AI Psychosis, Coordination Tax & the Limits of LLMs | Alex Telford (2/4)

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 growing up across Europe as an expat kid drawn to the biggest questions in science, studying biochemistry at Bristol and synthetic biology at UCL where the field's promises far outpaced its practical reality and the realization that pharma was the only place applied biology could become a career. If you missed it, check out part one. In part two, Alex talks about joining a boutique consulting firm mid-acquisition and spending nearly seven years doing substantive biopharma strategy work, development prioritization, regulatory strategy, and competitive intelligence, and the craft of delivering not just analysis, but an actual recommendation clients could push back on. He also begins tracing the throughline from his early fascination with NLP and word factors to the post-GPT moment that sparked what he calls AI psychosis and set everything that came after in motion.

Jon Chee - 00:01:39: And so you're wrapping up your time at your master's now. It sounds like you're just like, "Okay. I need a change of scenery. I need to have more direct impact." Talk about what was up next for you. What were you thinking? Where is your head at?

Alex Telford - 00:01:55: Well, at that point, I didn't really know what I wanted to do because I had assumed I always wanted to be a scientist. And then there was this—I remember there was this, like, fifth grade or whatever. We had this letter to a future self you had to write, and our teachers actually mailed it back to us 10 years later or whatever it was. So that letter was something like, "My dream is to be a microbiologist" or something. When I was 13, I wanted to be like a microbiologist professor.

And so I'd long had this idea that, oh, yeah. Obviously, I'm going to go into academia. Obviously, I'm going to be a scientist. I'm going to do research. But then I didn't like research at all in practice, so I was sort of at a bit of a loss where the practice of doing the thing I was interested in was not something I enjoyed, even though I was still interested in the thing. So I was looking for places where I could still do something science-adjacent and still engage with the science, but actually work on more applied problems.

The pharmaceutical industry is probably the only real place you can work if you're interested in applied biology. I mean, yeah, you can do, maybe actually work in crop science or something like that. Sure. And that might have been interesting as well. But I think medicines and pharmaceuticals are the ones where it's the biggest industry that takes in people who have an inclination to work in biology. So I was like, "Okay. I wanna work in the pharma industry." And I didn't know anyone in the pharma industry, but we did have a family friend who worked in life sciences consulting. And I ended up speaking to them. And they said, "Oh, yeah." And the way they describe the jobs: "Oh, you got to work with those different companies. You got to help them solve their problems. It can be many different problems. It's different from day to day. Sometimes you're working on commercialization. Sometimes you're working on R&D strategies. Sometimes you're working on medical affairs or document regulatory strategy, things like that." So, "Oh, yeah. Sounds pretty interesting." So I applied for consulting roles and ended up getting one back in Switzerland. So I moved back to do that.

Jon Chee - 00:03:37: I had that same moment where I was at the bench; I felt the exact same thing. It's like, I still wanna be involved. I was kind of at a loss. It was kind of like I had a mini identity crisis.

Alex Telford - 00:03:50: Well, I feel like that's so many people who go into science. You realize that you don't actually love it as much as you thought you might. It's kind of brutal, honestly.

Jon Chee - 00:04:07: It's brutal. You're just like, "Is this what it is on this side?"

Alex Telford - 00:04:07: It's funny because there's so many things that seem so obvious to me now about how careers work and how to structure your life and career and how you should go after opportunity. I had no idea about when I was in university. You're so naive.

Jon Chee - 00:04:19: It's hard for anyone to tell you, to describe it too. You have to just go try it for yourself. Honestly, that's what I've come to accept. Everything is just one big trial and error.

Alex Telford - 00:04:30: Maybe. I don't know. I mean, I feel like I kind of wonder what would have happened had I grown up in the age of LLMs. Would I have made the same decisions? I think very likely not, because growing up, I just didn't have access to good advisors on many of these topics. I didn't have access to people who knew how to navigate the UK university system or good advice on how to pick a degree, how to prepare for a life in academia, whether or not I even like academia, what kind of jobs you might have for someone who's interested in these topics. Consulting, I didn't even know was a thing. I literally didn't know consulting existed as a job until after I graduated university because no one had told me.

Jon Chee - 00:05:09: Yeah. I'm a pretty early adopter when it comes to technology broadly speaking. And maybe I was a little bit slow on the LLMs, AI-type stuff. I guess the reason why is because I saw during 2020-2021, things kinda went crazy. And I was like, I'm just gonna let this kind of evolve a little bit before I start engaging. And then just recently, I was like, "Oh, shit. This is so game-changing." And especially for young people, you can just get up to speed on virtually anything. And that probably—exactly what you said. It's like, "Oh, everything that this lab experience is being described, I'm just gonna avoid that. I'm just not gonna do that at all."

Alex Telford - 00:05:48: There's pros and cons. In the same way that because of social media and YouTube on your phone, no one really experiences boredom anymore because there's always an easy way to get out of it, I think no one really has to experience the feeling of being stuck on a problem for days because information's always accessible. And I think you do learn a lot from having to grind through being stuck and not just always having access to the answer. So maybe on balance, it was better not to grow up with LLMs available during education. Even if I made maybe suboptimal macro-level strategic decisions because I didn't have a good advisor, I made maybe more optimal micro-decisions about how I learn and that positioned me better to do well in the longer term.

Jon Chee - 00:06:31: I think about that too. It's kind of like that intangible. Let's just look at you unpack boredom. You're exactly right. I have not felt that boredom in a long, long time. But I'm thinking back when Excedr was started, when you're just left to your own thoughts and just having to grind it out—some really great ideas come from that. But the craziest part is I can't even remember the last time I was bored.

Alex Telford - 00:07:00: Can you? I mean, probably as a child, when your parent locks you in a room without access to the Nintendo and you can stare at the wall, I mean, that's probably the last clear memory of boredom. Certainly not since I've been living alone, apart from my parents rather.

Jon Chee - 00:07:16: That's wild. And exactly what you said too—when the answers can kind of just be available at your fingertips the whole time, I do really wonder what that's gonna translate into. Because for a lot of the kind of entrepreneurial lessons and insights that I've gotten, it was just a matter of just chopping wood and learning by getting my hands dirty rather than just being like, "How do you start a leasing company?" and then just let me go research all public information on how other leasing companies have done it. And what that has translated to is a very weird looking leasing company, but being weird and different is a moat.

Alex Telford - 00:07:58: Yeah. You got pushed towards these smoothed-out averages of how you might do something. And I think humans and any animal—evolution predisposes you to taking shortcuts and conserving energy. If there's a button that gives you an answer, there's a button that gives you food or dopamine, whatever, you'll press the button. You know, there's this experiment where people were tasked with sitting in a room, not doing anything, and they could press a button to shock themselves. And there's no reason to do that, but people will still press the button sometimes. The experience of just sitting in silence and being bored is pretty aversive to people. And I think just doing work is aversive in general, so people want to press a button and have the magic machine do the work for them.

I think one of the dangers of LLMs and agents is that you just have this magic machine that's like a slot machine that sometimes will do the work well and many times will not. And you just keep pressing the button until it does the work for you, and you never quite sit with the pain of having to figure it out and do it properly. You just press the button and hope it does something good enough, but you never find the really interesting niches if you're just pressing the "good enough" button constantly.

Jon Chee - 00:09:02: And, obviously, you're deep in the space. When it comes to these models and agents and LLMs, why don't they just get it right the first time? Why can't you just prompt the thing and then not have that slot machine effect?

Alex Telford - 00:09:18: Well, I think it's a very deep question that I don't know if I know the real answer to. I have some intuitions of why models struggle with that. Sometimes they can just do the answer. If you have a verifiable domain, you can break it down into maybe the data availability and then some components of the model's architecture and how they work. And maybe there's just the inherent difficulty of the task or the knowability of the task—epistemic uncertainty. I mean, those are the three things that contribute to it.

Sometimes you don't have data for how something should get done. The frontier model companies are training a lot on doing math right now and coding as well. Coding is a lot of rich data. The context is very rich. Whatever a program does is specified by the code, and you can run that code in an executable environment and then see if the code ran and did what you wanted it to do. So you have both this very rich contextual data in high volume, plus you have the ability to verify and build an evaluation to check if the model can write code that runs and does what you want it to. In domains where data is very accessible and it's very verifiable, we've seen a ton of progress. Models are doing fantastic in coding and frontier math. They're starting to encroach on other domains, but those are really the ones where we'd seen the most gains.

And then on the side of uncertainty, sometimes it's just unknowable what the right way to do a task is. How do you write a good piece of writing? If you're solving a math problem, there's one correct answer that can be verified. Or if you're writing code, it's verifiable. But in something like, "Does this piece of writing speak to your soul?"—how do you write an evaluation for that? And supply of really, truly great writing is very scarce. Even if there's a lot of writing available on the internet, it doesn't mean that writing is good. High-probability writing is not necessarily gonna be what people would consider great or aesthetic writing. Without the ability to verify what really, really excellent writing looks like—how do you encode taste?—then you're just going to get a model that generates very probabilistically likely writing, which may be smooth and maybe easy to read, but it may just be generic.

I think there's been a number of studies that have shown that the average human tends to prefer AI writing to expert-level writing. There's one study I've seen recently where people preferred poems written by GPT-5 or something to poems written by the masters. But if you're someone who has spent all the time reading and writing poetry and has engaged with the classics, then the AI-written work reads like slop. But if you haven't developed that taste, then the AI stuff is very easy to read. It's kind of smooth and fluid, and it doesn't really challenge you. So that's the aspect of unknowability.

And then there's things that are fundamentally unknowable. What's the function of a certain protein in the body? There's maybe multiple plausible hypotheses, but based on the available data, we don't really know what it does until we run experiments. We actually have to go and collect—actually run the experiments to resolve the epistemic uncertainty there because you just don't know.

The last one is just architecture. These very large hyperparameterized models encode huge amounts of information in sparse and distributed ways. Depending on how you prompt the model, you're going to enter into one particular region of the latent space or another. The models aren't coherent necessarily. Prompting is more like you're pushing it into a certain region of its state knowledge. Depending on how you prompt, you'll get very different and inconsistent answers. They're not like a human where humans are mostly coherent—they have a coherent identity, personality, and sets of views. The models will change. They're like this bunch of fractured entities that depending on how you prompt them, you can access one space or another. So they're very weird intelligences.

Jon Chee - 00:13:24: So trippy. And that was leading me to my next question. We talked about how you're in the wet lab and it's so complex; it's really hard to debug. These are evolving systems. You kind of got jaded by what synthetic biology promised. When you apply AI to that, it sounds like a mess. Obviously AI drug discovery and AI in life science is talked about a lot. Are we just way, way behind? Do we have a lot of work to do to actually have it be meaningfully beneficial?

Alex Telford - 00:14:02: So the technology is so general that you can apply it to probably any kind of task or workflow in some way. And I think there's some tasks that are good fits and some that are less good. In my company, we're trying to apply this knowledge to things that we think are good fits, like document analysis, document summarization, document writing, dataset generation. Can you find workflows within the industry where you do have to do things that are verifiable or easy for a human to go in and check? Can you pick use cases that naturally align with the strength of an LLM, like moving information from one format to another or synthesizing a document? There's a lot of procedural, rote administrative work in the life science industry, like creating a re-submission. That's a great fit for a language model or things like structuring data from an unstructured registry, like a clinical trials registry.

And then there's a broader class of AI and predictive modeling where AlphaFold and these protein folding models come in. They're not language models, but they use a similar transformer architecture. They're great tools for predicting protein folding and finding candidate binders for a target. And then there's bad applications, in my view, which are things like really pure basic discovery work. My bias as a life scientist is that you can come up with all sorts of hypotheses to explain some natural phenomena, but you have to test it. The smartest people in the world can come up with even more and more creative hypotheses that sound completely plausible and convincing, but they fall apart in contact with reality. I don't know how much intelligence really gives you the ability—I don't know if scaling intelligence means that you necessarily are gonna be able to derive these fantastic theories from first principles and don't have to do any experimentation.

In fact, I don't believe that's the case. I think you're always gonna have to do experimentation. We work in consulting and they have a lot of smart people. One thing they're very good at is constructing any kind of narrative to fit any available data. The language models are going to be very persuasive. They're gonna be able to construct these fantastically persuasive arguments fit on the underlying scientific data. But just because the argument is persuasive doesn't mean that it's any more likely to be true. So that's where I'm skeptical of the use cases of some of these deep learning-based models is when you're really asking them to step outside the domain and make a specific inference about what is the true state of the world when there are multiple potential true states. I don't think that's gonna be particularly fruitful. It can be useful in hypothesis generation—generate some ideas and go test them. But you're never gonna escape the experimental paradigm. You always have to go into the world and collect data.

Jon Chee - 00:16:53: Absolutely. And I couldn't agree more. It kinda gets back to what we were just talking about earlier—when younger folks might be relying on these models, what is gonna change in how we conduct ourselves in education or careers? Then you flip it and apply it bluntly to the scientific endeavor, what does that result in?

Alex Telford - 00:17:29: I wouldn't necessarily say that we were doing things correctly before. Given the constraints we had and the technology we had, it's unlikely we were doing things optimally. It's completely obvious that language models and AI generally are gonna become a fundamental part of how we work. We work with pharmaceutical companies and they're very enthusiastic about AI and LLMs, but the penetration of the technology into the workforce and their daily life is still very minimal relative to what it could do.

You're pretty naive if you're an educator or someone entering the workforce and thinking you can just continue to do things the way you've always been doing them. You have to change. Skills that were once valuable, like knowing a bunch of facts or being able to recite a particular section of an FDA regulatory document—that's probably not very valuable anymore. But being able to build systems and orchestrate systems and pick how a particular workflow should work are gonna become very valuable skills. Architecting is gonna become very valuable.

Software engineering is changing completely. Knowing the syntax of how to write a particular piece of code is a much less valuable skill than it was a few years ago. What's more valuable now is how do you properly operationalize many of these coding agents in parallel? How do you provide the right context? How do you properly spec out what you wanna build? I think work is gonna change from you being someone who does work directly to someone who orchestrates and builds and maintains systems. And then the other parts of valuable work will be going and collecting information from the world and putting it into whatever models you're using. Data that models can't get—how do you find it in the world, collect it, and encode it in a way where it can be operationalized? Those seem to be pretty durable skills.

Jon Chee - 00:20:04: Everything you said, I'm feeling it on my side even though we're not a software company. Did you game growing up?

Alex Telford - 00:20:11: I did. Yeah. Quite a lot.

Jon Chee - 00:20:12: Did you play StarCraft?

Alex Telford - 00:20:13: Yes.

Jon Chee - 00:20:14: Doesn't it feel kind of like StarCraft a little bit?

Alex Telford - 00:20:17: I think there's a sort of intuitive appeal to that analogy, but I actually don't think it's like StarCraft. There's an interim state where it feels a bit like StarCraft, which is kind of inefficient. You're maxing out your APM and shifting from tab to tab—kicking off one tab into codex, another into work tree, making a PowerPoint, and then back into our platform. You're just manipulating the computer to parallelize work, like checking how your miners are doing and scouting their base and microing some attack over here.

But I think that's a very temporary state. I think the actual better analogy is something like an idle game, like a Cookie Clicker or something, where you have this system that's running autonomously and you are an overseer. A lot of work is going into building systems that are automated, that run on a schedule for long durations and won't need human input as much. The phase of you tabbing around window to window is temporary; it's gonna become maybe more like a Civilization or a Cookie Clicker in how it feels. Especially as these things get better at computer use. Clicking around Excel and chat windows is gonna feel very antiquated because computer use is a constrained objective that is easy to run many parallel experiments for data.

Jon Chee - 00:22:03: It's funny because I'm starting to see that too, having our agents run on their own and I'm doing less. You bring up Civ, which is funny because my friend is basically vibe-coding his own ideal Civ in Claude right now. He doesn't know how to code at all, but he's just a Civ connoisseur.

Alex Telford - 00:22:40: And this is the thing where you don't actually have to read the code very closely anymore in many cases. Though you probably should have someone reading the code if you're a serious software company, because they actually do make mistakes quite often. The models are nowhere near perfect. They still have some flaws and sometimes do stupid stuff, so you can't just let them off on their own. But generally, you don't have to closely read the code to get some prototype working.

There's a broader point that it's probably gonna take a long time to retool how work is done. In the same way, factories had to be retooled after electrification. You didn't see gains from electricity for twenty years after electrification because you had to retool factories for assembly lines. You need to think about how do you build these autonomous factory lines for knowledge work that actually work without too much human oversight. Right now, I'm not that convinced that the efficiency gains from AI are that big. Maybe they're 20%, but we're still so far away from the promise, which is orders of magnitude productivity increase.

Jon Chee - 00:24:05: Interesting. We're talking about agents and all this stuff. You mentioned "harness," and this is actually maybe just my ignorance. What's a harness? I see people throwing around MCPs and stuff. What is a harness?

Alex Telford - 00:24:21: You can think about an AI application as having a few different components: the underlying foundation model, the data it has access to, and then there's the tools and the actual front-end application you interact with. A harness refers to the tools and the UI aspect of it. A helpful analogy is a playground. In a playground, you have a bunch of affordances—a slide tells you to slide, a swing tells you to swing. It's the environment that the agent operates in and the capabilities it has access to as a result. And then you have safeguards around it. So a playground is a constrained, safe, and observable environment.

It's the set of capabilities that a model's environment gives you the ability to do. Very few people are fine-tuning or training their own models right now because the pace of improvement in frontier models is so fast. Instead, you just drop a model into your environment and it gets all these new capabilities, as if it's putting on an Iron Man suit or something. That's what a harness is. The bet companies are taking is that there's durable value in creating these operating environments for these generic models and the data.

Jon Chee - 00:25:55: Got it. So what's an example of a harness out there, publicly spoken about?

Alex Telford - 00:25:59: Cursor is the one that most people know, the IDE. When you're working in Cursor, they've indexed your codebase. They have their own custom tools and workflows for how to search data, operate over data, and write code. Cursor is the main example of a harness that people use, and then there are things like Harvey and other domain-specific companies.

Jon Chee - 00:26:25: Got it. So it's kind of what stitches it all together and holds the foundation model?

Alex Telford - 00:26:31: Yeah. You have this fuzzy entity, which is a foundation model that can kind of do everything but doesn't necessarily do particular things very well or might go off the rails. Most people are not very good at using foundation models—they don't know how to use ChatGPT or an API to do work. You could use it like Google, but how you actually use a model to run my company for me—very few have any idea how to approach that or even something simple, like how to pay my taxes. You can't just go into ChatGPT and say, "How do I pay my taxes?" It'll give you an answer, but you can imagine a harness as being something where you drop in your tax documents and it automatically runs workflows to extract info, get additional information, read the latest tax regulations, populate the forms, and file them. You could use the ChatGPT API and string it together with some coding tool to build a tax-filing agent for you, or you could just use a harness someone else built that abstracts that complexity.

Jon Chee - 00:28:01: That makes a lot more sense to me now. People were just broing around with "harness," and I was like, what the fuck are they talking about?

Alex Telford - 00:28:08: Yeah. These terms get thrown around so much they lose meaning. What is an agent? What is reasoning? People say, "Oh, the models aren't reasoning." They are reasoning in some definitions, and in others, they are arguably not. This is what happens when something becomes a massive bubble. The terms that had specific meanings become abused and lose their meanings.

Jon Chee - 00:29:03: Exactly. I'm getting lost in the sauce on Twitter. It's used interchangeably all over the place. To get back to you—you wrapped up grad school and went to try your hand at consulting. Talk a little bit about the early days of being at CRA. What was that like for you?

Alex Telford - 00:29:27: I joined a company called C1 Associates. I only applied to life science-specific consulting firms because I had heard that the big ones, like McKinsey or Deloitte, might staff you on oil and gas or finance. I wanted to work in Switzerland, and I didn't want to work on back-office finance optimization for UBS. I think in retrospect, that probably was not correct. If you applied to McKinsey or BCG with an interest in life science, those are actually really great places to do that kind of work. But anyway, I just wanted to do life sciences. C1 Associates was a very small boutique firm based in Lucerne, a tourist town in Switzerland. I liked the team there; it was somewhat scrappy and small—maybe eight people when I joined.

Two weeks into that job, we got acquired by a larger company. I came in to work and there was some strange American in the office. They put us on a conference call and said we're being acquired. One of the founding partners sounded very tearful, and I was thinking, "Oh, great. I'm gonna get laid off" because everything gets laid off in an acquisition. Obviously, that's not true in consulting because you buy a firm for the consultants. All the IT support staff got laid off, unfortunately, but the consultants stayed. I didn't know that.

I remember one of the senior managers took me aside and said, "They're gonna give you new contracts. The acquisition is only gonna go through if a certain number of contracts get signed. You don't have to sign, by the way." I was like, "I'm new here. I'm not gonna try and take down this corporate merger in my second week as an analyst." That's a lot of responsibility. It's hardest for an acquired firm because you have a strong identity that stays even as you get absorbed into a larger company. We were like a satellite office in Switzerland, and that identity stayed very strong up until I left almost seven years after I joined.

Jon Chee - 00:32:30: So you continued to get the small consulting firm experience, it sounds like. They let you do your thing?

Alex Telford - 00:32:34: Yeah. I stayed part of the Swiss office. Maybe the office grew, but we were maybe 20 people at most. I love autonomy. The partner in that office was very successful, so we had a lot of autonomy and were mostly working on our own projects. The corporate overlords would come down from time to time and give us HR training, but we could kinda do whatever we wanted. I had a great time. I really liked my colleagues and the partners I was working with. I was continuing to get promoted and do interesting work with increasing responsibility, so why would I leave? I had no issues. It was a pretty good experience overall.

Jon Chee - 00:33:52: That's pretty rad. I guess as you were gaining your first projects, are you allowed to talk about some of them?

Alex Telford - 00:34:00: Well, I can't tell you about the clients, but I can talk in generalities. A lot of the projects were development strategy. Given a series of options, which indication or disease should we develop this drug for? What kind of patients? What product profile is gonna be attractive to patients, physicians, and payers? We'd talk to physicians to understand prescribing behavior, patients to understand treatment challenges, and payers to understand reimbursement. We'd do a bunch of secondary analysis: what's the competitive landscape look like? How is the market evolving? How many patients are there?

The first one I did was a lung cancer project back when the PD-1s were relatively new. Questions like: what cutoff should we use? Should we go for an all-comers population or a PD-L1 1% or 50% population? Should we do a chemo combo or dual immunotherapy? One thing I learned was that people are generally looking for opinions, not necessarily just information. I like to be given the info and make my own decision, but a lot of other people want to be given a recommendation to react to. It's not helpful to just dump information on them. They want you to come in and say, "We think you should go for PD-L1 1% or higher TPS first-line lung because of XYZ reason." You wanna come in with an opinion.

Jon Chee - 00:36:39: I'm just thinking about this within the context of an LLM. That's gonna spoon-feed them exactly like that.

Alex Telford - 00:36:49: You have to resist that, I think. An effective consultant is a combination of guiding them to the right answer without making them feel like they're doing a ton of work. You have to be a thought partner. You can't just be the analyst who dumps info, but you also don't want to steamroll a point of view that might be incorrect. It's quite hard to thread that with LLMs right now. LLMs aren't very well-calibrated on guiding people through the right amount of work; they'll either dump everything as a black box or make you do too much work.

A good consultant reduces decision friction. There is just so much decision friction in these big organizations, especially pharma. It's so complicated and there's so much uncertainty that you can always find a reason not to develop a drug. The easy thing to do is just sit on your hands and see what happens, but then nothing happens.

Jon Chee - 00:38:20: And what do you think that's a symptom of? Is it a symptom of being big? Do you have more to lose? Or is it cultural?

Alex Telford - 00:38:32: I think there's this inescapable attribute of large companies where there's just a lot of coordination tax because the network graph is very complex with many edges. I think it's particularly bad in pharma because the industry is so complicated. You have all these different siloed functions that need to collaborate: manufacturing, forecasting, sales, toxicologists, regulatory, and ClinDev. Everything is so complex and deep that no person can have a global view of a drug program. You only have this very thin window of understanding.

You can end up hyper-optimizing for your specific slice of the world. In ClinDev, you might not want a biomarker in the trial because it harms enrollment, but commercial folks want the biomarker for the companion diagnostic. Payer and pricing people want the biomarker to help get a premium price. You have to negotiate between all these parties who are optimizing for their specific slice. And then you have the problem of unknowability—you don't engineer drugs, you sort of discover them and uncover info over time. You can always find some reason, like a concern in one preclinical monkey study, to just kill a program.

Jon Chee - 00:40:40: And when you see this, what do you think it's gonna take to get rid of coordination tax?

Alex Telford - 00:40:47: I don't know if you can ever really get rid of it, but LLMs or AI as a technology could help. Rather than doing coordination one-to-one across an enterprise, you can coordinate many-to-one with some central entity that stores all your organization's knowledge. Most people in a large organization are effectively information-passers. Executives are the decision arbiters, but they have many people under them doing specific analyses that are being aggregated up the levels of hierarchy. At each step, some information is lost or hidden.

The functions of being an information aggregator and message-passer are things LLMs do very well. You could move those roles into other functions, flatten hierarchies across an organization, and have central nodes where people coordinate directly. If I'm working on a target and heard we worked on it in the past, I can ask the system instead of hunting down a guy who worked there thirty years ago who might not even be alive. I remember one company wanted to file old drugs for approval but realized they had lost the original clinical trial materials. They didn't know where the folders were, so they couldn't file the drug. If you had a system that indexed all your data, you wouldn't have to play political games to track people down.

Jon Chee - 00:43:02: This is really fascinating to me. We were talking to our outside counsel and she mentioned her parents were academics at Stanford and Oxford; they passed away recently and she still has all the papers they published sitting in a box. It leaves me wondering how much knowledge is just sitting in these boxes somewhere in the world. It's really hard to re-engineer a large organization. Do you need to recreate the org from scratch to do these kinds of things?

Outro - 00:44:24: That's all for this episode of The Biotech Startups Podcast featuring Alex Telford. Join us next time for part three where Alex breaks down why pharma organizations are so resistant to AI transformation, the coordination tax of siloed functions, and the tacit knowledge that evaporates when experts retire. We'll also discuss the near impossibility of getting people to encode their own workflows when most can't write a flowchart of their own job, and why he believes today's biopharma industry resembles the space industry before SpaceX, where every launch is still its own throwaway project with no compounding learning.

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.