S1, E38 - Reflections 5: How Specialized Does AI Have to Be to Actually Work?

Summary

---

Practical AI in Healthcare – S1, E38

Reflections 5: How Specialized Does AI Have to Be to Actually Work?
Hosts: Dr. Steven Labkoff and Dr. Leon Rozenblit
Release Date: May 24, 2026

---

Episode Overview

This fifth “Reflections” episode looks back over the podcast’s recent block of interviews, synthesizing hard-earned lessons from five influential guests—including leaders driving AI in pharma, a foundational figure in medical AI, a clinician-patient using AI firsthand, and two practitioners shaping ambient AI for care delivery. The central theme: after decades of hype and infrastructure-building, AI in healthcare shows real-world success only when solutions are deeply specialized—attuned to narrow workflows, expertise levels, and proprietary data. Labkoff and Rozenblit break down five emergent themes and candidly examine what’s working, what’s not, and why the human, organizational, and data factors now matter more than the underlying models.

---

Key Discussion Points & Insights

1. The Limitations of Generalist AI – Specialization is Winning

Timestamps: [04:35]–[07:16]; [20:34]–[22:23]

• Generalist AI and LLMs have rapidly commoditized; anyone can deploy them now.
• Success stories come not from “super agent” solutions but from specialized agents deeply embedded in particular workflows or use cases.
• Matt Trupo (Sanofi): Tried to create “one super agent to rule them all”—it failed:

  “I started out to create like one super agent to rule them all, and it failed miserably. The reason was it really wasn't specific enough in any single area.” ([05:23])

• Pivot is to multiple, highly specialized agents (e.g., for morning prep, governance, email).
• David Hidalgo Gado (Clio Health): Competitors try to generalize—Clio spent 9 months working iteratively with one ER group, “solved many problems for the same people,” not many people for one problem.

  “There are about 120 Ambient AI competitors generalizing across all of primary care. Clio went the other direction ... solved many problems for the same people ... spent nine months with one emergency department group, did 40 to 50 iteration loops before they launched anything publicly.” ([05:57])

• Proprietary, deep data—and closeness to workflow—create the durable advantage.

2. The Old is New Again – Symbolic AI’s Quiet Return

Timestamps: [03:36]–[10:50]

• Ted Shortliffe: The 1970s-80s “expert systems” in medical AI “didn’t fail”—they were just too early; now their methods (like knowledge graphs) are resurging, layered on top of LLMs.
• 2020s breakthroughs are often “1985 with better hardware”—combining symbolic AI and neural nets.
• Leon: “Winning solutions are going to be a mix of symbolic and AI and neural networks working together in various ways.” ([11:52])
• Foundational work deserves more credit; the methodology wasn’t wrong, just underpowered then.

3. AI Literacy & User Variability – The Human Bottleneck

Timestamps: [12:13]–[17:02]

• The gulf in results is often not model-driven, but user-driven: expertise, literacy, and willingness to experiment drastically impact accuracy.
• Barry Chaikin (Physician/Patient):
  Cites stark data—same AI tool, but physician-users get nearly double the diagnostic accuracy of patient-users.

  “The model was identical, the interface was identical. The variable was prompting skill, domain knowledge and the ability to ask the right questions to recognize the wrong answers.” ([12:23]) “[AI] is not an answering machine... it's just like that Excel spreadsheet. It's only a tool and it doesn’t know anything.” ([12:49])

• Practical advice: clinicians should use AI, then communicate outputs to patients—not the other way around.
• Organizational investment (like Trupo’s “Bilingual Workforce Program”) into AI literacy is essential pre-requisite for scaling impact.
• Leon: The true bottleneck now is “listening, workflow understanding, knowledge specialization.” ([06:26])

4. Patients as Pattern-Matchers – Data Fragmentation & Missed Diagnoses

Timestamps: [17:02]–[19:54]

• Even highly trained patients (like nurse/informaticist Danny van Leeuwen) struggle with system fragmentation and lack of data usability—his MS diagnosis eluded doctors for 25 years, despite pattern emerging from records.

  “I've been on a mission to gather my own medical data. Two months later I got a box, a four-pound box of paper ... not even in chronological order.” ([18:28])

• Data access ≠ data usability: overwhelming, unsearchable documents render modern patient agency mostly theoretical.
• Early adopter patients abandon the vast majority of healthcare apps after first use; the “rapid skeptic” posture dominates.

  “If the most engaged patient on Earth abandons 95% of health apps, what does that tell us about the consumer healthcare AI market thesis?” ([19:34])

5. Self-Experimentation & Executive Buy-in Accelerate Progress

Timestamps: [20:34]–[22:23]

• Notable “test pilots”: Executives, like Trupo, use themselves as live experiments to find the real-world limits and benefits of AI.

  “I challenged myself to say could I cut the amount of time that I spent in my day job in half ... spoiler alert is I failed. However, I got pretty close ... didn't get 50%, got 30% measured by AI analysis of the calendar over 12 months.” ([21:07])

• “Cloning” oneself (i.e., via AI-generated presentations) is already here, albeit with caveats.
• The future may hold AI-to-AI negotiations (regulators’ AI talking to pharma’s AI)—yet this raises profound questions of explainability, transparency, and trust.

  “It's funny because you can imagine a world where eventually you'll have regulators' AI talking to pharma companies' AI ... That feels interesting where you have both systems generating AI agents talking to each other...” ([22:23])

• Leon: Raises an ethical/practical warning:

  “If the AIs go into some dark impenetrable box and come out and tell us ... 'this drug is approved', I don't think anybody is going to be comfortable with that.” ([24:09])

---

Notable Quotes & Memorable Moments

• Matt Trupo (Sanofi) on “Super Agent” Failure:
  “I started out to create like one super agent to rule them all, and it failed miserably. The reason was it really wasn't specific enough in any single area.” ([05:23])
• David Hidalgo Gado (Clio Health) on Listening:
  “I'm overwhelmingly convinced that in order to add value, technologists in healthcare need to be tremendous listeners. We will never know as much about the best way to use our product as our users do.” ([07:03])
• Barry Chaikin on Prompting vs. Puppies:
  “I do not want you to agree with me all the time. I don't want a puppy. If I wanted a puppy, I would get one.” ([12:51])
• Danny van Leeuwen on Data Access:
  “Access to data and access to usable data are not really the same thing. They're very, very different.” ([18:44])
• On Self-Experimentation:
  “He cloned himself into what he calls Digital Matt, which is given three speaking engagements while the real Matt spoke elsewhere ... In a weird way, it's a little bit of a preview into what the future could be.” ([21:43])
• Ted Shortliffe’s Methodological Correction:
  “Expert systems back when he started, they actually didn't fail. It was just simply too early. Retrieval, augmented generation, knowledge graphs on top of LLMs—those are quietly the 1980s work coming back into fashion.” ([03:36])
• Leon on Industry Openness:
  “Sanofi's transparency is really unusual in the industry and the unusualness itself is a data point ... the rest of pharma is not following him into the sunlight yet.” ([28:13])
• Meta Reflection:
  “The technology is no longer the question. The specificity of the work around it is.” ([33:18])

---

Major Themes Synthesized (“The Five Reflections”)

| Theme | Core Message | |-------|-------------| | Specialization Beats Generalization | Building specialized agents for specific workflows/data outperforms one-size-fits-all AI solutions. | | Symbolic and Neural Blend | Methods from the 70s/80s (symbolic AI, knowledge engineering) are resurging alongside neural networks; old and new are powerful together. | | Literacy is the Bottleneck | Organizational and user literacy (prompting, understanding, contextual knowledge) now separate successful from failed deployments. | | Patients as Pattern-Makers (and Data Victims) | Even trained users are confounded by fragmented and unusable data—AI that merely provides access isn’t enough. | | Self-Experimentation Drives Adoption | Executives who use themselves as test cases accelerate understanding, surfacing real barriers/opportunities quickly. |

---

Conclusion

• Practical Win Condition: The real value from AI isn't in generic solutions or underlying LLMs—it's in “going deep” on workflow, data, and organizational context.
• The next challenges for AI in healthcare are social, methodological, and organizational: depth of specialization, human acceptance, explainability, and patient agency.
• The journey forward is “messy” (as hosts put it): true patient empowerment is “unfinished”, and durable progress will depend on transparency, literacy, and organizational openness.
• As summarized by Labkoff:

  “In that framework, the only durable advantage is that the depth of specific workflows and proprietary data, run by people who have actually used the systems themselves or on themselves ... that’s what's going to win the day.” ([31:33])

---

For Further Listening

• Blocks 1–5: Each reflected stage of the field’s maturity, from mundane wins, history’s lessons, the role of evidence, to infrastructural work, and—now—specialization and depth as the “limiting reagent.”
• Watch upcoming episodes for continued focus on explainability, patient empowerment, and how real-world organizations are navigating the zig-zag progression of healthcare AI.

---

Transcript

[A]

Welcome to Practical AI in Healthcare, the podcast that cuts through the noise to spotlight real world solutions delivering real world value. From patient care to clinical research, from life sciences to patient engagement, we focus on what truly matters in healthcare today. No hype, no theory, just practical insights where AI is making a true impact. Dr. Steven Lapkoff and Dr. Leanne Rosenblitt are your hosts as we explore what's real and moving the needle in this exciting new domain. Welcome aboard and let's get to it. As many of our listeners know, Leon and I work very closely with the DCI Network Division of Clinical Informatics at Beth Israel Deakins Medical center in Boston. This June, the network is hosting Patient powered Digital Health 2026. The conference will bring together patients, innovators, industry leaders, healthcare providers and policymakers to shape the next generation of real world patient centered solutions. The meeting will run from June 22nd to the 24th in Boston at Harvard Medical School. We've arranged for our listeners to get a discount on registration to the meeting. If you Register now before June 15th and use promo code PracticalAI June no spaces, you'll receive 30% off your registration fee. You can learn more at dcinetwork.org patients2026. In addition, we're always looking for sponsors. If you or your company are interested in becoming a sponsor a please reach out to admincinetwork.org see you in Boston. Hello, welcome to this week's edition of Practically in Healthcare. My name is Dr. Steven Lapkoff and I'm here as I am every week with my colleague Dr. Leon Rosenblit. How's it going Leon?

[B]

Oh great, Steve. I'm really excited for our fifth Reflections episode.

[A]

Yep. And today we are going to take that up. We're going to be discussing the last five or six episodes that we did. This block of episodes. We're going to be discussing Matt Troupo's episode. Matt was from Sanofi and he had a two parter that was covering a variety of topics. We also had Dr. Ted Shortliff, one of the founders of Medical AI. We had Barry Chaikin, who is a physician but also a cancer survivor and his episode discussed his journey as the physician slash patient and how he used AI in the course of his of his journey, David Hidalgo Gato from Clio Health and how they' an ambient AI model for emergency room medicine. And lastly a nurse of over 50 years, Danny Van Loen, who has been a serious patient with some very challenging issues for over 25 of those years. And we're going to Tease apart what the lessons learned from each of these discussions was and try to draw some conclusions.

[B]

Yeah. So Steve, what hit me preparing for this one is in Block 5, we try to make a point that infrastructure is being built, the technology is mature and the conversation has moved on. Block 5 says, sure, fine, but the infrastructure only pays off if the thing sitting on top of it is specific enough. As we'll recap, Trupa failed at building a single super agent. Hidalgo Gado spent nine months in one emergency department group trying to nail down their workflows before launching. Chaikin tells us that accuracy only works for the expert. Right? That accuracy depends on how literate your users are, not necessarily on the model. Van Leeuwen had Ms. For 25 years and the diagnosis was sitting in the records the whole time. Yeah.

[A]

And short live almost in passing handed us a methodological correction that expert systems back when he started, they actually didn't fail. It was just simply too early retrieval, augmented generation knowledge graphs on top of LLMs. Those are quietly the 1980s work and 70s work coming back into fashion.

[B]

Yeah, so the working frame for today is a little different. The LLM is the commodity. We established that in block 4. Block 5 makes the claim stronger. The generalist agent is the commodity too. What's not a commodity's depth in a workflow, in a knowledge base, in a patient's own data, or an executive who uses themselves as the test drag and is willing to really do like, you know, six months of AB testing on themselves.

[A]

All right, there we go. So we have five themes. Why don't we dig in and let's dig in with the strongest cross guest signal, that of specialization.

[B]

Yeah, I think the strongest signal across four guests in this block, not all six, but it's the generalist AI doesn't work. Right. So Matt Troopa Sanofi tried to build, in his own words, one super agent to rule them all. Love the Lord of the Ring references anytime. Right. And it failed. You know, and he was very transparent about how and why. So he pivoted to a system of 11 specialized agents like Morning Prep Governance, one on one specific strategy, sorry, scientific strategy, email others, you know, and that, that didn't save him 50% when he was aiming for, but spent some 30% of his calendar time over 12 months. That's amazing. Here's his exact line, and it's worth quoting verbatim. I started out to create like one super agent to rule them all, and it failed miserably. The reason was it really wasn't specific. Enough in any single area. That's not a junior engineer. That's the SVP of Sanofi R and D Tech telling us this. David Hidalgo Gado at Clio Health drew the same picture, but from the vendor side. There are about 120AmbientAI competitors generalizing across all of primary care. Clio went the other direction, he said, and this is direct, solved many problems for the same people. Right. Going a mile deep on the specialty. They spent nine months with one emergency department group, did 40 to 50 iteration loops before they launched anything publicly. And on the data side, Trupo's biologics work succeeded because Sanofi had hundreds, hundreds of millions of proprietary antibody sequences from the Kayam acquisition curated under one protocol. The advantage was the depth of the data, not the bigger model.

[A]

So we've been telling listeners since block four that the LMM is the commodity. Are you and I now ready to say on the podcast that the generalist AI is also a commodity? The generalist agent, that is.

[B]

Sorry, yeah, yeah. My sense is, from the conversations and from watching the industry, is that the agentic workflows and the harnesses that surround them are sufficiently mature where anyone can roll out a generalist agent. And you know, as we, you know, in fact, there are infrastructures for rolling out agent teams. Right. That they're built in, into, right into the platforms. So I think the answer is yes. And this is sort of natural part of technology progression, but it used to be sort of a barrier to entry to know how to do this, but, but currently almost any technologist can get that working. And the barrier is now back to the bottleneck is listening. The bottleneck is workflow understanding. The bottleneck is knowledge specialization of various kinds.

[A]

And that specialization piece, really, from my perspective, it feels like that feels right. And it feels like as we have been listening to our guests, they, they're echoing it. So Hidalgo Gado has this line that I keep coming back to. I'm overwhelmingly convinced that in order to add value, technologists and health, as a technologist in healthcare, you need to be a tremendous listener. We clearly now have never, we never will know as much about the best way to use our product as our users of the product. So that seems to me to be, you know, a signal in and of itself.

[B]

Totally. I also love that line and I really like the insight that his use case brought to our attention of how important listening has become. Right. I sort of drew a little picture of it and we posted in one of them about the pipeline for generating automation and the Bottleneck used to be the building of the automation, right, the programming itself. And it, you know, there was another bottleneck upstream, but it was much less tight, which was figuring out what it needs to be built. So although people in technology understood that sort of, so requirements gathering and analysis and, you know, product definition were important, they weren't the critical bottleneck. Well, guess what happened as new technologies expanded the pipe of, of for building that initial bottleneck, which became the critical one. So I think we're just seeing the transition in where the resource constraints in producing useful, useful systems are. So I, I would still, I would argue the specialization is now doing the work. And Trupo's 11 agents, Clio's mild deep emergency department focus, Sanofi's curated biologics. Here's the twist. The kind of specialization it's winning isn't all neural net. Some of it is the old symbolic AI. Traditionally, you know, that's coming back sort of the, you know, I, I don't want to call them zombies. I, you know, they're not back from the dead. But you know, you get a little bit of that image like, okay, symb, like AI is back, but they're coming in through the side door. And, and Ted Shortlift told us why.

[A]

Well, yeah, and you know, if you go back to Ted. So I've known Ted for well over 30, almost 35 years. And you know, I was a student of the kind of thing he was doing when I was a resident. And, and it wasn't that he had it wrong in any stretch. I mean, the work that he and his colleagues did back in the 70s and 80s was, it was underpowered. Basically. The idea was right. The they did what David Hidalgo Gado had done. They studied the expert work and tried to dig into how the experts were thinking. And as a result of that, they did. They created the expert systems of what they called them in those days, but they just didn't have the computing power to bring them into fruition as they do today. And I think our guests are telling us, both with Trupo and with Hidalgo Gado, effectively the same thing, that if you have the specialization, you dig deep enough, you're going to get a much, much better result. And if you look back and say what, what Ted had said, when I've heard people say, well, the expert systems failed and we had to do something different. And my reaction was, well, yeah, there was all this AI winter that was going on between 1888 and 95. I think his point was methodology wasn't wrong. The surrounding compute memory and data really wasn't ready for the task.

[B]

Yeah, I mean short lived was kind of gentle about this on tape, but the implications are sharp. Is a whole generation of AI researchers was told that the expert systems era just failed and you needed neural networks. He's saying that the framing was wrong. How much of 2026 healthcare AI is actually returned to 1985 with better hardware? And if it is, who's giving those guys from the 70s and 80s who did all of that foundational work enough credit? I am not hearing it, but I think they deserve it. So I think that idea of it didn't fail. It built the foundations with inadequate tools is really important because I think I'm going to say it's obvious. So that means I'm probably going to be wrong, but I think it's obvious that winning solutions are going to be a mix of symbolic and AI and neural networks working together in various ways along the line that sort of Gary Marcus has been advocating for example this

[A]

issue of the Specialist's return and we have the methodological correction from Short Live. Both of these are about what's happening on the AI side. But the most uncomfortable finding in this block is not about the AI at all. It's about the humans that are using AI. You know Barry Chaykin put a number on that and it keeps coming back in my mind.

[B]

Yeah, those damn humans, right? There's an old saying in it is like man, a lot of work would be so easy if it wasn't for those damn users. But joking aside, Barry taken sight of the study he attributes to nature and the numbers are stark. The same AI tools but physicians had almost doubled the diagnostic accuracy of patients. The model was identical, the interface was identical. The variable was prompting skill, domain knowledge and the ability to ask the right questions to recognize the wrong answers. So Chaykin's metaphor for what AI actually is, which stuck with me, is like a da Vinci surgical robot. Right? It's not an answering machine. So his quote was AI is not going to be able to answer my questions and do this kind of work for me. It's just like that Excel spreadsheet. It's only a tool and it doesn't know anything. And it's prompt for configuration is one of the greatest anti sycophancy lines we've recorded. I do not want you to agree with me all the time. I don't want a puppy. If I wanted a puppy, I would get one. So I'm adding that to all of my prompts. So his practical advice to clinicians sits underneath all of this, right? Because he's a clinician and a patient. Don't give the patients the AI, use the AI yourself. Generate the patient information sheet, give that to the patient. And the kicker, quote, talking about clinicians who do this well, is if you don't already, if you don't already think highly of you, they'll think even more highly of you. I think you meant if they don't already think highly of you, they'll think more highly of you. Trupo, on the institutional side, made the same investment, but a different way is Bilingual Workforce Program. You know, hundreds of executives trained and thousands of employees in online tools before they try to scale the technical work. Literacy clearly matters to people working on this.

[A]

And I'll just add to that, like we've, you know, Leon, you and I at the DCI Network, one of the huge findings that was a surprise finding for us six months ago when we were doing our work at the Single Tooth Noise conference, echoed that at all. And we sort of tripped into that. Right. We didn't expect that AI literacy was, was even a thing. And then as we went through the conference, you know, with folks like Chris Demick, who is xhrq, we realized, yeah, in fact that specialization actually matters. That degree of literacy matters, and it enhances the output. And I think both Trupo and Barry Chaikin basically have said exactly the same thing in different ways in different domains, in different walks of life and different jobs. But the message is clearly the same. The degree of literacy that you have in your tools really matters. So, you know, with regards to, you know, I want to move on a little bit so we can cover through everything we have in our, in our episode today. Getting back to Hidalgo, he saw the variability inside one product mattered. So Clio scribed the same workflow. The savings range from way less to upwards of two plus hours in a clinician's day, depending on how the given clinician used it. I don't know if that was an accuracy thing from literacy or what, but it feels like it might be part of it and it might be the fact that he and his team spent so much time working on it. But that really is a vendor side literacy variation. Right. So they became literate in how the clinicians were doing their job and translated that into the tools they built. And that degree of literacy mattered because it gave the folks using it that much more energy to be able to get more out of it. That, to me, feels like an important piece is this A new performance variable that we should be tracking on in every deployment. Not does the tool work, but who's using it well and who isn't and why?

[B]

Yeah. You know, as a social scientist, I think the concept of tracking variability is just always really important. Right. So when we look at how something happens, the, you know, our initial instinct to look at the center of mass of the distribution. Right. Then sort of look at the mean and the median. But where people actually land and who lands where with a new deployment is probably more important. So I really think that's a powerful insight. If you can look at the variance in how different users and different classes of users are actually adopting to whatever the technology is as a vendor, you're probably going to get further along. And if you can focus in tightly on one class of user and help them with their specific problems, you're more likely to actually jump to something that delivers really high value. But that sort of brings us to our next theme. So Chaikin showed that the gap between trained users and untrained users. Danny van Leeuwen showed us something different and harder. Even when the patient is a trained user.

[A]

Right.

[B]

A nurse of 50 years, an informaticist, full access to his record, the system can still miss the pattern for 25 years.

[A]

Yeah. And that's a consistent problem that his dog medicine. As long as I've been practicing, you know, the whole issue of things hiding in plain sight, especially complex diagnoses, is really, is really difficult. I think that when these new systems come online, if they're able to take an objective look at data from, like, the records that you can feed them, it's highly likely, especially if you're well trained. And I think that's another important feature. If you're well trained in being able to interact with them in a literate way, it shouldn't take 25 years to get to a diagnosis. Even these complex diagnoses should, in theory at least, float to the surface. Danny was a nurse for 50 years, and he's also been an informatician. He's also been a caregiver through members of his family. His Ms. Went undiagnosed for so long, the episode triggered every episode that he had was triggering a cardiac workup, largely because he had a cardiac event in his father when his father was a young man. And therefore anything that happened sort of tripped people up to think, oh, it must be a cardiac thing. But the cardiac wasn't the issue. They weren't piecing it together. You know, he's a licensed nurse with full access to his own record up in Boston and he, he couldn't get the dots connected. I think when he tried to get his data so he could, he could synthesize it himself. Here's what happened. I've been on a mission to gather my own medical data. Two months later I got a box, a four pound box of paper and it was paper that was not even in chronological order. That's a data problem. He also then went on to show that of another practice he had 296 pages of redundant, non searchable PDFs. I'm betting where I know where that came from, just from three months of visits. His line is that access to data and access to usable data are not really the same thing. They're very, very different.

[B]

Yeah, that giant box of papers really struck me. Danny also said, and this is interesting on its own, that he's tried over a hundred healthcare apps and uses five more than three times out of trying 100. Right. So there's most of them used once and dropped just five stuck around. So the quote is, I'm both an early adopter of technology and a rapid skeptic. If the most engaged patient on Earth abandons 95% of health apps, what does that tell us about the consumer healthcare AI market thesis? And is the rapid skeptic posture actually the right posture for valuing these tools from a patient's point of view?

[A]

Yeah, so that's actually what I do if I get these new apps personally, if I can't figure them out in just short order, I abandon, I try to get the things that really work and I consider myself pretty literate in the use of these apps and in AI. And it's challenging. So we have specialization, the return of symbolic methods, the literacy bottleneck, and the patient as a pattern maker. There's one more piece that we haven't really talked about today and it's the most personally audacious thing I think a guest did in this block and it was Matt Troupo in part two where he, he decided to use himself as the tester for what he was trying to do.

[B]

Yeah, I just loved his audacity and his scientific spirit applied to his practical a to day work. He just said himself an explicit goal, cut his work day in half. And he told us on tape, I challenged myself to say could I cut the amount of time that I spent in my day job and half so I could refocus those efforts and stuff we're talking about now. So the spoiler alert is I failed. However, I got Pretty close in the first iteration. So you didn't get 50%, you got 30% measured by AI analysis of the calendar over 12 months, six months before and six months after. So here's the architectural lesson told from the inside. And it's the same one that opened our episode today. He started with one super agent. Right. Which I think a lot of us do. It's like build an agent that does all of my day to day stuff. The pattern monolithic generalist fails, but the specialized ecosystem works is now his explicit read of where the industry is going. And he went further and he cloned himself into what he calls Digital Matt, which is given three speaking engagements while the real Matt spoke elsewhere. The clone writes its own scripts from project documents trained on Trupo speaking voice, his framing on tape. It's still a little bit of a joke, however, in a weird way, it's a little bit of a preview into what the future could be. And the long horizon read from him about orchestration is, you know, I'm quoting here. Yes, we're getting periods of acceleration within a function, we're getting more efficiency within a function. But to really bend that cost curve and the time curve across the entirety of R and D process, it needs to be done in an automated way. And I think that requires that orchestration of all of those agents working together. Hidalgo Gado gave us the vendor side version of the same anxiety. What keeps me up at night is just understanding that the pace of change is greater than we as humans can really comprehend.

[A]

Yeah. And you know, at the end of the day, I think both of them, and it's interesting that they both arrived at the same place from a different perspective. You know, Troopal also said this, and I want to read the verbatim because it's a little uncomfortable. It's funny because you can't, because you can imagine a world where eventually you'll have regulators AI talking to pharma companies, AI, and they'll be asking questions and they'll be responding and then you're getting everybody in the room at some point. That feels interesting where you have both systems generating AI agents talking to each other and then eventually you've got to come together.

[B]

I mean, it does make me wonder if that's the future we want. It also reminds me about our earlier thread where we had, we talked about what's going to happen with provider AIs talking to payer AIs to negotiate pricing. Right, Right. I mean, it's happening. We certainly have people, we've talked to people on one side building the tools and we know that their tools being built on the other side. So it's only a matter of time before you have AI to AI negotiations. And it's possible that those will develop into AI to AI trusted spaces that provide the opportunity for moderated conversation that humans can understand and trust. Right. Because I think we need to build these out in a thoughtful way that provides human oversight and a level of transparency that makes so comfortable the decisions make sense, they're audible and we get them. Right. So I think there are one vision that I think would be unacceptable. If the AIs go into some dark impenetrable box and come out and tell us where the decisions are. Like this drug is approved. I don't think anybody is going to be comfortable with that. Right.

[A]

No, I don't either.

[B]

So I think we need to be realistic about engineering for human acceptance and not just like, well, what's technically possible. So you know, so that's one sort of big theme from, you know, that I always walk away from, but let me kind of bring, bring us together to our meta reflections. Right. So we've had five themes. Specialization beats generality. Symbolic methods are coming back. Yay. Literacy is the bottleneck. Patients can pattern match better in the system when they have the tools and the executives who matter are the ones who are using themselves as a test platform. So let's talk about what that does to the thesis we've been building since episode two. Explainability just, I mean it modifies a lot of things, but one of them that jumps out for me is explainability is not the same thing as transparency. So short lived says human explanations are also post hoc reconstructions. And maybe we should not be too worried if LLM narratives doesn't match. It's actual computation. Trupo says regulators accept modeling provided you can explain how you did it. Chaikin cites open evidence linking not the company, just actual open evidence linking every recommendation to a reference, three different bars and block three row and white. Discussion of liability becomes harder, not easier under short lived framing. Again, you know, if transparency isn't a requirement and justification isn't a measure, are we comfortable with that?

[A]

Well, you know, that's a really interesting thing and I think the tension on that front is going to be, get, get harder as we move forward, not easier because both transparency and explainability are still lagging. Although in a future episode we're going to be in, we're going to be speaking to a company that makes the claim that they're able to explain how neural networks work. That'll be coming up in about four to six weeks. I screened them earlier this week. But at the end of the day, until those issues are there, I think it's going to get harder, not easier. But, you know, if we go back to patient agency, you know, with caveats in Block 4, Shakar and Price gave us patient authorized data and participatory medicine. In Block 5, Chaikin and Van Leeuwenhad the patients who can take advantage of all this are highly self selected. How do we hold both truths together? Patients should own their own data for sure. Most patients are not yet equipped to do it directly. They're not literate enough in the use of the tools without sliding into either paternalism or naive empowerment. Where do we go?

[B]

Yeah, I go back to A.B. price's quote, which is, it's not broken, it's unfinished. We are. The process of giving the patients the tools and teaching them to use their data is going to be slow and messy. Right. And I think it's going to have a lot of dead ends and a lot of rough edges. But the only way to get started is to get started. And, you know, so what's happening now is going to have downsides and we are going to have to be very deliberate and careful in finding ways to address those downsides. But in my mind at least, I just see as an unfolding process that leads to better places, even though it's going through some rough terrain.

[A]

Yeah. When you get into that, though, I think it raises a, you know, how permissible. You know, we're in this environment where we have the challenges of the legal system, we have the challenges of accuracy, we have the challenges of patients pushing harder and harder into the world that they want to see better medicine. And at the same time, we're going to have to tolerate in some degree inaccuracy. We're going to have to tolerate the, you know, where Barry Ch. Where Barry Chaykin was illustrating how the experts get better results versus the patients who are getting somewhat worse results. We have to live through that in a way we can get through it intact. And I'm not sure how that's going to play out. It's going to be. It's going to be messy.

[B]

I think it's messy. And this whole thing is generating an amazing amount of political resistance. Like if you go, you know, on Reddit or online spaces, you know, the discussion among young people is like, why do we need this AI thing at all? Like, it's, it's just very Unpopular, you know, you and I aren't living in that world. But unless we both solve the problems and articulate the positive sides of the solution and provide transparency, it's going to encounter increasing political resistance for people who kind of just freaking out. Right. So it's a, it's a, it's a lot of technical change. It's a foundational set of technologies. So I mean, one example of that is what you're seeing from that Trupo. You know, Sanofi's transparency is really unusual in the industry and the unusualness itself is a data point. Most pharma companies treat this kind of AI work as a competitive advantage. Right. It's like, you know, we've, we talked to many pharma and you've worked inside pharma. They're like, we're doing AI right? And you know, Steve's, Steve's our guy to do the AI and, but we can't talk about it outside the, the company. The Trupo treats it as something to talk about in a podcast. Either his openness is generous, which I think it very much is, or strategic, which I think it very much is, or both, which I think it very much is. Right. Either way, the rest of pharma is not following him into the sunlight yet. Right. I mean, we are working with several with pharma in pre competitive spaces where we have open conversations, but none of it is meant to be public. So what does that say about the maturation story we told in Block four?

[A]

Well, it's a great question. I think that, you know, there's not a straight line through all of this, right. It's going to be zigging and zagging. And I agree with you that I thought Matrupo's openness and Sanofi's agreeing to have something that public at this stage of the life cycle is remarkable. I mean, when I was inside various companies, we weren't allowed to do things like that. To go out in public and basically show what's underneath the kimono, so to speak, that was a really remarkable thing. And frankly, we didn't reach out to Matt. Matt's and his organization, they called us, which blew, you know, to even now blows my mind. But it's going to be zigzags, right? We're going to be, it's going to be like skiing down a mountain. You can't go straight down. If you do, you're going to wipe out. And if you go slowly side to side, you're much more likely to get to the Bottom without a broken leg. And I think that's where we have to. Have to navigate and we have to be patient with that. It's going to be something.

[B]

I love that metaphor.

[A]

That's a.

[B]

That's a really good metaphor.

[A]

And.

[B]

Yeah, I haven't heard that before, so thanks.

[A]

Well, so where does that leave us now? We have several blocks that we've done. I'll recap them. In block one, the big theme was that mundane matters and mundane solutions matter. And I think we've seen plenty of evidence that that's the case in block two, that history doesn't repeat itself. It tends to rhyme. And we should learn from the EHR experience. In block 3, evidence constrains. In block 4, infrastructure is where the actual work happens. And then block 5, depth. Symbolic methodology or symbolic methods. Literacy and self experimentation are limiting reagents. The LMM might just be the commodity, and the generalist may. Well, the generalist agent might well be the analogous commodity. In that framework, the only durable advantage is that the depth of specific workflows and proprietary data run by people who have actually used the systems themselves or on themselves. I think that's what's going to win the day. And that means specialization and special training and literacy actually matter. And I'm thinking we're seeing evidence of that in multiple places, in multiple guests, from multiple different perspectives in their own fields, and they don't overlap. It's really interesting to me.

[B]

Yeah, I'm getting the same read from this section. So to bring it together, block five gives us the limiting reagent. The AI works. That was block four. The infrastructure is being built. That was block four, too. The question now is whether people, the methods, and the organizations around the infrastructure are specific enough to extract the value. Trupo and Hidalgo Gado say depth beats generality. Shortlist says the old symbolic methods are coming back through the side door. Chaikin says literacy decides who gets high accuracy. And Van Leeuwen says the patient is the pattern maker. The system's been missing. And truepoint Part two tells us the executive has to use themselves as the test trick. Right. You know, it's an uncomfortable reality we're living in. We're all our own guinea pigs at this point. Usually we experiment just in our children with technology. Now we all get to experiment in ourselves. So the thesis we've been building since episode two now has five layers. Mundane winds, matter, history, rhymes, evidence, tests and constraints. Infrastructure is where the real work happens. And depth, which is specialization, symbolic knowledge, literacy, self experimentation is the limiting reagent and where there any of it actually pays off, that's where we are. After 37 episodes, the technology is no longer the question. The specificity of the work around it is.

[A]

Yeah, and that's really insightful, Leon, and I love how you are able to really tie things together like that. So thank you for that. And, you know, we'll be back over time. We're going to have really a whole array of additional discussions in the coming weeks, as I alluded to a few moments ago. But before we go on, I just want to say thank you for listening to us and thank you for tuning in. Every week the podcast has grown rather remarkably. We started off with, you know, 80 people listening to it back when we started in September, and we are kicking out thanks, Mom. But now we're, we're measuring our listenership in the thousands. And that's, that's remarkable. And I want to just say if you have a guest that you'd like us to interview or you want to be a guest on the podcast, send us an email. You could reach me or Leon. Our respective emails are Steve at practicalai and healthcare.com or Leon at practicalai and healthcare.com and if you have a topic that you're interested in, send us a note. If you want to recommend somebody again, send us a note. Yeah.

[B]

And if something in today's conversation resonated, share it with a colleague. And please join us next time on Practical AI and Healthcare.

[A]

Thank you for joining us this week on Practical AI and healthcare Care. If you're ready to go beyond buzzwords and hype and explore how AI is truly transforming healthcare, stay tuned for more conversations that get us to what works. Until next time, stay, practice.