How Tempus Is Using AI To Transform Diagnostics With Eric Lefkofsky

I'm good. I'm good. Thanks for having me.

In my case, I'd been in tech for a long time. I got into technology in 1999 and had built a series of technology companies, all kind of logistics, moving trucks around, and then media buying. These spaces were all very antiquated and had not adopted technology. We had developed some expertise at building these hybrid systems that were half human, half automated, and made these marketplaces more efficient. Those companies grew and went public or got sold. When Groupon came around, it was a similar problem: we were trying to apply technology, a different kind of technology, basically tipping point technology, to local businesses and local problems. So I found myself pretty early on leveraging a lot of those same skills that I had developed over a long time. From that point on, I never thought I would get into healthcare. But about 10 years ago, my wife was diagnosed with breast cancer. I found myself spending a lot of time in a clinic while I was CEO of Groupon, and I was just amazed at how little technology was a part of her care and how antiquated what I thought was a very sophisticated process was from a technology, data, and AI perspective. I think I just had this aha moment that my entire career and all these skills I had learned at applying technology to spaces that historically had not had technology could be applied here. So I started immersing myself in oncology and have been immersed ever since.

The company was founded in August. This August will be 10 years, so August of 2015. Pretty early on, we started off trying to apply technology to a space that didn't have technology. It was actually a much narrower focus than that. It was really trying to figure out how to make next-generation sequencing reports intelligent. When my wife got her sequencing report back, it didn't know if she was male or female, didn't know anything about her, what drugs she had taken, whether she was eligible for trials. It was a remarkably non-personalized report for what at the time was the state of the art of personalized medicine. I thought, this is crazy. Why don't they know that she's a woman? Why don't they know what drug she took? Why don't they know she's not eligible for that trial? They just didn't know those things. So we set out to connect those dots. At the time, having been in technology and starting Tempus as a tech company, I didn't think we would be sequencing patients. I assumed we would be building the intelligence layer that sat on top of other people's sequencing. I went to those folks who did the sequencing and asked them to give us the molecular data so we could contextualize and make those reports intelligent. They refused. They refused to give it to us, refused to give it back to the hospitals or doctors that were ordering these tests. They had it locked. That opened up a point for us to enter the market. We began sequencing, and we started in oncology. We still have our biggest roots in oncology, but probably five or six years into Tempus, we realized that the platform we had built that makes these diagnostic tests intelligent and contextualizes them was equally applicable not just in cancer but in cardiology, neurology, immunology, any disease where there's a lot of different data.

Everything. Even today, I get sent articles and people talk about how technologically sophisticated this space is because these are very technically sophisticated people, but they are technically sophisticated in a completely different discipline than the one at hand. The challenge is it's very hard to be a craftsman who can make cars with your hands and then suddenly become a software engineer. You need bioinformaticians, translational researchers, folks with PhDs and MDs to apply cutting-edge AI tools and techniques, especially those being used today for large language model building. I think the first step is appreciating that and being like, okay, I'm an expert here, I'm not an expert there. How do I find people that are and build those bridges? That's what Tempus does for a living. I always say to people, swim in your own lane. I don't want software engineers pretending to be PhDs or MDs and vice versa. One thing that's magical about Tempus, now that we close this acquisition, we'll have over 3,000 people working together to solve these problems. You need the expertise of people with a deep science background, especially biology and chemistry, but you need the expertise of people that understand how foundation models work and how they can be modified in an agentic environment.

If you had said to me 10 years ago, design an ecosystem that would impede progress at every turn, I would hand you the US healthcare system. This is almost masterminded. It's genius in its complexity to destroy progress at every turn. It's like a movie. I don't think this process fixes itself from within. It's set up not to do that. I think it's going to take private commercial enterprises that have a totally different incentive to solve this problem. Then you have the incentives in terms of biopharma, which are typically to generate a drug that maximizes its value during its patent life. The incentives of payers are complicated at best. If you look at the system today, it just isn't set up to adopt technology in big ways. Yet I'm quite sure the only solution to our problem, which is maybe among the biggest problems facing the US government, is technology and AI. You can't support a $5 trillion healthcare spend that's growing at 7 or 8 percent a year. It'll engulf everything. The only solution is technology and AI. I think it's going to take companies like Tempus to make everybody more efficient, including people that make drugs, people that treat patients, and people that are treated.

You start by saying, what am I trying to do here? I'm trying to make diagnostics intelligent. How do I make diagnostics intelligent? The diagnostic has to know who the patient is and then recontextualize itself around that patient. What diagnostics? Every diagnostic: blood test, CAT scan, MRI, mammography, genomic tests. I want every diagnostic input to be intelligent and contextualized. It's the only way to help physicians and patients get navigated to the right care because diagnostics sit at the center of every major decision. No one makes a decision without ordering a myriad of tests, interpreting that data, and then setting you down a path. We want that to be super intelligent. The problem with that is to make every diagnostic intelligent, you need to have all the data in one place and be able to generate insights in real time. That's no simple task. Our approach has always been to figure out how to build products that are sustainable as one individual puzzle piece in that bigger puzzle and have them build on each other so that the puzzle becomes clear at some point. We started with sequencing, saying, can we build a good business in sequencing that grows, delivers margin and cash flow? Then we said, if we can do that and that data is connected to clinical data, so we have rich molecular data connected to outcome and response, who would find value in that data? Biopharma. So how do we build products that biopharma wants that allows them to use that data? We built pipes to generate molecular data and clinical data flowing in real time. If I'm building these pipes, what other businesses can I layer on top? Can I help get patients routed to the right clinical trial? Can I help close care gaps? Can I say something algorithmically that I can get paid for? Each of these is a puzzle piece that ultimately, if you do it well, generates lots of revenue, gross profit, and free cash flow that can be contributed back into the whole. Probably if I look at Tempus today, the two biggest hurdles that we've gotten over that others have not is one, our broad connectivity to about 3,000 hospitals in the United States, and two, we're now generating enough gross profit and cash flow that we can do the math. We're generating maybe $800 million of gross profit. It's a tremendous amount of money. If you look at the amount of money we need to run the business, it's a minority of that gross profit. The majority of the gross profit dollars are being reinvested in product, engineering, science, and studies. We're at a point now where we're able to invest $500 million, $600 million, eventually a billion dollars back into growth in a sustainable way. If you look at the inflection point of truly great businesses, and I hope Tempus is one of them, that math continues. If you take a look at Google, they're investing not a billion dollars into new products, they're investing tens of billions, eventually $100 billion. If you want to know why search is so good, it's because Google probably invests twice every year what we invest in the entire NIH budget on curing humans, just in making search good. So it's really good, and people die. We need to create a model that allows that same kind of flywheel of investment.

We've always had a very large product engineering team and a team with deep AI expertise. If you look at our investments on an annual basis and the size of our investments since we've begun raising capital, this is where the majority of our money has gone. We just spent a lot of money on tech. It's hundreds of millions of dollars, some enormous amount of code. That said, we don't look to rebuild core underlying technologies that are perfectly good and operating in a commoditized manner, meaning multiple people have them and they're quite good. We didn't build our own optical character recognition system; we used one that was available. I think of some of these large language models in that same vein. There are now multiple models, whether it's Gemini, Llama, GPT-4, or whatever, that operate really well, low cost, high fidelity. So we use them. But the magic is in applying them toward multimodal healthcare data, which is radically different than the data these models were trained on. You can't just drop one of these models on a DICOM file, an electrocardiogram raw wave file, or a BAM file and say, what does this mean? It would be like, I don't know, there's three trillion A's, T's, C's, and G's in random order. It would build the largest four-letter sentence in the world. It doesn't know that that's actually counting base pairs to figure out if a gene is mutated. So we have to make that data accessible. There's no need to invest money in building an entirely bespoke foundation model. That said, there are certain data types where these models don't work at all; they just weren't designed for that. So we will likely have to build smaller models that are tuned to a specific data element for healthcare. But in general, we apply models from all the big players today, and they are just getting better by the month.

There's no reason why at some point, given the size of these models, the increasing scope of context windows, and the ability to run compute at low cost, I won't be able to draw associations between lots of molecular data and outcome response. That's literally lights on versus lights off. I think companies like Tempus will make that real very soon, in the next one, two, three years. That will be considered the first true moment of precision medicine. The fact that I found an EGFR mutation that gave targeted medicine is kind of cool, but so is chemotherapy targeted to you. We now need molecular data because that's how we're going to actually determine your individual course of therapy, which will not be the same as Brett's or mine even though the three of us may have the same exact disease phenotypically. That's a huge deal, and it will work equally well in oncology as it will in diabetes, cardiovascular disease, and neurological disorders. The second part of the aha moment will be when we have all these different data modalities being fed in—your pathology slides, your scans—those same models will likely be able to predict things before they happen. For example, you were riding your bike and a little piece of something went in your eye. You had a retinal scan six years ago. I'm looking at that retinal scan, your A1C, and three other things. I run some algorithm in the background and say, Charles, you're going to develop severe type 2 diabetes in three years. You'd say, whoa, okay, I'll stop eating donuts every morning.

I think the models get better and better. Hallucinations are already logarithmically down. I don't even know what percent they are now. But I think there will always be a human in the loop. I use the analogy of a plane. We have 777s with auto takeoff, autopilot, and auto land. These things have very automated systems, but I want two pilots up front. I'm not advocating that pilots can not be on the plane. I wouldn't advocate that a physician is not in the middle, no matter how sophisticated these systems get over the next decade or two. I very much want a doctor saying, yeah, makes sense, I'm going to hit go. Maybe one day that's unneeded, but nowhere in the near horizon. If I had a kid going to medical school today, I would not have any hesitation to say you have a great job. Radiology is an example where even five years ago people were saying, with image classification, what do we need radiologists for? Actually, radiologists were spending something like five seconds per slide. There was a lot of stuff that they would prefer to have a system that can cleave off those easy cases so they can devote their training to paying more attention to the things that matter more. A lot of times, medical professionals don't have time to read my entire chart or my mother-in-law's entire chart, which is incredibly complex. They can't actually process all that information in the time and monetization scheme that they're afforded when they meet with her. If they had a better mechanism to summarize and present that information, it would be a huge win.

It's the augmented human, 100%. When you start with a system that is somewhere between 30 and 50% inefficient, the improvements you get, you won't even feel them at the macro level. Here's a great example. I'm 55, so I grew up in a world with no smartphones and no GPS. You basically got lost everywhere you went. You were pulling over to a gas station asking someone where to go. Now, all the miles we drove in the wrong direction are gone. Yet it's really crowded and gas prices are really high. I think it's the same thing here. There's going to be massive efficiencies, but I wouldn't structurally worry about what all the doctors are going to do. There will be a shortage of nurses a decade from now, I guarantee it. Hopefully, what happens with technology is you get significant improvements every year so that the $5 trillion spend either stays at $5 trillion but can handle another 50 or 60 million people, or it goes to $6 trillion but we have a system that's way healthier.

The road map in oncology will likely set the stage for other areas. Oncology had to be first. Change is always hard. Before Tesla, if you said there's no money in making cars, what a horrible business. The biggest company in the world was worth $25 billion and had been around for 100 years. You'd have said the same thing about movies. Blockbuster wasn't that great, and here's Netflix worth 20 times what Blockbuster was ever worth. When you apply technology to a space, you can create a better product. We made better diagnostics, and that produced some growth. But long term, all these other things you layer on top create the sustainable advantage. In the case of Tesla, things they're doing in autonomous driving, recharging stations, robotics, all get layered on top of the core of building a better electric car. I think it's opened up the world's idea that you can have a trillion-dollar car company. Diagnostics are the same. They sit at the heart of all major decisions. Why would they just seed all that insight to somebody else? Why wouldn't they become the people that deliver that insight? The fact that some diagnostic players historically have been rooted in very low-cost, high-volume tests is the same as the fact that we had car companies that didn't become Tesla. Sometimes you need a new entrant to change the game. There's also a lot more complexity. Diagnostic tests are becoming more complex, harder to do, and therefore harder to commoditize. We've seen that in some of the MRD and early detection work. People said, don't open a lab, NGS is being commoditized, this is a race to zero. All that's happened in the past 10 years is the exact opposite. When we decided to buy Ambry, people said hereditary testing is a commodity, its best days are behind it. I said, how did you conclude that? Public investors might say, I bought Invitae stock and it didn't work out, but that doesn't define a space. Right now, some micro percentage of people get whole genome, whole exome, or large panel hereditary testing. If you and I could fast forward 20 years, it would be like the standard of care. I don't know why that wouldn't happen. We're going to ultimately understand from that molecular data, because of these large language models and large image models, everything that's going to happen to you. You're at high risk of developing AFib, or early onset dementia. You're going to want to know that and take action. When you get disease, that same data can be used over and over again. That's the best part of molecular data. For example, what drug are you taking? Do you metabolize that drug well? Maybe you need double the dose, maybe half the dose. All of that is going to be informed by your molecular data. If you don't have that data, you're like a vehicle with no battery, hoping someone else would give you a battery. That would be a tricky place to start.

We built agents that run models today that would have been cost prohibitive when we started. We started building agents pretty early. We've been at it for a few years. Some of the models we were running, we would run a query and it would cost us $2.7 million to run a very tiny query. We said, this is bad. Even we, who spend a lot of money on cloud and compute, can't absorb this. Now those same models that cost us $2 million to run cost $2,000 or $10,000, some tiny fraction. The cost reduction is that extreme. The vision we have is to generate real-time insights so everyone is on the right therapeutic path. To do that, I essentially have to run compute across all healthcare in real time every time there's a diagnostic insight. That's a layer we're not ready for. Tempus isn't ready, the US healthcare system isn't ready. But I think it will be at some point. You're talking about a scale that's really unimaginable, a different level of scale. We spend a lot of time thinking through with our cloud partners how that's going to work, because it is going to come. People like Google and Microsoft will have to help us figure that out.

It was a personal story about a friend of my daughter's who passed away. It was a great opportunity to recognize her. The challenge her parents had, and the challenge we all have if you've ever been in this situation, is managing complex care is miserable. If you have multiple doctors you are trying to get opinions from and weigh in on what to do, it is no simple task. Especially when you're dealing with a loved one who is sick. We built an app called Olivia to give patients access to those same tools. They can click a button and get not just some structured data from MyChart or Apple, but really all their data, literally all of it. If you want to move from one doctor to another, get a second opinion, or just have an archive, it's there. You have it stored in this little personal locker in your pocket. You can send that data anywhere you want because it's your data. You can ask it questions. As these models get better, the questions you'll be able to ask will keep getting smarter. We want patients armed. I've been saying this for a long time. I started Tempus because I saw a tsunami coming at the speed of an asteroid. This was really not a fun tsunami, and they needed to get ahead of it. I couldn't really articulate what it was 9 or 10 years ago, but now I really feel I can. That tsunami is that the average patient who today walks into their doctor completely uninformed, looking to put all their trust in somebody else with no understanding of what's going on, will in a short period of time walk in as if they just met with a hundred of the smartest people on the planet who spent weeks understanding their condition and have given them the 10 questions they need to ask. The doctor is going to think, oh no, no, no, don't do that. That's going to happen. At the end of the day, you're going to have really sophisticated tools that patients can use. We view information flow to everyone as nothing but a good thing, period, end of story. I want payers paying for the right drug, not the wrong drug. I want drug companies being as efficient as they can be, never having a failed phase two or phase three. I want patients to know everything about their care. I want good doctors to be great doctors and great doctors to be super doctors. You just want information flowing. The only people that should suffer in that world are the people that say, nah, forget this technology, it's garbage. If you look at the Fortune 50 or Fortune 100 CEOs back in the year 2000, if they thought the internet was a fad, the stock market cap of those companies stratified by the CEOs that said it's a fad versus those that said this is going to change everything, it's a perfect correlation. Those people that saw it coming got ahead of it. Those people that didn't, didn't. We want everyone in healthcare to get ahead of it.

Yes, it's going to take us. We were flooded with people that wanted to come in. We made a decision to give everybody a free trial. I think it costs like $12 a month or something. We wanted to give people a 14-day free code because we didn't want everyone to have to pay upfront.

I couldn't agree more. For us, Olivia is like chat 1.0. This is very early. It'll keep getting better. What I really believe is if you look at Google for a moment, Google has really two kinds of searches: regular search and incognito. In a world of healthcare, they actually need a third kind of search: a search that knows who you are. I want to be able to say things with that in mind so that you're not recommending something that conflicts with my meds. If you know I'm taking an antidepressant, don't send me to something else. I want you to render results that are not just semi-personalized but entirely personalized. The challenge is some of the big search engines and LLM companies may have a hard time with that. They're like, privacy, I don't want all that data, it's a lot of PHI, what do I do? I don't know how this is going to shake down. It may be that companies like Tempus end up building more impactful products because we want to have that relationship with the patient in a very identified way.

The biggest single thing will be that the error in the current system, the US healthcare system, is massive. I think roughly one out of every four decisions is entirely the wrong decision and sets a patient and a system down a really bad road. In roughly 10 years, I could see that being mostly gone, where errors are very infrequent. Certainly by 20 years. In a world where there are no errors and people never get the wrong drug in the wrong dose at the wrong time, you're going to have real improvements to longevity. You could easily see a three to five year extension of life. You're not going to see 10 or 20, but you'll see three to five. You feel three to five. One year felt like COVID was one year. Three to five years is a lot of lives because when they have problems, they're going to be on the right path. The next big one after that is being predictive to avoid problems before they occur. That's also coming. We started sequencing patients 20 years ago, and it's only now that maybe we're getting somewhere. It's probably the same thing with early detection and predictive tests. The early entrants typically aren't the solutions; they change over time. But I think 10 or 20 years from now, we will be very good at predicting lots of stuff and avoiding it.

Thank you.