just last comment, what we felt was we have built the fastest genomic platform. It's like the fastest jet we have built, but the airports are not there, right? 

Vishal Gulati  32:09  
Yeah... 

Audience Question 3  32:09  
see, the systems are not there to use the technology we have built, as we're finding globally, that is true. So I'm looking forward to, yeah, see seeing the work your companies are doing. 

Vishal Gulati  32:18  
Thank you. 

Audience Question 3  32:19  
Thank you.

Audience Question 4  32:22  
Hi. Thank you for the panel today. So at the beginning of your talk, when you answered the first question, you talked about how medical device is a bit of a laggard in this space. 

Vishal Gulati  32:34  
It's a very high risk for me to say it at this point, you do know that, right?

Audience Question 4  32:38  
she agree. I'm coming forward to agree with you, because you're in complete alignment, and from my experience, and I just would love your insight on this, because I think, you know, I'm from a very large medical device organization, and I can tell you, we are laggard in space, because the entire organization is not at the competency level that we need to be at to be able to converse in the language of AI. And if I think about your intro and how you even set the stage right on how did we even get here, I haven't heard anyone explain it in such a kind of methodical way. And so I think for us to not be laggards, we have a really big task in front of us, which is to lift some of the larger organizations, not from an R and D side. Let's put the that group aside, but it really is from the executive side, the commercial side, the sales side, the people who have to carry the message of the product. And so any recommendations or your insights on, how do we start to lift the med tech organizations from a competency standpoint in AI, and any resources that you're like this is, these are golden resources?

Vishal Gulati  33:59  
Well, I'm slightly conflicted in answering your question. In fact, it works very well for my business that med device industries are not so I think that so this is actually in my interest, that that is but I will still answer your question. So I think that, in general, what I would recommend is to, and this is something which, when software innovation was also starting to disrupt healthcare, I used to say that is that people should experience software and try and understand, be able to imagine, I don't expect everyone within a medical device company to be able to use models and build these, these applications on their own, although it's there, it's not that hard. But I do want them to be aware to the level that they can imagine a future with it. And I think that that level of competence is absolutely essential within at least a part of your. Organization where they can try and imagine what what is possible. And I think the the way VCs do that is that we don't always just wait for founders to come and say, this is a great idea. Oftentimes we are expecting them to come with an idea. So you know, you have a list of things that you know will happen, and now you're waiting for the right person to do do those things. So I think that that that kind of thinking is absolutely essential, and I think that every organization should encourage people to do that.

Audience Question 4  35:30  
Well, I don't think you have to worry. Even if we encouraged everyone to do it, we are not going to be disrupting anytime soon. You're very safe

Amrit Panjabi  35:40  
to add to it to Michelle's comment just briefly, we evaluate AI readiness within companies. So we do study how some of these big companies are positioned for an AI ready future. And pharma, the big hire has been chief AI officer and really buffing out their their AI teams within those pharma companies. So we track the partnerships as well. So they have been partnering with startups from small to too big, and with Nvidia as well, laying the groundwork for maybe future applications of AI. I think speaking to Michelle's point, just having people who are conversed in the or well versed in the AI knowledge is just as important to laying the groundwork for the future.

Audience Question 5  36:21  
Hey, thank you for the talk. I did have a question around the burden of proof in healthcare. We know that you know AI, skepticism may be high, especially among clinicians. So my question for you is, what are some of the emerging best practice ways of building AI that's verifiable and explainable? Last year, it seemed like rag was really popular this year, maybe not so much. So I'm just kind of curious your perspective, if there are sort of emerging areas around that?

Vishal Gulati  36:48  
Thank you. I think that there is, there has been. So the two areas where there has been a lot of lot of work done in the last decade, in AI and especially in healthcare, is generalizability and explainability, because without those two things, it's impossible to have a medical product. And so I think that it is so many of our companies spent a lot of resource to build more generalizable AI so that they were able to so one of the ways you do that is you essentially have a wider range of data that you, that you expose your your training to, and so on. And explainability is is another thing, but in it has not been a huge issue, because if the model is very highly predictable, then explainability becomes less important. If it is less predictable. And that's what's going to happen when we come to multimodal models, and we are giving it all type different types of data, and then you can expect a very wide range of answers, and in that situation, explainability will become more important. So I think, again, as I said, the product is not the AI model. The product is all of these things, predictability, explainability, that's the that is the that is the product. 

Audience Question 5  38:10  
Thank you.

Audience Question 6  38:14  
Good morning, and thank you, gentlemen. My name is Ali, and the CEO of cloud caf, we started off with building a medical device that sits in home dialysis patients, in the in the home of peritoneal dialysis patients, and we pick up on infection origination three days earlier than the standard of care. That's fortunately been going great, and I go to market closing the year at about 5 million. Arr, now we're going upstream. We're targeting the chronic kidney disease population with a patient management platform to incentivize more of them to start at the home. And this is basically a six month period through which we need to increase their activation, provide specific education, and it's a very tailored experience per patient, depending on where they're starting. We need to sort of meet them there. We're in the very early stages. Do you have any examples of either companies or projects that have successfully delivered behavioral change at scale across a wide population, like this one? 

Vishal Gulati  39:11  
I think behavioral change is hard. That's the first thing you want to start with. So there have been, there was a lot of excitement around behavioral change in in obesity, there was a lot of excitement behavioral change in mental health, which, through CBT and other types of platforms, some of them have scaled others haven't. I think in the in, in case of obesity, I think GLP ones disrupted most of most of those. So that's that's not happening. And I think in mental health, some of the CBD platforms have been, have been very good at that. So that's all I can I can tell you

Audience Question 6  39:49  
Awesome. Well, in our case, it's a four month period that we're trying to reach a single binary output of, yeah, do it at the home versus crash start in center. So if you get any call. Please let me know, but thank you, you're welcome. 

Vishal Gulati  40:02  
You're welcome.

Amrit Panjabi  40:03  
I think that's that's time for us. So thank you so much for an engaging discussion. And yeah, thank you. If you have any questions, we'll be here. 

Vishal Gulati  40:09  
Thank you.