Pharma's Coordination Tax & the 2 Paths That Could Actually Disrupt It | Alex Telford (3/4)

Intro - 00:00:06: Welcome to The Biotech Startups Podcast by Excedr. Join us as we speak with first-time founders, serial entrepreneurs, and experienced investors about the challenges and triumphs of running a biotech startup from pre-seed to IPO with your host, Jon Chee. In our last episode, Alex shared the nearly seven-year consulting run that began mid-acquisition, the craft of giving clients opinions rather than just data, and the GPT-3 moment that reframed everything he thought he knew about how work in his industry would be done. If you missed it, check out part two.

In part three, Alex talks about the structural reasons pharma is so difficult to transform, the enormous coordination tax of siloed organizations, tacit knowledge locked in people's heads, and the challenge of getting teams to encode how they actually work. He draws a vivid analogy to factory electrification: the efficiency gains from general purpose technology don't materialize until you redesign the whole organization around it. He gets specific about the two paths that could actually disrupt incumbent pharma: genuinely differentiated target discovery science or vertically integrated direct-to-customer commercialization, and why competing on operational software efficiency alone isn't the lever.

Alex Telford - 00:01:41: No. I do think you do need to recreate the orchestration. Like, I made the point about electrifying the factories back in the onset of electrification. I think you have to retool the organization around these new technologies, and that has happened in many technologies—like any big technology wave, if we have a new general-purpose technology, you need to rearchitecture the organizations. And some will not be able to rearchitecture, and they will die. Some will be able to rearchitecture and they will thrive. And there'll be new organizations that will come and disrupt the incumbents that have failed to rearchitecture.

I fully expect there's gonna be a huge effort over the coming years to go and digitize knowledge and to find the data in the boxes in the warehouse and put it into the system—that's gonna happen—but also to encode the information that is not written down somewhere. It's tacit knowledge in people's heads. Now if you work with a drug discovery or a regulatory expert, they're gonna have thirty or forty years of knowledge about how to interact with the FDA and why we made all these decisions about our drug programs and all this knowledge about the organization and the world and how it interacts with the world. And when they retire, that knowledge is—it may be somewhat disseminated before they left, but most of it is gonna be lost. The really deep, tacit expertise is gonna be lost.

It seems obvious to me that it is better to find a way to encode your employees' knowledge in some kind of system that everyone can access in the organization than to not encode that knowledge. It's like a strict benefit. If you think about what are the jobs humans are gonna do if AI gets very good and takes all the message passing and information synthesis jobs, it's gonna be the finding and encoding the knowledge jobs because there's still so much to do. There's so much knowledge. People say, "Oh, we're running out of tokens" or whatever to train the model. I don't think we're even close to running out of tokens. There's orders of magnitude locked up within enterprises that is digitized or not used to train the models yet, or is not digitized. And there's even probably orders of magnitude more than that in human heads that they're not being distracted in any meaningful sense.

Jon Chee - 00:03:34: Yeah.

Alex Telford - 00:03:35: So I think, when we work for the company, a lot of what we're trying to do is figure out, if you try to automate some process, how do you do some process? How do you think about this process, and how do you do it, and how can we replicate that in some semi-deterministic workflow or get an agent to do it? And when you actually get people to try and write down the flowchart of their job or their task, it is not easy. People do not generally have a good mental model of how they do a task that they can write down as a flowchart. Maybe if you're an engineer or something, you're used to thinking in that way. Most people are not used to thinking in that way. So, extracting the flowcharts of tasks and putting them into automated systems will be a major effort of the coming years—this workflow design or process encoding or something, whatever you wanna call it.

Jon Chee - 00:04:24: That's funny that you bring that up. I was literally doing that with the team this morning. We have these Wednesday meetings to just chat about—just no agenda in the meeting. And it was literally this topic. And there are two things I've been thinking about on this front. First, on the large pharma side, they have so much—when you think about a large pharma, rearchitecting the org sounds painful as hell but necessary. But, also, they are the ones that are just sitting on so much data. So I was thinking, for this exercise: do you think it's freaking possible to just get all that knowledge into this digital brain and then rearchitect an org that has, whatever, 40,000 people?

Alex Telford - 00:05:11: I think it's possible in the sense that anything is possible.

Jon Chee - 00:05:15: If there's a will, there's a way.

Alex Telford - 00:05:16: Yeah. I think it can be done. There's a feasible path to do it. I think you do have organizational resistance. It is challenging to change people's ways of working. Most people do not seriously want to engage with the topic of, "How do I automate my work?" And if you've joined a job—let's say you're a medical writer at a pharma company—you joined to be a medical writer. Most of the time, you wanna work nine to five and go home to your kids. You don't wanna do medical writing plus have another job as a systems engineer vibe-coding. That's not what you signed up for. And most people who are medical writers are not that into vibe-coding medical writing agents. It's just a different skill set and a different set of interests.

So that's hard to forcefully do that. And then, you know, if you have these top-down initiatives, like, "Oh, you have to use these systems," it's hard to get effort out. You need to take effort to get value out of AI systems. They are not like people that you can just drop into a task and do the task for you. They have weird failure modes, and they don't really understand things in the way that you would expect a human to understand them. They have very deep knowledge of facts of the world, but they also have very weird misunderstandings of things. And they're kind of naive in many ways. So they'll give you bad results if you prompt them badly. What I see very often is someone tried to use AI. They're like, "Write me my FDA submission," and it puts out the stuff that plausibly looks a bit like an FDA submission. They're like, "Oh, this is not very good. This is not reusable." So I'm just gonna go use my medical writing agency. But then you don't wanna think about, "How do I architect the system to actually properly write the submission?"

Jon Chee - 00:06:55: Yeah. You just want it to one-shot the submission perfectly. And I think about that. I wonder if you're looking at the crystal ball—is there gonna be a new—because I always think about this, large pharma. Pfizer was founded a long time ago, and Pfizer became Pfizer over a very long period. Will this enable a new, quote-unquote, "large pharma" to appear that has taken everything that you described from an organizational structure to actually compete with a large pharma? The closest ones are Alnylam and BioMarin. Right? But if you have the ambition to compete on that stage and now you're doing everything that you have just described to reduce all this coordination tax, isn't your cost structure and efficiency so great that you can feasibly disrupt and compete on that stage?

Alex Telford - 00:07:52: I think maybe it seems plausible, but my intuition is that that's not the axis you wanna compete on as an incumbent pharma company. You're basically arguing that you should compete as a software developer. So you should build systems for information. What is software? It's information manipulation. So the way you're gonna win is by more efficiently manipulating information than your competitors. And that's gonna give you some enduring advantage; it's gonna help you scale from nothing to disrupting incumbents. I think that's probably not true.

I think the way that a pharma company—like a new biotech—would disrupt a pharma company is probably more aligned to the fundamental way that value is generated in the industry, which is target selection and maybe some aspect of commercialization. But it's ultimately target selection and being able to capture differentiated data that helps you identify and validate targets faster. So I think that'd be one way you could disrupt. Another could be some DTC type thing where you start with going direct-to-consumer, then you go back into developing your own drugs.

The software stuff is nice. And obviously, I think it's useful because I'm working at my company. But I don't think it's like—if I was gonna start a pharma company from scratch with the express goal of trying to disrupt the incumbents, I would not start by going, "Well, how do I properly use AI agents to write regulatory submissions?" You only get benefits from these things at scale. It only helps you when your information is so large that you then need the system to—it's like a cold start issue. Actually, you'd much rather spend your time on, "How do I find a fantastic target that's gonna give me an amazing drug that's gonna generate incredible cash flow?" And I can use that cash flow to find my next opportunities and scale my business and become this engine of running many programs.

I think you definitely invest in AI and automation as you build, but I wouldn't put that as my number one thing. Or you would find some way of collecting consumer data or commercializing with consumers that's differentiated. Like LillyDirect is very interesting because they have their GLP-1 franchise, and consumers can go into a digital platform and get access to it and get prescribed. And then they're buying this orexin drug, which potentially has applications in sleep beyond rare sleep conditions. And then maybe they'll buy a hair loss drug. And you can imagine this world where Lilly becomes your partner for just general health and well-being. And it's like, as I'm aging, I'm just gonna go to Lilly to get my weight loss drug and my hair loss drug and my skincare stuff. And maybe if I get really sick and I get cancer, I'm gonna go to the specialist doctor and get my drugs with a typical pathway of prescription. But for most general health things, maybe I'll just go through LillyDirect.

So, another way to disrupt pharma would be something like you bundle up a bunch of these OTC-type or DTC-type molecules with a telehealth platform, and you vertically integrate. And then you become a preferred partner—I can just order medicines on my phone. So, yeah, those are the two ways I see it. It's like you do real novel science, fundamental target discovery—you have some great dataset for a kind of radio flywheel of data—or you just go direct to the consumer and you disrupt the commercialization model.

Jon Chee - 00:11:09: I mean, just look at Hims & Hers.

Alex Telford - 00:11:11: Yeah. I mean, they're—there's a partnership with Novo, I think, recently. I think that's one way you might do it. It's like a Hims & Hers type shape of company. I mean, I don't love how they're going about it.

Jon Chee - 00:11:21: For sure. But just DTC in the zeitgeist.

Alex Telford - 00:11:25: But maybe that is a way you could do it.

Jon Chee - 00:11:28: For sure. And the other part that I was just thinking about—we've always been a remote-first company, even before COVID, and very much in written chat, and we use Gong for call recordings and stuff. And we're currently trying to figure out ways to—all this tacit knowledge that you're talking about, just get all that data so we can act on it. And I'm biased. We don't go into the office. I'm just like, but you could probably just lose the tacit knowledge if you're not writing words and recording these calls over Zoom and stuff. You just lose it. And it's just not accretive to—if we're living in an agentic world, it's just not accretive to that.

And at first, when Bridgewater talked about things—like Ray Dalio—it was like, "I record every meeting ever." Okay. Well, you probably have some crazy data if you've been doing that for a really long time. It's kind of like when the founder of Shopify said, "I've had a keylog and screen recorder on my laptop since the dawn of time, and I just use that to train." Anyways, you're talking about where do we get the data becoming increasingly important. I can see it firsthand.

Alex Telford - 00:12:42: The way I think about that is I think the fundamental engine of the business has to be very strong. And how you actually generate value for customers—it's probably not by better summarization of your calls. Better encoding of your meeting transcripts and transcription of your transcripts and access by some agent is gonna generate some business value. But it's not the core edge of the business. So I think it's pretty easy to get distracted with this task of, "Let's encode everything and let's have incrementally better knowledge." But I think it's better to be thoughtful about what you're encoding and how does that tie into what is fundamentally driving value for the business.

I didn't fully answer the question about data: "Can a pharma company just ingest all the data and get value?" I actually think a lot of that data is probably unusable. It's not been collected with the right structures. It's not been annotated properly. The context in which the data was generated is lost. So I don't think just putting in more data is good. Actually, the problem—honestly, a lot of the problems we have in building these systems is: how do we get data out? How do we improve signal-to-noise? Signal-to-noise ratio, I think, ends up being much more important for getting good performance out of LLM-based systems than just the raw amount of data.

One of the pathologies that language models have, which may be impossible to get rid of, is they attend too closely to information that you've put into the context window. And maybe it's something to do with how they're trained. Typically, when a model is trained, especially as they're doing more reinforcement learning, the information that models are fed in these task environments are often toy examples or synthetic examples where the information that's relevant to solving the task is contained within the environment. But in the real world, often the information required to solve a task is not within the original set of information you're given—it's outside; you need to go and find it.

Models are trained to act as if everything they've been given is important in some way, and they'll just run with things. But actually, in the real world, often the answer is, "None of this is useful. I have to go and do something else, something fundamental. I have to go collect some new data." Or, "This data is just flawed in some way. It's fundamentally unusable." But the models don't do that. They just try to run with whatever you give them. So optimizing signal-to-noise is a far better way to spend your time than let's just ingest every piece of possible data and just let the model figure it out. Because typically, I find that they don't just figure it out. They just get confused. Even though in isolation more data is better, it doesn't follow that you want to put every piece of possible data into the system.

Jon Chee - 00:15:28: And is getting the signal-to-noise ratio to improve on that—is that just literally: we just need to roll up our sleeves and just clean this up and get it legible?

Alex Telford - 00:15:39: Yeah. I think it's, unfortunately, kind of boring, and that never changes. You know, there's always some, "Oh, we have Power BI, and we have Tableu, and reimagine enterprise analytics with intelligent dashboards." And then you have this cool, fancy dashboard and the underlying data is garbage. Then you have to go and clean the data. But people don't wanna do that. That's hard. But now we have AI, and AI is gonna do everything for us. But now if you give it garbage data, it's not gonna give you good answers. So you have to go back and clean the data.

Same with Renaissance Technologies. There's a pretty famous anecdote about them where it's like: why do you hire PhDs to do all this work? What is their edge? The edge is just going in and getting these incredibly smart people to clean data and try to understand simple regressions—simple models on very clean, very highly selected data. Picking what variables to regress. It could be one X against one Y. And building these simple models on very high-quality, robust, highly selected data—optimizing signal-to-noise. And that was their edge. It wasn't necessarily some hyper-complex analytical models. It was good data cleaning, robust data architectures—the basics. So I just think doing the basics really well is incredibly difficult and boring, and it's not a satisfying answer in some ways.

Jon Chee - 00:16:58: I actually find that incredibly satisfying. Maybe I'm just a weird dude.

Alex Telford - 00:17:03: Well, I like it too, but...

Jon Chee - 00:17:05: So we're weird, basically. At the end of the day, the way I think about it is—this is the same thing when I was starting the company initially. It's just: you guys don't wanna do the work. You don't wanna do it. I will gladly do it, and I actually find this entertaining. And that's my edge. I will sift through the mud for long periods of time. I know what to sift through and what to organize in order to make this effective, but you guys don't wanna do it.

Alex Telford - 00:17:33: Yeah. I think now we're in a bit of a mania phase where I'll get on calls sometimes and people will show me their vibe-coded apps. "I spent twelve hours vibe-coding this crazy custom solution." And it's like, okay, well the underlying data is not ready. But you could just create these crazy programs. "I can make my own software, and I can make my own dashboards." If it's disposable software—disposable software is great, I think. If it's like a one-off analysis or you wanna make a slide deck, these AI tools are great for doing disposable analysis. But I think you can get lost in the sauce where you think that just because you can make these cool front ends and dashboards, you're doing something useful. But actually, what you need to do is go back to the fundamentals of: how should this process work, and what data do we need, and how do we clean this data, and how do we ensure that it's robust? And that is just not very sexy. It is much more fun to just vibe-code something with Cursor or whatever and have a cool dashboard. You're like, "Wow, this is sick." And then you just throw it away because it's actually fundamentally not that useful. But for a few days, you felt like you were making a lot of progress.

Jon Chee - 00:18:48: Exactly. I sleep easy at night knowing that people don't wanna do that unglamorous, unsexy work. And if you're willing to do that for a really long period of time, you just pull ahead. It compounds. For anyone who thinks they're behind, just note: there's a lot of stuff that you can just chop wood on and pull ahead because people don't wanna do it.

Alex Telford - 00:19:10: Yeah. For people who are scared about labor automation, I just think there's a lot of wood in the world to be chopped, and there will always be a need for people to go and chop the wood. I just don't see a path for LLMs obviating all work. Certain types of work will go away if you're like a dashboard engineer or something. Sorry, probably that's not valuable anymore—data science dashboard generation—or if you're just a specific front-end engineer, probably your job is at significant risk. If you're someone whose job is you know facts about the world—yeah, probably at risk, unfortunately. But there's just so much to do, and the penetration of these tools is so low, and there's just so much work in the world to be done. We have so many problems. LLMs are not gonna solve all the problems in the next ten years. I mean, some people are AI boosters who think it's gonna automate all labor, but I just don't see it. There's so many problems.

Jon Chee - 00:20:04: Likewise. So, you're at CRA. You spend a decent amount of time there as a consultant. When did you know it was time to start your own company?

Alex Telford - 00:20:14: I didn't really set out to start a company. I've been really interested in natural language processing for a long time. I'm pretty interested in the science of the brain and consciousness and those kind of questions.

Jon Chee - 00:20:24: And that was on the side?

Alex Telford - 00:20:26: Yeah. Because it's just like: how does a brain process language? That's an interesting question. And then how does brain process language map onto how computers process language, or how they did before transformers and NLP tools. I was pretty into NLP. And I also felt like NLP was very applicable to the work I was doing because in pharma, so much of the data is unstructured textual. You spend a lot of time extracting information from text or summarizing information, or comparing strings against each other.

A disease in one paper might have a different name than the disease in another paper. It might be muscular atrophy in one paper; another paper, it might be muscular atrophy, comma, spinal. Just doing the string comparison is something that takes a lot of time to clean these datasets. So I thought, "Okay. Well, I spend a lot of time cleaning datasets. Surely, NLP means the computers can help me clean the datasets for me, and me and my teams don't have to do it." That's what got me interested in the domain initially.

I was super interested in word vectors, which is this idea that you can encode the semantic meaning of a word as a string of numbers. There's a famous example of being able to do math with words where you encode the meaning of "king" and "queen" and "woman." And if you do this operation where you take king plus woman or something like that, then you get something close to the vector for "queen." So it's this very cool idea of, "Oh, I can do math with words. I can encode strings of these mathematical objects I manipulate." And then I can do these comparisons—"king" is similar to "queen" and "man" is similar to "woman"—and I can cluster it. I can apply that to drugs as well. I can use that to disambiguate terms and things.

So I got pretty interested in the topic. LLMs built on that premise. When GPT-2 came out, that got me pretty excited. And when the GPT-3 API came out, I was playing around a lot with that and trying to figure out use cases. And I think that was just ultimately what got me really excited. After GPT-3, that's when I had my AI psychosis moment, like a lot of people were having. I was playing around with this stuff all the time. I was like, "Oh man, this is how a lot of work is gonna be done in the future. And clearly this is so applicable to my industry." And I feel like I'm spending a bunch of time making PowerPoints and doing things that feel like they are not the future of work. I've been doing this job for seven years, and it's fine and I like it, but I just wanna explore what this technology means for how work is done in the industry I care about.

So I didn't necessarily set out to really start a company. I just felt like I wanted to build something that uses technology in my field. So I just quit my job pretty cold turkey with nothing lined up. I said, "I'm just gonna do some writing for a bit, do some blogging, and then I'm gonna play around with LLMs." This is around GPT-4 time. And maybe I'll make it a company—I don't know. I'll come up with some ideas, make some prototypes, and see what excites me. And then I ended up meeting my cofounders and getting some investment, and eventually a company coalesced around the idea.

Jon Chee - 00:23:24: Did it start from your writing? Is that how everyone gravitated, or was it a more concerted effort?

Alex Telford - 00:23:30: Well, my cofounder emailed me because he heard me on a podcast, and I was on that podcast because of a piece of writing that I did. So in many ways, the writing ended up nucleating a lot of these things. And one of our investors—our lead investor—found me through writing or through Twitter. Would I have done the company if I hadn't written anything on the internet? Probably not, because again, I wouldn't really have considered venture capital. I probably would have just played around with the idea for a bit and maybe started a bootstrapped company. I don't know what I would have done. I only got conviction in the company idea later when I was building all these little prototypes.

In building the prototypes, I was learning that the problem is much deeper than you can just build a few simple applications. Actually, the problem is fundamentally: you have to rethink how data is architected and organized at organization scale. You need systems where this data can be robustly collected and validated and served to many different downstream applications. So that enterprise context layer—whatever you wanna call it—is where the real opportunity sits, not at these individual point applications. Because so much of the context that you need for any one workflow is context that is shared across many workflows in an organization. If you wanna build one point solution, you have to—it's like to bake a pie, you must first create the universe. It's sort of how I feel about it.

Jon Chee - 00:24:51: Yeah. I love that the writing was the catalyst for all this. It was something that you're passionate about and found really fascinating. I encourage anyone to just put it out there. I used to be like, "No, I'm not gonna put anything out there." I was afraid of looking bad or whatever. You just gotta fucking do it. Even for the podcast, I was like, "Shit, I'm kind of a perfectionist." And you're talking about not being an optimizer. I'm a perfectionist-optimizer. And I'm just like, "God, this is never good enough." But sometimes you just gotta hit publish and just send it.

Alex Telford - 00:25:27: I'm in a little bit of two minds about that. I don't think you wanna put out bad stuff.

Jon Chee - 00:25:30: I'm not saying bad stuff, but is it that "great is the enemy of good" or whatever?

Alex Telford - 00:25:36: Yeah. You gotta put stuff out there. The nice thing about starting when you don't have an audience is that it doesn't really matter if the stuff is not that fantastic because no one's gonna read it anyway. Maybe a few people will see if there's something interesting in there and they'll read it. It'll take a while before you actually start getting decent readership and recognition. So you learn—you learn from your audience, and your audience is very small initially. But you wanna put out stuff you're proud of.

Jon Chee - 00:25:58: Yeah. For sure. Absolutely.

Outro - 00:26:02: That's all for this episode of The Biotech Startups Podcast featuring Alex Telford. Join us next time for part four where Alex recounts his cold-turkey exit from consulting—no job lined up, just a conviction that the intersection of LLMs and biopharma was too important to explore from the sidelines—and how a period of writing and building prototypes in public nucleated the cofounder relationship, early investor interest, and the $8,600,000 seed round behind Convoke, built around a vision of an autonomous drug development operating system that treats every manual knowledge workflow as a problem waiting to be reengineered.

If you enjoy the show, subscribe, leave a review, or share it with a friend. Thanks for listening, and see you next time. The Biotech Startups Podcast is produced by Excedr. Don't want to miss an episode? Search for The Biotech Startups Podcast wherever you get your podcasts and click subscribe. Excedr provides research labs with equipment leases on founder-friendly terms to support paths to exceptional outcomes. To learn more, visit our website, www.excedr.com. On behalf of the team here at Excedr, thanks for listening.

The Biotech Startups Podcast provides general insights into the life science sector through the experiences of its guests. The use of information on this podcast or materials linked from the podcast is at the user's own risk. The views expressed by the participants are their own and are not the views of Excedr or sponsors. No reference to any product, service, or company in the podcast is an endorsement by Excedr or its guests.