pean_1
===

[00:00:00] Welcome back to Behind The Knife.

I'm Amin a PGY four and Behind the Knife Fellow here at Duke. Last month, Stanford's Institute for Economic Policy Research released a working paper that showed approximately 76 to 176% efficiency gains in productivity for home tasks by adopters of large language models. There's a lot to say about this, but for busy clinicians, every second that we can get back in our personal lives is cherished.

So how can we start? Well, today, Dr. Patrick Geoff, co-director of Behind the Knife and Trauma Medical Director at Duke and myself are excited to talk to an early adopter and truly an AI power user, Dr. Christian Peon. Dr. Peon is an orthopedic trauma surgeon at Duke, faculty member at the Duke Margolis Institute for Health Policy and founder of Revel AI Health, a company that targets many of the tasks that directly lead to physician burnout.

And he's also the author of Sub of the Substack. The Techie Surgeon. Yes. I'm excited about this one. So the pace of AI development [00:01:00] is dizzy, right? Every single week there's a breathless new proclamation about what AI can do, and it's really exciting, but it's also extraordinarily overwhelming, especially if you happen to be a busy surgeon, you feel like you're getting left behind.

You can't keep up. And in some ways it's at odds with what we do as surgeons. We strive to be very thoughtful, very meticulous, very safe. We wanna do the same thing over and over again 'cause we know it's gonna work for our patients. Meanwhile, AI is kicking down the door and daring you to go faster and faster and do more.

I saw a the other day of the, , back in the day, Christian might appreciate this, the Kool-Aid guy where Oh yeah. Kicks down the door and that was AI kicking down the door. You can't hardly keep up. , And so we thought it'd be a good idea. To do a, a very practical episode in which we focus on AI tools that you can and really should be using today.

So we've got Christian, as Iman mentioned, an AI super user. He really is a super user to enlighten us. So Christian, we want to pick your brain [00:02:00] and talk about some of the basics and how listeners can get into the world of ai. I've heard you talk about the mental model surgeons use to approach ai. And you encourage them to blow it up by thinking about using AI as an operating system for your information life.

So what do you mean by that? First off, , super excited to be on on this. I mean, I think Behind the Knife is doing amazing things, , and I really appreciate what you all do and love the show. Love all of the continuing education you do. My personal perspective is that not just surgeons, but people in general have thought of.

These large language models and as gen of generative AI as a search engine. I think that that is the wrong way to think about these models, and I know we're going to jump into all of the different use cases. You really need to think of these large language models as almost thought partners and infrastructure, in my opinion, for anything that has to do with knowledge work.

I like to say that as humans we kind of really only do two [00:03:00] things. We, , transfer and exchange information and we manipulate and mold matter and the transfer and exchange of information. These models are so, so good at it. , So that's what I mean when I say we need a different mental model. We really need to, this is not Google on steroids.

This is, , a system that if you implement the right way, can help you augment so much of the knowledge work that you do. Yeah,, I think that , , the search engine analogy is unfortunate, but I think that the reason it gets made is because there's nothing else that has changed the way we interact with the world as much as, , the search engine, right?

Like you think about what has, what has really changed human use of, , information gathering or anything else. And you have the search engine and now you have large language models, but they're fundamentally different concepts, but only comparable in their impact.

Yeah, and I think it's funny, um, you know, you mentioned that as surgeons we tend to be very regimented and think about step A, step B, step C, and yes, large language models Jet, GPT, claw, and they're all [00:04:00] stochastic models. So like you get, you put one input, you're gonna get different outputs. And I think that kind of rubs us the wrong way, but that's, that's also part of the power of these tools.

And I like to say, if you're a clinician. Just, just think about a process that you have to execute over and over again that involves the exchange and transfer of information, whether it's writing notes, , trying to figure out what the proper billing code is for your operative note. , Maybe it's putting together an Excel sheet of information week to week and.

Put a list of those processes down on, on paper if you want, or speak them into a large language model. And I think that you will be just shocked at how good these systems are for those repeatable processes that are related to information transfer and exchange. Yeah, it's all about kind of getting over that hump, right?

It's, you have a few good use cases and you're like, oh good lord, this is amazing, and you're off to the races. , We were just talking before we hit record about the fact that we [00:05:00] see shockingly few staff for one, but also current residents. So kind of in that this age group we're talking about using LLMs specifically, or AI in general on a regular basis.

, Shockingly few. , Once you kind of get down to the ranks of the current medical students and undergrads, there seems to be something about that generation where they're picking up on using these tools, , more readily. And so that's especially why, , I think it's important to, for this episode to happen today and, and for folks to get involved because, , the power of these tools are, are, are just amazing.

And so upfront, are these frontier LLMs, right? The original. Groundbreaking large language models. The three main ones that we all talk about are Claude Gemini and Chat, bt GPT. These are the workhorses of the AI world, and as, as recently as one to two years ago, these were just more of a large language model.

More along this line of, , uh, a more analogous to this Google search. And they do so much more now. So Christian, how [00:06:00] should folks decide which large language model or which ones to use if you can use multiple. And are there any specific hacks for these models that you'd like to share? Yeah, these models, they differ in what they're best for.

I think that the same way that you use a Bovie cautery and it's one case, , maybe me a, a mallet in another and then a 10 blade, , in for a another part of a case. I think that it's important to try to use these models for what they're best at. I have my own stack and I rely heavily on Claude Cowork personally, and I would encourage the users to.

In my opinion, use, use that as your entry, if you will in your initial kind of large language model, thought partner. It's very good at writing, it's very good at a variety of different outputs, whether it's an HTML. So, and when I say HTML, it's like a little website, , that creates visuals for you. It's excellent at that.

It's very good at repeatable structure through what's called skills, and it ends up being the workhorse [00:07:00] of my personal AI stack. So. Let's say I want every email that comes in from my research student to get, , scanned and then to extract the, , the word document that was attached and then have it waiting for me in a readable document.

Uh, highlighting all the, the prime points and taking the perspective of a peer reviewer. I use Quad for that. I think that there are other tools that are better at visuals. So if you go on my newsletter, I have this thing, I call them visual prompts. I make prompts in Claude in order to express myself visually through a slide deck or maybe an image, , a data visualization.

And then I actually will take the prompts. That I've made in Claude and usually go into Gemini or their, , image creation tool called VO three. And that's a fantastic tool to make really compelling images. They can actually come pretty close in some cases to drawing some of the anatomy, if you will, on an image.

, And it, it's fantastic for that. . I've stopped using chat GBT [00:08:00] personally and look, I was a mega user. I, , received a notification from them that I was in the top, I think 0.75% of users impressive. No, nothing wrong with it, but I just ended up transitioning over personally, , to Claude. So that's those, that's my stack.

, I will also say I use open evidence, , for clinical use cases around medical answering as well as Doximity GPT, but open evidence tends to be, my go-to. Um, and then, and then the other tool that I've, I've used is just the ones that are embedded within my, in my electronic health record system.

And that's an ambient scribe, but in this case it's a bridge. But we've other sites are using rev, AI's scribe. You can use whatever one is appropriate for your practice. I think. Right. And so, , what about when folks say, well, I, you know, I, I only have so much time. I'm going to jump into a Gemini, right?

Yeah. And I'm gonna learn the interface. I learn what kind of tools it offers. And then two weeks later, Chad, GPT drops, the next biggest thing. And this is a horse race, right? Each, yeah. These, these three, these three companies [00:09:00] are massive. They've done incredible. And every week and every month it's a little bit different.

Do you have any recommendations for jumping around between models? Again, especially when something, right now, this is April, 2026. Claude Claude in general has, has a, a lead for some good reasons we'll talk about. But any recommendations for that? Because it can again, be overwhelming. It could be overwhelming now, now I'm using the worst, you know, I need to be using something different.

It can be so overwhelming. And what I would say is you can, it's very easy with these tools to, to just get analysis paralysis. And I've seen actually from some of my, , Substack readers, I see them following along, following along, trying. They're downloading all these skills files, right? These workflows, you just have to use.

And it's okay if it's chat GPT by the way. You just have to use one of these tools and integrate it into your workflow. And I think that that will go a long way. I personally would start with Cloud cowork because it is very, is a very friendly user interface that can manipulate files on your desktop.

And for [00:10:00] repeatable knowledge work that we're doing as clinicians, I think it's going to be the most powerful, , platform to just introduce yourself to the world of large language models. I will say though, I personally, I jump all the time between Claude and g Gemini are VO three because I'm going from written input to visual, visual output.

And, , there's nothing wrong with that. You don't want 10 tools, but two to three or four is, is really powerful. And, um, just find your workflow for those repeatable tasks. Stick to it and just keep iterating. Like, just think about the use case. I, I want to put together a morbidity mortality conference.

De-identified and I need to put this case study together. You're gonna, I'm gonna take the evidence from open evidence on that topic. I'm gonna package it and copy, paste it into Claude Cowork and have a file ready for me that's got my Duke branding. You know, like, just think of the, the end product, the inputs, and then try to let these large language models arrange things, , for you.

And then I say, , you know, never trust, always [00:11:00] verify. Um mm-hmm. Your domain expertise is actually really powerful here, and you know the right question to ask. You still need to be willing, , to spend the time to come up with the workflow and check the final product. But before we go any further, I mean this, this is all great.

I think that people would really like to hear some specifics on how you use. These tools practically. So you mentioned, for example, building a case for Eminem, or you mentioned reading through one of your research students', , emails. But, and you know, this is a question for both you and, and Dr. Geoff, but what, what are some good other examples that people can start with maybe today or tomorrow?

Um, just to, just to dip their toes in the water. Patrick, you got something? I got, you know, I got, yeah, well because I got something, 'cause I was literally just working on it right before we, we came on, , for the show today. So as trauma medical director at Duke, we're interested in looking at our trauma triage guidelines.

And so we have three years worth of data, over 12,000 [00:12:00] patients. , They come from a trauma registry and that's a painstakingly. Data, , this is a manual process. And before, if I wanted to thoroughly analyze that using different, , triage guideline, , different triage guidelines or resource based tools that show you whether you're over activating, under activating, somewhere in between, that would take.

Literally an expert in, in statistics and likely weeks of of work. And then when you want to iterate on it a little farther, you go back to the drawing board, wait a couple more weeks and back forth. It's extremely labor intensive. Instead, , I'm using actually a combination of Claude and Gemini and I've got this data set and.

Constantly iterating from it, saying, okay, tell me last year's, , a breakdown, my level one twos and threes, how am I over under triage according to this criteria? How about this criteria? Okay, where am I falling out? How are we under triaging here? Which mechanism of injury is it? And it is remarkable 'cause all I'm doing is typing or even now speaking [00:13:00] questions to the computer.

He goes and does his thing. And it's, it's kind of gotten ridiculous to the point now where I'll have Claude doing some heavy duty work. And in between then I'll be on the fast version of Gemini Cross-referencing some topics to, as I wait to kind of fill myself in on that next paper, the next thing that I need to, , that I'm about to ask.

And so that, that's an example of workflow that, I mean, even six months ago or a year ago, I should say, couldn't have done it. You couldn't, you could not do it. And the beautiful part about it is that too, is it spits out these PowerPoints. , Slides and say, Hey, make me a slide. Write me a brief executive summary.

So when I go back to my colleagues, , in this very multi-institution, or multi, uh, I should say, specialty practice of trauma surgery. That we have some really great deliverables to share and say, Hey, look at all this work, and this is me some, some data Luddite working on this stuff. I think it's, it, it, it still actually blows my mind that this is possible and, and so powerful.

Christian, I know you have, you have endless, , you've been plenty. But I will, I [00:14:00] will try to be like basic, if not basic. I'll try to be practical, right? , A lot of us are trying to keep up with the literature. One really cool thing about cloud cohort work is this thing called scheduled tasks. And it does have what's called again, and don't worry about some of this terminology, but it has an MCP, , basically a connection.

To PubMed. So one thing that I've done is, hey, I need to stay up to date on orthopedic trauma literature. , And actually I'll tell you like we could even create this now live. I use, um, a dictation tool, whisper flow, , and I speak instead of typing where I can, I was like. Every week I want to be updated on the Journal of Orthopedic Trauma's, latest and most relevant topics for an orthopedic trauma surgeon.

And I wanna receive an email with links to those, , PubMed links for that journal. I'm an orthopedic trauma surgeon about five years into practice, so. Try to format the email to me in a format that's going. I'm gonna be able to scan really quickly, and that is relevant to my practice, and also make for me an HTML that I can [00:15:00] interact with and has some of those key points from those articles.

You hear the little bleep, that's the dictation. I don't type because it lets me iterate faster. And actually Claude Cowork, I'm in it now. I'm gonna send that and it's gonna make a scheduled task for me. So it's thinking, it's putting it together and it's asking me, okay, what day of the week do you want this to happen?

I'm gonna say every Wednesday at 7:00 AM Now every Wednesday, 7:00 AM I'm gonna receive in my email inbox, uh, and in my Claude Cowork instance. The latest and greatest in orthopedic trauma, and I can augment my knowledge that way. I'm not doing it manually. So that's one use case. I think another great use case to think about, you know, if we're really thinking about clinicians and residents, I use it as a coaching tool.

So, , you talk about like stringing together things. I, I use granola, that's. That's a meetings application. And again, I use it, , because it's very simple, but it also pulls information, , into Claude through, , MCP. , I might say, Hey, you [00:16:00] know, junior resident who just came outta my service. Is it okay if I record our conversation?

I just wanna do some, some targeted feedback, , together. Let's, let's talk about the cases. We've gone over this, this week, and I'll turn my granola on. I'll have the recording of the transcript. I'll give my feedback. I'll say, great. Hey, like. I'm gonna get back to you with some of the feedback here and, and we'll do this again in four weeks and, and we'll build on this existing conversation we had.

And rather than going and typing, 'cause I'm a busy surgeon, I'm gonna forget, right? I'll take that transcript and I have a skill in my claw, but it doesn't matter. You don't need a skill. You just say, I just had a meeting with my junior resident. I got permission from them. Please outline some of the strengths and weaknesses that we discussed, uh, in a way that I can send to them in an email that I can build on the next time that we meet.

And by the way, go through the transcript and give me some feedback on whether I gave good feedback as a mentor. , Did I give enough insight and advice to this student that's going to be relevant for them, and how can I improve on it next time? So that's like another use case and [00:17:00] example of how I'm using these tools to augment my daily interactions.

Right. And that feels maybe a little out of the box. Again, it's the transfer and exchange of information. It's a task that I'm doing over and over again, and it's one that I'm using a large language model as a thought partner so that I can get better over time. So those are a couple. There's so many more man, there's so many more patient education.

I think we're gonna, let's, let's go around the horn again. Imon. I want to ask you about one or two. Yeah, I'm gonna share another one or two. Chris, you get to share another one or two because I do think this is, this is. , Important to hear how people use these tools really opens your mind up. Maybe you might want to use it in the exact same way that, you know, Kristen, you just mentioned, or maybe there's something that similar that you go, wait a minute, I, I, I see the, the connection here.

Let me try this new process. Yeah. You, I, I think something Dr. Pan said earlier, he said, you know, he, he, he says the word repeatable a lot, right? You, you keep saying anything repeatable and that's, that's a concept in computer science that goes back decades. It's, you know, the, something I was taught very early on is if you find yourself writing the same thing twice.

Can you just write it once? Right. Is that [00:18:00] something that you can turn into a function, , is what we would say. And so one practical example that I did ju just a year ago is I think that these tools are fantastic for administrative tasks. Mm-hmm. So one, one administrative tasks that we have to do as residents is you need to schedule medical students to different rotations.

Now, historically, this was a massive. , Deal in our program, you know, our residents would be late occasionally in, in slotting out. These medical students we're all busy. , That would delay the, , the program coordinator that would delay the anesthesia department. That, I mean, it, it was just this huge consequence.

And I looked at the problem and I thought, this is such a simple thing to just, , to just, just write as a function. So it only took me, you know, a, a, a very small amount of time, probably less time than just actually scheduling people myself. And I built a website. And, , and I, and I just gave it to our program coordinator, and now instead of residents doing it, he does it and, , it takes him one second of just plugging in an Excel sheet and he's done.

So, you know, it it, this was an, an administrative task that took [00:19:00] residents, you know, 10, 15 hours each, every single month. And now it's completely gone and it takes about 15 seconds a month. So I think anytime that you have something administrative. And you find yourself doing something like copy pasting or anything like that.

These are things that LLMs ex excel at. , And they, and they excel at transferring knowledge as well. So you know, if you have a task that's very repeatable, but for some reason someone else just can't do it. And again, scheduling kind of comes down to that. Well, I mean, most likely an LLM can do it really, really well.

Right? Those nuances that you think that only you have, it's probably something you can transfer. , So that, that's something I use it for a lot. Of course, I, being in computer science, I use it a lot for research, but really it has, it has reduced my administrative burden, , dramatically. And that paper I referenced earlier with 200%.

Efficiency gains and productivity for home tasks. For me, it's gotta be like 500% things that took me six hours. Now take me one, I mean even something like editing this podcast, , in terms of just totally just, [00:20:00] just making little corrections. You know, maybe we had a blooper or someone's, , someone's kid came running in the background Yeah.

Historically that, that would take, you know, 15, 20 minutes to edit and now it's just a two second thing with an AI tool. So, those are some of the ways I use it, but again, infinite examples. And we at, you know, behind the knife, , making content. We have a, a growing tech stack, tech stack being just the tools that we use to create and it's multilayered.

And I think Dr. Pan, you'll give us a, an example of a, of a fairly simple three step process for a AI Grand Rounds presentation, which I think is just a really nice example of how you can kind of daisy chain these things together and. , One of the biggest projects we're working on, , Iman and Dr.

Swensen at Behind the Knife is the AI board simulator, and that's a, a an LLM based, , uh, project. Uh, that I think I've learned a ton from Matt Iman about how these LLMs function, because we're trying to, you know. Tweak these models to match what, [00:21:00] uh, the nuance and the detail and the richness of an interaction with a board examiner.

Mm-hmm. And that is certainly much easier a said than done, but, you know, that's would be an, an an advanced level thing i'm gonna go backwards a little bit farther before it's more personal things. I just moved into a new home. And again, congratulations. Talk different models. You know, for this, I use Gemini.

Gemini, you choose fast thinking, , et cetera. So I use the fast function of Gemini on my phone and I've had to fix a whole bunch of different things, do a bunch of different gardening, and. It had some tree that was dying to take a picture of it. Well, why is this dying? Next thing you know, I got the stuff ordered through Amazon and I'm treating it with the, the right insect repellent thing.

And, and lo and behold, it's a few weeks ago Now, the tree looks beautiful. , We, had to fix the, some covers and had some very specific hinge on it, which I've never seen before. Take a picture of it. Instantaneous. This is how you fix it. Step one, step two, step three. So, so it doesn't have to be. , The most advanced, you know, the most advanced [00:22:00] usage, but this is part of the door that opens once you start using it, you can also personalize things, , by creating your own chat, , functions within chat, GPT Gemini, they call 'em gems, et cetera.

And certainly within Claw there's an extensive variety, which we're gonna touch on. When it comes to creating things like artifacts and projects and skills. , But, recently I got a pool, which is exciting. I have no clue how to take care of a pool. But now I have within Gemini a gem that says I'm, it knows all my pool equipment.

It's, I've taken pictures of it, it knows how many gallons there are. So when I test the chlorine, I just say, Hey, my chlorine's this, and tells you exactly how much to put in it. And I actually compared it to, 'cause I, I got this pool and I was like, oh my gosh, I'm overwhelmed, downloaded all these different apps.

And those apps are, are garbage compared to the highly personalized, , information given through, , what took me literally, I think three or four minutes to set up. But so I love that man. I love, I love that actually. 'cause of the last thing you said. So this idea of a personalized software that [00:23:00] is serving your needs uniquely, I think that is where a lot of your listeners can leverage these tools to come up with things that.

We just would not have ever been able to have come up with on our own. And that honestly, no developer's going to either 'cause they don't have your lived experience, , or your lens of your unique problems. , And so, you know, I I think that, , my advice would honestly be just chat with these tools with some of your common problems, , and see what comes out of them.

Be, , be willing to scrutinize, but then, then. Explore and see if you can get to that repeatable process that's information laden, , in a way that is more efficient, , or creative than you would've been able to on your own. . I think these tools are great for Excel sheets, you know, and scheduling, like you mentioned.

Right. , And, , there's so much administrative work that can be automated. , Just another thing, like if you do have something you wanna explain to a patient, right? And maybe you're struggling for the words, , to explain to them, I mean, these tools are great for that. [00:24:00] I'll do that in open evidence. I'll go in and say, Hey, I've got, you know, this.

Issue or complication for a patient, what's a good way to explain it to them? And sometimes I'll take, go a step further and take that explanation, put it into Gemini, ask for an image that might help conceptualize it for the patient. , And I've used that before and actually it's been, it's been pretty great.

So just new ways to communicate and conceptualize information. Yeah. None of us are talking about clinical use cases really. Right. And this, right. Well, we have plenty to, plenty to talk about before we get into your AI ground, , ground grounds presentation. But when it comes to the clinical use cases, a few months ago we had an episode on, , , the ambient scribes, right?

, And I don't know if there is a more. Direct way for physicians who spend any time in clinic, but also, it's, it's actually quite useful for inpatient encounters too. To save more time, if you wanna, we, we, we have an audience of clinicians, right? Mostly. And if you want to save time, take a half day clinic.

I would say if I do a half day [00:25:00] clinic, I'm saving about an hour and a half to two hours. Of time, I'm kind of verbose when I dictate and I dictate my notes and I like to talk a lot and kind of fill it in. And I think as part of this, , Christian, I got to thinking about this when you mentioned how you communicate with patients.

It's also remarkable. And as you're talking, like to your phone, right, to get your, your scribe to create a note that you want by invert, by in, by doing so, you're also talking to the patient in a way, , that's much more clear and you're, you're sharing a lot of information that maybe you would not have shared before.

And I found that those interactions to be better, I can spend more time with patients and then the notes are better. It says exactly what I told them. Right. It, it says exactly what they told me. There's no, , there's no, , making things up in between. Right. , I think it's an incredible tool. You know, I, the, the ambiance grab one is fascinating because I read Dr.

Pan's article from January where it's, it's, it's actually, it's, and trials have shown this. It's, it's actually not even the [00:26:00] time that's, that you saved. It's, it's the, and you actually, you should probably explain this since you wrote the article. No, no, no. That I read. But, you know, it's, it, it, it, it's not even the times, it's the burnout, right?

It's like, it's like what is really. What is really the thing that burns you out? It's not necessarily looking at the note, it's the process of sitting down, typing out the entire thing. And even if you're not saving time, you're saving your mental load. And I, that's just one of the most powerful things you can have.

Totally. Yeah, I was gonna say, it's interesting, it's. I feel I would inherently say that the ambient scribe saves me hours. Most of the studies are suggesting maybe 20 to 30 minutes a, a marginal increase. Oh, I feel like it's an hour. No, me too. Well, that's what I was just, I, I feel like that I really do, but, and maybe it's not, maybe it doesn't matter that it's.

20 minutes in aggregate in these studies instead of an hour and a half. But it be, it, it has completely changed definitely the way that I, , you know, all my notes are done on time and they, they didn't always used to be. , And I agree with you. I think the other thing is, you know, you're, you're mentioning [00:27:00] something, , that maybe a super user is more likely to do.

I've heard people complain about the ambient scribe, and I, I always tell them like. Did you say what you wanted it to do? Did you, before you went into the room, give a little bit of context saying, I'm gonna go in to see a 68-year-old female who had a hip fracture two weeks ago, as you got these comorbid conditions.

Then go in and then give a little more context on the way out. I personally read the radiographs. I need the physical exam to specifically emphasize X, Y, Z, because if you don't. Then yeah, you know, garbage in, garbage out to some extent. And if you're in the, in the room with the patient and you, , you know, give a very cursory, have a very cursory conversation, that's all the LLM has to go off of.

But it sounds like you, Patrick, have honed in on it and maybe that's why it is saving you and me hours and others, you know, 20 or 30 minutes. . The last thing I'll say is like, so that's a skill to learn when you're using these ambient scribe tools is you have the ability to basically talk to the scribe to give it instructions around the note that you're trying to create.

The other thing I'll say is I do think that the technology, , [00:28:00] itself of ambient scribing is getting commoditized, but it is the launchpad, in my opinion, for deeper workflows like that's coming, right? Like whether it's one of the big systems of record like Epic or, , some of us that are working in these, in, in integrated verticals of care management with ai, it's gonna be like scribe informed a chart review.

That was done by an agent that ended up suggesting an order that did an asynchronous outreach to close the care gap. I mean, like that is where the, that's gonna be the launchpad I think for, for clinical intelligence. , My impression is that I think, as you said, the reason you two are really good at using this tool is because you've used other LLMs.

And the skills you learn when you direct an LLM are very horizontal. They kind of apply to any tool that's built on top of an LLM. , So. It, it the, the reason that we, you know, I think that we, we are really trying to encourage people to use these, , to any degree is because as they get deployed, the way to get the most out of them is just to get a feeling for how they work.

And you can [00:29:00] only do that by interacting with it. And, , that goes back to exactly what you said. The more context you give these things, the better they are. The less context you give, the more variable the result and the less reliable the outcome. So, , I think that's just one big, , reason to start.

Tomorrow using these tools You do today? Yeah, yesterday. Yesterday. Right. But, but, and we said the top that, , Christian has a, a Substack called Techie Surgeon.

It's a fantastic resource to learn how to use AI tools like we're talking about today, but it's also a really fantastic, , look at the interface between clinicians, ai, and the types of, , tools, , that we're reviewing, like LLMs, , and, , healthcare policy and , the financial aspects of a hospital.

And so if you have any interest in that, I, I highly recommend you check that out. I appreciate that. Yeah, and, and that's the thing about I think technology is that it's intersecting with our patient care delivery and our patient care delivery is, I. Extremely impacted by policy. And so [00:30:00] I've tried to write about the intersection between the three kind of clinical ai, , care delivery and health policy because they all shape one another.

, So yeah, please. , And by the way, if people do subscribe or check out some of the live sessions, I would love more feedback in insight from people like. Your, your listeners at Behind the Knife? , Because I wanna empower people to use this technology to become more knowledgeable about health policy and shape the field, their own lives and the lives of their patients.

And I, I think this is just a great tool to do that. , It's not an end all be all. , You know, you, you, , need to scrutinize the output from these tools, but I think that we need more creative clinicians. Using these tools, figuring out the next big use case. It's not gonna be a a techie person. I think it's gonna be someone with really unique domain expertise and insight to unique problems.

, Those are gonna be the real, , I think transformative, , applications of this technology. Right. And, and we all have domain expertise. Right? Yeah. Whether you're an intern, a medical [00:31:00] student, , yeah. You all have your own lived experience. You know, you're a, a senior surgeon, uh, everyone has true domain expertise, and as surgeons, it's pretty awesome, right?

Because there there's not that many people, , who are surgeons, number one. And then if you learn how to interface with this technology, how that becomes such a powerful tool to, to leverage that, that insider information. That, , you know, the next tech startup founder doesn't have, 'cause they didn't do a surgery residency, they didn't, they haven't spent years taking care of patients, you know, in the middle of the night, you know, fixing femurs on, on Saturday afternoon, et cetera.

So, . Christian, let's, let's focus on Claude Cowork for a minute, because that is all the rage right now is something that you like to use and talk about, and I'd like you to explain to us a few of the features. So I think most folks know you, you fire up an LLM, you can type it in, you can talk to it, and you can have this ongoing conversation.

You can do some really neat things for you. But within Claude Cowork, you hear people talking about projects for one. Yeah. , That you hear them talking about artifacts. And skills. Can you explain what each of [00:32:00] those are? Absolutely. In your simplest of terms. Absolutely. Yeah. So, and Cloud cowork I think is really useful for a couple of reasons.

Just quickly. , One, it can actually manipulate your files on your desktop. That's intimidating at first, but it's a mechanism for the large language model to organize information, and that's where projects come in and. Projects are basically folders of information and they allow for you to work in a certain domain or subject and not have to reiterate over and over again to the chat bot what you, what it is you're working on.

And by doing that you can, you know, have a folder for personal reasons where you're, where you're workouts are getting downloaded and maybe the chores task list, the, the pool cleaning instructions for Patrick. And you can work in that domain. You can switch to another project. Let's say that's your academic life and you've got a folder of research.

You've got a folder of your personal ambitions for your career, and you've got your branding guidelines from your academic institution, your CV and your working, and the [00:33:00] LLM begins to know you to some extent, and that's really powerful. The other thing that I think is really powerful about Cloud cowork is the ability to run tasks in parallel.

So a lot of us are used to like, you know, I'm gonna open up a chat session, I'm gonna work on something, and I'm waiting until the chat bot is finished cooking whatever I'm doing. Or I'm trying to open up multiple tabs in my browser and trying to work on things. But oftentimes the, you know, the chat or the task gets stuck and Claude Cowork, you can fluidly.

Start multiple sessions. So maybe one is writing that prior authorization appeal letter. Another session is editing a research paper. The third session is building a, a visualization or artifact, , explaining a new research project for you or students art. So that's really powerful. Artifacts. That's just the output from the model that lives in a sidebar, and that might be an Excel sheet.

It's oftentimes an HTML, which is basically a little web-based, , widget that gets put out [00:34:00] from whatever you have done. I like the HTML output. Because from a workflow perspective, I may say, give me an interactive HTML with prompts embedded in it, and I may take that HTML and then those prompts and do something in another tool like Open Evidence or like Gemini in order to amplify the work that I'm doing.

So there's tons of po, tons of possibilities. The other thing that's nice about Cloud cowork is it has these scheduled tasks. You heard us talk about them earlier. You can have tasks. Running at certain times. Let's say every day you want your emails to be sorted, , and brought to you in an HTML.

Prioritized reminding you which emails you need to get back to and to have drafts of responses to emails you haven't gotten back to, , Claude Cowork schedule tasks can do that. , The last two things that I'll mention is dispatch. Dispatch allows you to use the tool remotely from your cell phone as long as your computer is on.

You can run a complicated task on cloud cowork while you're on the go. , And then skills are repeatable flows. [00:35:00] So, . You mentioned, , earlier, , this scheduling task, you could have a skill, right? Let's say you worked on this for a long time in a session and you're like, I need an Excel sheet that looks like this.

, And I, I, you know, I need to make sure that I've sorted the PGY years differently and the services, and you like the output. You could say, Hey, like, I want that to be a skill from now on. When I say scheduling Excel skill, I want to be able to have a similar output, but it may be for a different task and.

, You can do that and that skill, the next time that you invoke it or mention it will be applied to a completely different task, but with a similar output. You can do that around your branding and much more. So that's, that's hopefully a reasonable overview and I know it's a lot in a short amount of time.

, There's so much good material out there to learn each one of those different little facets of cloud cowork and to leverage them for your own workflows. . That's why I love this tool so much is it's so versatile in its output and its applicability to different tasks. Yeah. Barely scratching the surface.

But if you want to get into it, , watching, I'd probably [00:36:00] say two hours, maybe worth of YouTube videos, , you find the most popular ones or the highest views, you can kind of choose a few, kind of what you might like the style of, I'd say about two hours, get you started, and from what Christian you said.

The, the simplest approach, I would say, if you have a project, especially as thinking about our listeners and surgeons being mm-hmm. Folks who, you know, administrative duties, clinical duties, et cetera. If you have a project you wanna start with, I think you can start simple with a folder where you store PDFs.

Yeah. And PowerPoints and stuff like that. And where you could talk to it about a skillset, Hey, you are the chair of surgery at such and such institution. Uhhuh. This project is about optimizing, or rooms. There's these stakeholders that are involved. Okay, you, you have this. Fixed set of knowledge and tasks go with, you have a certain number of PDFs and whatnot that you wanna work from, and then within that conversation, even alone, you can just start.

Iterating and then you can move on to these next things where you're automating your workflows, where, you know, again, it's kind of a bigger leap to say, well, I'm [00:37:00] okay with, with connecting different tools you mentioned. Mm-hmm. Connecting different agents and tools, like, okay, I'm okay with this folder being shared with Claude.

I'm okay with this Gmail account being shared with. Et cetera. And so, so that's kind of, I think, more advanced and next steps. And then culminating the, in the, the fact that Christian is an animal and he is got like 10 different things happening at once on these scheduled tasks, you know, that that's super, super high level stuff.

But if you back it down to just a single project, for example, I think a good one is that, that trauma optimization thing, that's something that anyone, , I can do. And, and, and yeah, I, you know, you'll be really pleased with that. So one thing I wanted to, , have you talk about Christian is the word age.

Agentic is everywhere. What, what does that, what does that mean to you in the setting of, , LLMs and Claude or kind of the future of how you look at this? Moving from a chat kind of thought to a doing, feature of ai. Yeah, I really think, , of agentic AI as just AI that's moving [00:38:00] from chat and surfacing information to doing tasks.

, You know, I think that this is about a lot of the things we've already been talking about. , This is where you have AI that is actually taking the next few steps. Maybe it is, , that you have the ai like we were talking about earlier. Going into your files, pulling out the information, creating an actual Excel sheet and firing off an email.

Like those three steps to me, make me think of, , agentic ai. And, you know, I try to tell people, don't get too hung up on labels. , Claude Cowork in and of itself inherently is really agentic because it's, it, it's executing multiple steps and tasks and it's going beyond that sort of search engine.

Mentality to operating system. That's really the way that I think about it. , Once you're starting to delegate tasks, , that are multi-step to ai, you've kind of crossed over into the agent framework, , in my opinion. , Now some, some skeptics would say, you know, oh, that's CR [00:39:00] jobs and just individual tasks.

It's not really agentic, it's not autonomous. But that's the simplest way that I think about it. You know, that, uh, I don't know if you actually, maybe I, Amon you may have a different, , different definition. I, I mean, no, that, that was great. I mean, I think if you want to go, the most simple way to think about it is, , you can treat it like a medical student first, then you can treat it as a resident.

, And then, you know, one day you can maybe even treat it as a co-partner. But, , I think of it as, you know, maybe you're an attending and, , you delegate some task. Maybe you're only comfortable with, okay, go take down the dressing for now. , But eventually you'll work your way up to, okay, why don't you go round on the patients and let me know who's sick.

, Yep. So, , that, that's kind of how I think of an agent, the same way you would think of any sort of, , any sort of person working up the medical pathway. , But that's awesome. But, , but maybe to, to wrap up, I , I want to ask you about Revel ai. So, what is it?

, , and why did you start it? What was the inspiration? Yeah, absolutely. Yeah. Rev AI is really an agent care coordination platform. We've [00:40:00] really thought about the different regulatory policies and the common patient facing tasks that we face all the time, and we've wrapped AI on top of those interactions to try to improve care coordination.

Most of our work is in the musculoskeletal space. So we've looked at some of these regulatory workflows, like collecting patient report outcome measures, or coordinating care for patients before and after surgery. And we're really equipping AI to educate patients with content that we create, , con converse with the patients and collect information from them that also maps to some of these policy regulations.

And then automating what. The clinician would be doing. So we've got ambient scribing technology. , We have, , voiceover internet protocols. So you can make phone calls yourself or your staff can, and documentation automatically happens. But we always wrap it around episode of care management. And what inspired me to start the company, , you know, I wasn't honestly out.

To start a company, I was [00:41:00] trying to solve problems. There were these value-based care models that orthopedic surgeons needed to participate in. We all of a sudden were responsible for optimizing patients preoperatively, trying to contain costs in the postoperative period. And I saw a huge gap, especially actually for my underserved patients in the communication that we had with them collecting the information for them.

And I said, you know, why don't we have technology that can do some of this? , And it started off as a project. We got selected for a Techstars Accelerator, which came with some funding. Ended up, ended up going the venture backed route. Raised a a round of 3.1 million this last year, and now we're getting close to being in almost 20 hospitals and care organizations.

We're integrated into Cerner and Athenahealth and we're getting into Epic. , It's been a remarkable experience. And also, by the way. One that I would encourage, , listeners to at least get exposure to. Our platform now is interacted with over 200,000 patients. Our agents have interacted with over 200,000 patients.

[00:42:00] We're automating a lot of compliance and regulatory workflows for clinicians and hospitals, and that scale has been super gratifying. As a clinician, I still operate, do a lot of cases, but, , we need more surgeons and clinicians, , at the innovator's table. I kind of fell into it, but I would do it over and over again if I could go back, even though it's been a rollercoaster.

So that's, that's Revel AI in a nutshell, would love, , on another day to give you some examples of some use cases and how we're using agen AI there. But,, I'm hopeful that it's just one example , of many other clinician founded companies leveraging AI that will spring up in the future. Nice.

Well, it's, it's, , a pleasure having you on Christian and, , and chatting with you about all this. , I hope people got something outta this as we kind of blathered on about different, , use cases. But again, this is something that is extraordinarily important, obviously excites all of us. , Again, check out Christian Substack for more, and until next time, dominate the day.