Tomorrow's Medicine

Personal work is essentially focused on workflows. I built a search engine, search and discovery engine. I built fine tuning models. And if you are interested in building stuff with me, join one of the e-commerce, like Econel workshops on workflows.

I have a version for consumers where you learn to build your own AI agents. And I have a version for enterprise agents where we talk a little bit about scale and that vibe coding is not the end of the story. Now, when we talk AI, very often people talk about chatbots, co-pilot, apps, vibe coding. But that’s not the full story.

AI goes way beyond this. It goes beyond a normal chat conversation. We have a physical world. So, let’s talk about AI in the physical world.

Let’s talk molecules, materials, medicine. So that’s new frontier. And we can call this bio AI. Artificial intelligence applied to biological data, biological problems.

Let’s move out of the research labs into the real world. And let’s see what this means. And obviously the stakes here are high. And we know one of the typical problems to address is cancer.

Let’s talk about AI. And we have been working for decades on it. So what will change now when we have better technology to address this? So with that, I’m very delighted to be joined today by Cyriac Roeding.

He’s co-founder and CEO from early. He is, as you well very quickly figure out, a fellow German. He has- Yeah, I’m sorry about that. Not that German.

Don’t excuse yourself. I’m sorry for that already. No, but he is the CEO of a bio AI company developing methods where you essentially create little switches or robots inside from cancer cells. It’s super fascinating.

Now Cyriac is in his fourth company. He’s one of those amazing guys who go through Silicon Valley and create one venture after the other and created a lot of value. He raised by now more than 150 million from investors. Including A16Z, Khosla, Perspective and J&J.

He sits on scientific advisory boards, included its Nobel Laureate, Jim Allison and Moderna co-founder Bob Langer. Before early, Cyriac built Shopkick in Germany and sold it then for 250 million. And therefore he was, or still is, I think, an early investor in open AI. He is an LP at many different funds.

And he co-authored the Harvard Business Press book at 25. This is like, I mean, I don’t really know how to introduce you otherwise. Like in other words, then he’s kick ass. So here he is, Cyriac, so glad to have you here.

All right, it’s a pleasure to be here. And thanks for having me. I’ve heard a lot about this show from you. And so I’m excited to be here today.

So before we go into the typical story flow where we talk about you and who you are. But let’s first start with the big news. You closed Series B, congrats. That’s a big step.

Like, tell us a little bit about it. Oh, I wish you were right. We haven’t closed Series B. We just started it.

Oh, you’re close. We just started fundraising on it. We have raised 105 million for early so far. And now we have to raise more money because we’re going to take this whole thing actually into humans.

And for that, you need a little bit of money because clinical trials are not the cheapest thing on the planet. So that’s what we’re going to do next. And just one thing, the last company you mentioned was Shopkick that was in the US, actually. We made it easy for people to get rewarded just for walking into stores physically, into a real store you can touch, which is the topic of today.

You walk in, use your smartphone as a digital player on top of the real world, the non-digital world, the physical world. And then you get rewarded. You get rewarded just for being physically there in points. And the points turn into gift cards.

And that company was acquired for 250 million. Kleiner put 7 million in, got 78 million back out. And then I wanted to do something different. But everything I looked at, in hindsight, I literally just did what I wanted, looked at what I was excited about.

And I came to very different things. Like brain. You told me about this, right? You kind of like, you went around Silicon Valley essentially trying to figure out what’s the next cool thing.

That’s right. That’s right. It led me to brain machine interfaces, computer interfaces. I literally helped a guy sitting in a dilapidated building in Berkeley shooting microwaves at his own head, trying to decipher his own thoughts.

And this was in, yeah, it’s true, in 2016. And it was very, very early. I also looked at the other thing that is my passion and my love is robots, consumer robots. And on that intersection, it’s obvious you have the physical hardware that atom and the digital software, the bit and the atom meets the bit.

And then the third thing I looked at was biology meets AI. And that’s a really interesting world. Sorry? Now, in all three, like in all cases, also, if you compare this with a shopkick before, you always try to figure out how can I get AI workable or how can I get technology workable in real world, right?

That’s exactly it. I am fascinated with the intersection of the digital bit and the physical atom. Where those two meet, I think you get the real fundamental innovations because we live in the physical world. We live in no other world.

We live in the actual touchable world. And when software can start manipulating the actual… Then it becomes very, very interesting and potentially game changing in how we live. So I’m actually much more interested in that than I am in pure software alone.

And that’s, by the way, a great other analogy is what is more interesting, VR or AR? Virtual reality or augmented reality? Augmented, if you ask me, absolutely augmented. What’s more interesting to you?

Is it more interesting to sit with a cap here and you look like a weirdo and you look at your own world and you live in it and you have no interaction with the actual world? Particularly Apple glasses, right? Yeah, it’s really awkward. Or are you more interested in glasses that actually put an overlay on the physical world on top of your glasses and you can see information about what you’re looking at?

You can connect with people, get their backgrounds based on their LinkedIn profile when you talk to them at a conference. What if you can have, while you are working as a surgeon, get information displayed in front of your eye while you are actually operating on the patient to make sure you’re cutting all the tumor out? So that’s where it becomes very interesting. And that to me is the classic example of physical meets digital is more interesting than just digital alone.

Now, before… And I didn’t plan to dig into this, but if we say physical meets digital… Where do you see the big blocks of what’s happening today? And you know Silicon Valley pretty nicely.

And so obviously we will talk cancer for the rest of the show, but are there other blocks? You mentioned the glasses. What else is in your mind when you… Yeah, it’s a great topic.

So I think in a nutshell, to keep it really brief, I think the big… Massive innovations are going to happen. A, in robotics. And I think in 20 years, not having a robot living in your house is going to be as strange as not having a washing machine today.

Yeah. Two, glasses are the next big computing platform. This is the next platform. Yes, I totally agree.

I actually tried very… Like I was… This is the first version. Like, dear Meta, can you please fix those glasses?

They are terrible in the user interface. They’re still too clunky. But yeah, like, yes, I agree. I want…

I need a new pair of glasses. So if you know any good system, let me know. So, you know, whoever owns this platform is going to own the next big substantial platform. It is…

We started with the watch. And the watch is very interesting. It is very interesting for fitness and for health monitoring. But the glasses are the direct connection to what you’re thinking, what you’re doing.

And it is… It’s remarkable. And of course, the next thing… And they are always…

I mean, we… Like, some of people have like the… Like, whatever… Like, the necklace.

Don’t… Will not work so much. Glasses are actually… They’re always pointed in the direction you are actually looking at.

That’s right. And a lot of glasses… A lot of people wear glasses anyway. And…

But even those people who don’t wear glasses all the time, they usually have sunglasses. So it’s not… And it’s not an unusual use case even for people who don’t wear glasses every day. So then the next step after the glasses, what could that be?

Well, that’s obvious. It’s the combination with your brain. Yes. So that’s brain-computer interfaces, BCI.

And BCI is a rapidly evolving field, not just Neuralink, which has made some very impressive progress, but also Science, Science Inc., which I’m invested in. It’s a very exciting company that actually puts an overlay here into the iris and allows people who can otherwise not see to suddenly see again. And it’s a very exciting company that actually puts an overlay here into the iris and allows people who can otherwise not see to suddenly see again. By connecting directly to the nerves.

So it’s a very, very exciting field that will help people initially who are in some form limited physically. But in the future, it will connect to everybody who wants it. And basically gives you an… Imagine if you could surf the web.

And instead of always finding the next thing or giving a prompt… What if you could just think? And the next thing shows up immediately that is interesting to you. Just by thinking about it.

So that’s that. We talked about this novel interface so many times. We had MIT lab trying to do this with the hand movements. All too complicated.

Also, the AI tool where you then double tap on your hands. Doesn’t really make sense. Like to have a brain connection will be definitely the next helpful interface. Right.

And some people will say that’s a disaster. I never want that. And of course, I understand that. Look, there are two…

I mean, there are two different ways of thinking about this. One is the invasive way, which is the world that we’re going to live in for the next 20 years. Because it is very difficult otherwise. But then there will be probably some breakthroughs that allows us to be non-invasive.

And actually just re… And then of course, if you want to go even further out, if you can read, but you cannot write, that’s a big limitation. What if one day you could actually write thoughts into our brains? So now you’re talking about Big Brother and what are we supposed to think and forced to think.

And there are all kinds of problems that come with that. And I want to be also clear about that. Also, when we talk about biology, this is all… It’s exciting and amazing, but it has to be handled with great care.

Totally. And the reason I’m saying that is it has always been like that. Whenever we, as humanity, develop new ideas, new products, new technologies, it was always both a good thing and a dangerous thing. Yes.

For example, in the past, you had horses on the street and then cars came. And one day, the first person died. Yes. No, but you have to develop the safety standards around it.

But the difference is with every new generation of technology that we’re developing, the stakes get higher. Totally. Because the scale gets higher. The scale gets higher.

Now, before we dig into this, because I obviously want to talk about it early. You have a nice swing in the background. We should cover this. But if the media team could just play again the question thing in.

Because, as usual, you can ask questions. You can ask questions about scale. You can ask questions about whatever you want to ask questions about. We will see it.

We will bring it in here. Let’s dig into the topic of today, which is essentially the whole idea of bio-AI and AI interfaces to the human. Tell us a little bit about early. And tell us about how you met this initial idea.

Because that’s a super fascinating story. All right. So first, the one big area that you asked, you asked about big areas, physical AI. One of the most fundamental ones is the living atom meeting the digital bit.

And that is your body. And that’s where we’re going. That’s where we are in biology. And that is precisely why I’m doing what I’m doing.

Because I think one of the most foundational changes that we’re going to see in the next 20, 30 years is the combination of the living atom with the digital bit. So here’s what happened. I was looking for my next thing after my last company got acquired and I was looking at over 200 ideas. And when I, whenever I look at a new idea, my next company is going to be the digital bit.

And I’m going to be looking at over 200 ideas. And whenever I look at a new idea, my next company is going to be the digital bit. And when I look at a new idea, my first filter is to try to forget it as fast as I can. And most ideas don’t pass that filter.

They’re forgotten very quickly. And only those ideas that I simply can’t forget without, despite trying, might be interesting. So in general, the underlying theory is that I generally believe there’s an ocean of good ideas, but there is a dearth of outstanding ones. And only outstanding ones are worth pursuing at all.

By the way, not even every outstanding idea is worth pursuing. So you have to really have a very tight filter. So then this is what happened. I looked at all these different things.

I already mentioned brain, brain computer interfaces. I mentioned consumer robots. And then I was really excited about biology meeting software. And I only live four miles from Stanford in a little town called Potola Valley, 4,000 people.

And so I rolled down the hill and went to Stanford. And I was like, oh, I’m going to be a robot. And I stumbled around the campus. And I met really interesting people.

And, you know, it’s interesting. I don’t know how many people I met, probably over 100. And each person was telling me they were working on the world’s most important idea. And by definition, everyone except one or everyone except everybody was wrong about that, of course.

But the problem was I’m not an expert in biology. So I couldn’t. I couldn’t differentiate between a good and an outstanding idea. I could perhaps say whether something is good or not very good, but not whether it’s outstanding.

For that, you really need to know your stuff. And so it became really frustrating. And so, like, after the first four months of not doing anything, we’re really exciting. Went to China and packed up our little micro children.

And my wife and I went there and we lived in Beijing. For three weeks and met entrepreneurs from tiny idea to big idea. Went back, went to Europe, did the same thing. Wonderful.

After six months, I became very, very antsy. I wanted to do the next thing. But I couldn’t find it. And so now, mind you, 11 months in, I was really depressed.

It was like now Thanksgiving. And Thanksgiving is basically the end of the year in the U.S. And I was sitting there thinking, I still don’t have it. I can’t believe it.

The year has passed. And on that day, the door opens and my wife comes in, Angel. And we lived in somebody else’s house as a sub lessee at the time because our own house was being remodeled. So we got other people’s magazines that were not meant for us.

And she comes in and she gives me this magazine from Stanford Medical School. It says, she says, I think you have to read this story. So I look at it and it’s called, and yet you try. You should Google it.

It’s a story in Stanford Medical Magazine. And in the story, it explains how one of the world’s most preeminent early cancer detection and treatment experts, Dr. Sam Gambhir at Stanford University, lost his only child to cancer at age 16. Yes.

Millen. And how he was searching for new ways to detect and treat cancer early, which is obviously viscerally important to him. So I read this story, Lutz, and I was shaken by it. Can you, like, and it is in your name, early.

Like you said, like he was researching how to detect it early. Yes. Many of, like many folks who are not from the medical field might kind of think, well, what we are looking for is a treatment. But early is about detection early.

And the guy, like the. Professor was searching for ways to detect it early. Why? That’s a slightly shift of notion here.

Why is that? Well, the biggest lever is to treat your cancer early. If you, I give you an example. Lung cancer.

Lung cancer is the number one killer of all cancers. It kills twice as many people as the second biggest one, which is colorectal. And it’s not just smokers, by the way. So lung cancer, if you detect it at stage one, you have over 60% survival rate after five years.

If you detect it at stage four, you have an 8% chance of living in five years from now, 92% will die. Yes. So it is absolutely critical to treat it early. You can only treat it early.

If you diagnose it. Okay. So you can only treat it early. But to be clear, early started as a diagnostics company.

It’s now a therapeutics company. And we can talk about this in a bit. Do you want me to explain how this thing came about? Yes.

So you actually met this professor, right? So I decided in that moment on this Thanksgiving day to send a cold email. So I sent a note to Sam and I said. It is not a coincidence.

I’m sending you this email on Thanksgiving day. I can only imagine how hard this day is for you. I don’t even know your son, but I felt a gripping sense of grief when I read your story. But I was also very inspired by the work you’re doing.

I’m not a biologist. I’m not a scientist. I’m a serial entrepreneur and I’m looking to build my next thing. I would like to meet you.

It took. Two months to find a spot on his calendar. And on a Saturday morning. We met for breakfast at a small restaurant in Portola Valley.

And the first thing we found out is. He also lived in Portola Valley. Oh, fun. There are only 4000 people and we were neighbors and didn’t know that.

It takes three three minutes by car from one house to the next. So my first question to Sam and this is might be relevant for the audience because. I am not a biologist. I’m not a scientist.

What the heck am I doing here? And I said to him. I need to ask you something and I’m not fishing for compliments. I want a brutally honest answer.

Should someone with my background or better lack there off because I’m an engineer. Not a biologist bother the world of biology with my presence. And he looked at me and he said. I have two answers for you.

The first answer is the world of biology needs more people like you not fewer people like you because we have a lot of experts and very few generalists. We have even fewer people who know how to build a startup. And the second answer is the world of biology will always find a way to screw you over. It will take longer than you want and it will be harder than you want.

So I took the second part of the answer as the cost of doing business didn’t like it but went with it. And before I knew it every other Saturday morning. I was like. Oh, it was a Saturday morning.

I found myself sitting at his kitchen table at his own house and he was trying to teach me biology largely in vain. But to give him credit for trying it. Then he started introducing me to a bunch of awesome scientists at Stanford and clinicians. The problem was I still couldn’t find the right combination of an outstanding idea.

And an outstanding scientist or team made or possible co-founder because each one is one in 10,000. the math. Yes. You multiply them.

So three months later, I went back to Sam and I said, Sam, I’m so sorry. I still don’t have it. May I ask you a simple question? You’ve seen hundreds of ideas in your own lab, outside of your own lab, which one has the highest potential of all of them?

And he looked at me in his scientist underselling type of way. And he said, there is this one idea that I believe has some real potential. And that became early. Now let’s explain, by the way, we have again, like audience, please ask questions.

I have already six of you, like six questions here from the audience, which is amazing. But looking at the questions, folks understand very much what early is doing. At least the ones who asked, thanks so much for it. Now give us first this beer.

What is early? What are you guys doing? All right. I’m going to keep this brief.

So very high level. The number one problem with cancer drugs is what? It’s systemic toxicity. You kill like cancer, like cancer cells, which the buddy didn’t identify as bad cells.

And you need to kill cells. And you just need to kill cells. With radio, you kill not only the cancer cells, you kill everything around it. With chemo, you’re not only killing everything around it, you just kill everything.

So yes, if you don’t have a biomarker, like a CD20 marker or something to actually attack it directly on the cell, you are screwed. That’s exactly it. So the top 10 cancer drugs produce 87 billion in revenue per year. And these top 10 cancer drugs, which are very potent drugs, give you on average 12 months more to live.

And they give you 15 percentage points more survival rate, which leads you to 49% in total, which means every second person still dies. So why is that the case? The main problem is when you dose someone, the drug doesn’t only work on the tumor. It also works on healthy cells.

That’s by the way, why in chemotherapy, your hair falls out. The hair doesn’t have cancer. The chemo drug is focused on fast dividing cells and your cancer divides fast, but so do your hair cells, of course, because your hair is growing. That’s why your hair is falling out.

So in other words, extraordinarily imprecise. And by the way, this is why one of the first questions we got here in the chat was from a rub up. Thanks for asking, by the way, how do you, how do you avoid that this is leaking into benign tissue, right? How do you exactly that?

He just asked the million dollar question. So I’m going to try to get there really quick. Okay. So let me just explain briefly.

So you understand the size of the problem. Now, if you give somebody a drug and it has the so-called on target, but off tumor effects, in other words, your target is not only on the tumor, it’s also on healthy cells. So you only have one shot, one chance of addressing that. What do you do?

You lower the dose. And when you lower the dose, you lower the efficacy of the drug. So you’re not giving the person as much as it should. And then the second double whammy comes in, the drug goes into your bloodstream.

You have five liters of blood in your body. By the time it reaches your tumor, there is this much left. So now I’m going to ask you a simple question. The question we asked our selves, what would be the ideal cancer drug?

How would that look like? That is, um, so far before I knew that I got to know you so far, it was understand whether the cancer cell has something you can dock on like CD, 20 marker, typical thing. So it has some structure, which identifies the cell as a cancer cell. And then you create a drug only focus.

On that genetic structure of the cancer cell. The only problem is most cancers don’t have that structure. That’s right. They either don’t have it or it mutates a little bit from patient to patient.

So it doesn’t work everywhere. And then it, unfortunately, not only on the tumor, it’s also elsewhere. And that creates the so-called on target of tumor toxicities, which lowers the dose. We’re back to the same problem.

So now let me ask you a simple question. What if we could make the drug. Only where it’s needed, meaning in the tumor, have it be produced by the tumor against the tumor in the tumor, by the tumor against the tumor. What would be possible if we could do that?

Well, it would identify the cancer cell, make the cancer cell destroy itself, right? I mean, that would be personalized cancer treatment. Yeah. So it would essentially unlock extremely potent therapies that would otherwise be too potent because they also go elsewhere and they’re too toxic.

Yes. So we’re going to try exactly that in a lung cancer phase one trial, starting next summer, dosing the first patients. So how does this work in order to make this possible? We had to create three major innovations.

We created the first genetic switches made out of DNA that only turn on in cancer cells. When you inject them into the body, they go everywhere, but they only flip on in a cancer cell like a light switch. And that required massive amounts of machine learning and AI. Otherwise we wouldn’t be here.

In fact, the interesting thing is we try to do it without AI and it failed. I’ll get to that in a moment. The second big thing is somehow you got to get this DNA thing to the tumor. How the heck are you going to get it there?

And that’s where lipid nanoparticles come in. You might’ve never heard of a lipid nanoparticle, but actually most likely you’ve had it in your body already with a COVID vaccine, because that’s how they became very well known and billions of people have received them. Here’s the problem with that. They are excellent cargo containers that take DNA or mRNA and they put it in a container and then they transport it through the body.

The problem is in 10 minutes, they are in your liver and they are out of your body. Why the liver? Because the liver is responsible for flushing everything out of your body that doesn’t belong there from alcohol to lipid nanoparticles. And so therefore, if you want to treat any cancer but liver cancer, you’re going to have to have a completely different lipid nanoparticle.

And for 15 years, this has been called the billion dollar barrier, because it’s excellent. It’s a great way to treat cancer. This is chemistry, not biology, by the way. Yes.

And I mean, these are two things, right? First, you need to reprogram the cell and the other one is you need to do it in the body at the right place. That’s right. So first you got to get the material there.

That’s the lipid nanoparticle. So you need like a cancer lipid nanoparticle that gets to the cancer and into the cancer. Once you’re there, you got to make sure you have a switch that only turns on in cancer because the lipid nanoparticles also going to go elsewhere, not just to your tumor. Totally.

Not just to your tumor. Now that switch has to be exceptionally accurate. So how do you do this? And then the third innovation is once you are actually in the cell and the switch turns on, you have to program the cell to make its own drug.

And since you’re only getting into 1% of the cancer cells with this transfection, with this, with these lipid nanoparticles, the big question is how in the world are we going to kill the other 99% that we don’t get into? And the answer to that question is think of a domino chain reaction. Dr. Salim which is identifiable.

That’s exactly right. So the cancer cell needs to produce a message, a message, think of it as a smoke signal that comes right out of the tumor and then have the immune system attack the cancer. But you have to be very careful that what you’re doing is staying only in the tumor so that you don’t have these off-target effects elsewhere in the body, these side effects that you don’t want. It has to be exceptionally local.

And all that brings us back to the switches that you asked about. So I have two questions. One is in order to make clear how do you do this? And the other one is actually one of the audience questions here is from K.

Carr. He said, OK, this is all cool, but how do you communicate this? Because that’s complicated, right? How do you educate and influence all the decision makers in the chain?

Because what we are discussing, is novel. It’s a novel way of doing therapeutics. So far, we could not reprogram the cell. So far, we could not get it down to the level of one.

So far, we could not have the cell producing their own drug. They all need a license. Every cell needs now an FDA approval license here. Like, think about it.

No, but joking. But still, tell us how. The short answer is very simple. We’re killing tumors.

And that’s, at the end of the day, the only thing that truly matters. Would you like to make the drug only where it’s needed and then kill the tumor, but keep it out of the body of the rest of the body? And the answer is obviously yes. And then the question comes, well, how are you doing this?

And then I can explain, well, it required three innovations. And they took years to accomplish. This is not something that you can, yeah, it’s easy to be said, but it’s exceptionally hard to do. So it took us years to do it.

But, you know, in the beginning, you find some people like Marc Andreessen, Jorge Conde at Andreessen Horowitz, Marc Benioff, who literally signed on to the concept itself because Sam Gambhir was brilliant and had shown at Stanford with very early mouse models that there might be an idea here. And they signed on to that and kudos to them for taking the leap of faith with a $20 million seat round on that. When we then went in the A round and we brought it to Khosla, they were betting on this at a time when we just had some early proof of concept in mice, but not a lot of data yet. And they led a $40 million round with Perceptive.

And then afterwards, we started going to town and we had to use tons of machine learning, 20,000 cancer samples, sift through them with machine learning because the original concept was to use a switch that nature provides. It’s called an endogenous promoter. It looked really good in cell lines that never change in the lab. But the moment we tried it on individual real cancer patient samples from the operating room, the whole thing failed completely.

We literally hit a brick wall. It was a lot of… A lot of brick. A lot of brick.

And how did… Because you have been an entrepreneur for so many times, you said it by yourself, you need three miracles. Most startups fail on one miracle and you actually have now three in a chain. How did you go about the brick and how did you make this feasible?

I hope nobody was aware it was three bricks. I didn’t even know that it was… I knew the major risks. That’s one of my main thing when I do, when I look at an idea.

I will never mine three main risks in order of priority, delivery risk, specificity, sensitivity, and that has never changed. So the analysis is accurate even so many years later. So now to your question, how do we solve it? We actually said, look, we could look at individual mutations like lung cancer has over a hundred of them.

The problem with that is once you do that, your construct gets bigger and bigger and bigger to check out all these mutations and you can’t get it into the nucleus anymore of the cancer cell, which is where the dysregulation happens that causes the disease. So you’re stuck in a, between a rock and a hard place. Either you’re accurate enough, but you can’t get into the nucleus or you can get into the nucleus, but you’re not accurate enough. So what are we going to do about it?

And then it took us a year of failed attempts until we asked the right question. As you know, oftentimes finding the solution, is mostly dependent on asking the right question. So we said, if we cannot interrogate every single mutation, what if we didn’t interrogate any of them? What if instead we looked at the downstream commonalities of these dysregulations, the things that are caused by them, not the original mutation and looked at the so-called hallmarks of cancer.

So that will be mine time. Once we had done that, we had a shorter list and we identified so-called master transcription factor binding sites that are dysregulated. We then assembled a combination of those, but then the real work began. And this is where it gets really tricky.

So let’s say you find the dysregulation of these transcription factors. You can now put them together, but you can space them, direct them differently. And it’s a three-dimensional system. So there are literally trillions of possible combinations, especially when you add a core promoter like we did and an amplifier and so on.

So you have such a large universe of genetic sequences, you can’t get through it. Now, I love this structure you just described. And by the way, we probably want to level it up a little bit more in terms of the technology description. But if I just point to the flow that we are having, very often people say, oh, generative AI, they will solve all our problems.

We will have those amazing new drugs. Listen to Cyriac here. Yes, he described very well how he uses strong machine learning power and large clusters. But as he said, the real work really starts once you have the understanding.

Because the machine in a bio AI in the interface is not just solving, it’s not solving for the application. It is giving opportunities to figure out this would be one way to address it. So there are many roads to go down and the machine helps us to go down those roads faster. And that’s the real value when we talk about AI.

Now, fast forward like you actually figured out a way and then you went ‑‑ because I have here one question from Chris about like how to apply this leukemia in children and so on ‑‑ and mine was like this is early days. Cyrik is not yet ‑‑ Cyrik ‑‑ ‑‑ Cyrik ‑‑ ‑‑ Cyrik ‑‑ ‑‑ Cyrik ‑‑ ‑‑ ‑‑ started with dogs, right? Yeah, we actually have a simple north start early. We want to be in patients as fast as we can.

So we’re taking our construct and we want to be in humans by summer of next year. Summer of next year. Amazing. So we actually want to run a clinical trial on lung cancer patients, a phase one trial, where we will intravenously inject the compound and we will try to show that this is safe and show some early efficacy data or indicators that point to that this can actually cause effect on the tumor as we are intending it to do.

Because the goal of this thing is not to be a theoretical research project in a lab forever. The goal is to save people’s lives as fast as we can. And I should probably point out that we’re not going to be doing that. We’re not going to be doing that.

We’re not going to that. Sam Gambhir that we spoke about died from cancer two years into the company. He had a cancer of unknown primary, which is even worse because you can’t even detect it early. By definition, it was stage four and it was very sad.

And my other co-founder, David Sui, who’s a brilliant gene therapy expert, and I have built this together with Badri Anantharania who’s our SVP platform into the company it is now based on the original vision, but in a completely different implementation. We haven’t even gotten to LLMs yet, which are also part of our game now because the wet lab, this is something I should probably say, because I think it’s really important for this topic of physical AI. Physical AI is only interesting when you train it on physical world data. Yes.

By the way, this is a general problem for the whole space. I get asked very often what is at the moment the best startup idea. And obviously, I don’t know because I’m not a wizard who can look into the future. What I do know is that all models need real data and therefore physical data.

And like the physical world models, Yann LeCun and so on talks much about it. They need data and they need real world data. That’s right. They need.

They need real world data. So what you in biology need to do to get real world data is to have a wet lab with pipettes. Yes. And a vivarium with mice.

And you need to get all this data generated in actual wet lab situations and generate high quality data because otherwise you’ve got garbage in, garbage out. Everybody knows that. So how do you generate that? That is the problem.

Because normally when you do a sequencing or a test on a sequence, you can do like if you’re really, really intensely working 50 a week. Yes. Well, unfortunately, we have millions to test. So what are we going to do?

Be there for 100 years? So we hired an expert from Austin University of Texas. She came in and she started building a so-called massively parallel reporter essay, which allowed us. To look not at 50 a week, but to look at 250,000 per batch by genetically barcoding them like a product in the store that you get checked out at the register.

Each one has a barcode, except it’s not a black and white code. It’s a genetic sequence. It’s a DNA sequence. So we throw them all in the same pot and the ones that are winners that are total outliers that are way better than others.

We then read them out based on the DNA barcode and that data, the best ones get fed into an LL. Why an LLM? Well, because it’s all structure. Let me guess.

What is, what is DNA? It’s text. The sequence. So it’s a sequence.

It’s a sequence of text. It, and this LL alphabet of DNA has four letters and it, they keep mutating permutating over and over again. So it’s a, there’s a perfect problem opportunity to be solved by an LLM. If it’s trained on the right data.

So we started with this, like, we need, we need to make this clear that people don’t get confused when serious means LLM. He means it like a technical structure, which is a transformer architecture, which is essentially trained on sequence pieces. The LLM you guys have from Anthropic, Claude may I, whatever tool you use in a chatbot is trained on human data. So it’s life is like a box off next word is chocolate because it’s trained on it.

If you. Look at a sequence of, um, uh, gene sequencing and you are trying to understand what’s the next possible mutation, right? You have the same thing. Like, like it’s like, it’s like a box of chocolate.

Or if you look at fraud data and saying like, these are transactions, what is the next transaction to be? We use large language models as well. Like we use transformer models for those types of problems. That’s a great way of explaining it loads.

Essentially. We’re. Predicting the next possible sequence for genetic switch. Yes.

And here’s the interesting thing. All the models out there, they are okayish for DNA. There’s one that’s specifically trained on DNA, which is Evo 2 by the Arc Institute. And that’s actually not one of the models that everybody uses as a chatbot today.

It’s the best one that’s out there right now, but still 99% will be useless stuff that comes out. So how do you predict what is probably not useless stuff? And for that, we build a proprietary Oracle, a prediction model that is much better than market standard to predict what comes out. And that what comes out, we then use bioinformatics to decluster it, to make sure it’s not so concentrated that they’re all pretty much very similar.

We spread it out. Diversity gets bigger. And then we feed that data into our wet lab. The wet lab that produces the outliers, the outliers go back into the LLM.

And so with. Every spin of the wheel, this thing gets smarter. It’s a learning loop. And so the last thing I want to let me, let me put like so much amazing stuff is happening here.

I want to underline this because this is very, very cool. Many folks saying, okay. And may I launch just now yet another model. What is out there for us as an entrepreneur or will like a Google and may I take over the world?

And you just heard the very clear explanation of why this is not so easily feasible. Because Cyriac uses the transformer, the one which is the basis for may I as well. However, he uses his personal data, not his personal data, but the data from his lab data from the wet lab data, which nobody else has to train a fine tuned model that can answer information that nobody else has. And then he does a feedback loop to improve it.

And we, have seen that so many times the future at the moment is. Is really on so-called fine tuned models that are basis that are trained on specific data. Nobody else has. And therefore there is a race on data.

There’s a race on fine tuned model. And the example Syria just gave is for healthcare and it’s very specific, but the idea, the underlying idea is happening in all industries across the globe. Yeah. You, you probably remember what alpha fold did for proteins.

Alpha fold made proteins programmable. So what do we do? We are essentially making DNA programmable. Totally.

Yes. And therefore you have the ability to do this feedback, which is stunning. Very, very cool. And the, uh, Oracle turns out is one of the most important elements.

Having a prediction model that is proprietary and better than what’s out there is really important because otherwise you have too much. Garbage that you can’t use. And then the wet lab is the one that qualifies it again. And then it’s the learning loop.

You need learning loops in order to accelerate the progress. So that’s what we do. Uh, and that’s essentially where we are now. We’ve tried this in, in thousands of mice to kill tumors in various mouse models.

We have shown safety in monkeys, which is exciting. That just happened the last, um, four months. Oh, wow. Now we can show that it’s safe and safe in monkeys.

Uh, it was so safe that the body temperature. Yeah. our clinical candidate, which means the product we want to do. And now we want to take it through the FDA and manufacturing process.

This is already started. We have a very strong CMC manufacturing team on board, and now we’re taking this to the clinic. Now, because we are like an AI engaged audience here, we have certain questions which are interesting in a healthcare setup on here. So Veena, thanks Veena for your question.

She actually asked about liability. She said, okay, well, if you now do an AI design genetic switch and let’s say like, you know, you produce false positive, false negative, because it’s, you know, that’s in any system, the fact, how do you actually treat this in terms of liability and responsibility? And it’s interesting because it’s a question you have across all AI design products, but healthcare is completely different in a, in a certain way, which makes it more complicated or less complicated. It depends on.

Yeah. So, you know, healthcare is different in, good and in sometimes not so good ways. Uh, as I mentioned before, even the best cancer drugs, oftentimes, unfortunately, don’t work today. This is the reality we live in today.

I’ll give you an example. The, you, you know, to what checkpoint inhibitor drugs are. Yeah. Um, I do, but, uh, plan.

The checkpoint inhibitor drugs are the biggest innovation in cancer treatment in the last 20 years. They essentially use your immune system to attack the cancer. The way they do it, and the big insight came from Jim Allison. Jim Allison basically found out that you don’t just need to stimulate the immune system to do more.

You need to first take the foot off the brake of the immune system. That means the cancer is really good at telling your immune system actively not to intervene, not to attack the cancer. He found a way to take that foot off the brake of the immune system, and he got the Nobel Prize for it in 2018. All of these checkpoint inhibitor drugs make up over half of the top 10, over 40 billion, 50 billion a year in these top 10 cancer drugs.

Now, here’s the problem. Even the best drugs like that have, on average, a 35% response rate. 65% don’t respond to it. Yes.

Because, again, you’re going on such a meta-level end that you’re missing the molecular structure threat. Now, the biggest problem is that if you take the foot off the brake of your car, but then you don’t push the gas pedal, where does your car go? Not far. It stands still.

You also need to stimulate the immune system. But when you stimulate and you push the gas, you get these cytokine storms throughout your body. That’s the problem. That’s precisely what Early is doing.

We want to be the second piece of the missing puzzle. The second missing piece of the puzzle. Somebody takes the foot off the brake, but somebody needs to push the gas, and we do it locally on the other tumor. The reason I’m using this as an example is because our audience member asked us about liability.

The biggest problem is that a lot of drugs do not work today in general. Yes. these drugs work. So now you have to be safe.

And that’s where a lot of mouse tests come in, a lot of animal tests, monkey tests, all kinds of safety tests before you introduce a new drug to the market. That’s where the FDA comes in to make sure that you are taking the right approaches. And safety is first for any new drug. And every single drug gets tested this way.

Like, it’s a given that from everything which we deploy has false positives and false negatives in any industry, right? And what I learned at my time at Google Health by working with the authorities, actually, no area in our life is so well thought through like healthcare in terms of when do you start screening? How long do you start screening? What does it mean for population health setups?

We think about healthcare, in this respect, so much. It’s very strict. Very, very strict. And it needs to be for safety of patients.

Yes. So this is actually like, Peter had another question. He talked about what happens if you during those tests discover conditions that you haven’t expected. It’s the same thing, right?

You’re trying to kill cancer. If you’re realizing that there are side effects, then it becomes a very strict protocol, which makes absolutely one of the best AI regulations. So you have to have a very strict protocol. That’s right.

I mean, that’s why clinical trials are there. And I also need to be clear, our genetic switches are not all machine designed. We are constantly giving our own input and modifications to it. This is a combination of humans and software, humans and machines.

So the biological knowledge is the critical factor here. You don’t just come up with random stuff, because by that, honestly, if you try that, that’s called brute force. You’ll be here in 100 years from now. It won’t work.

Yes. Now, I have one other question because you are going now towards a clinical trial. You narrowed the cancer you’re trying to attack brutally from them. So many cancers you could have used to attack and you narrowed it down to one area which you’re now going to clinical trial with.

This is a little bit of question every executive in this country has because AI can do so many things. And they need to figure out how do I deploy. So if you think outside of healthcare for a second, as the executive Cyriac who you are in terms of how do you have focus AI resources, how did you decide on the actual trial structure where you wanted to go? What was for you the driver?

You really have to get your hands dirty and understand. At a very fundamental level where the main problem lies. And for that, you need to really dig into the business you’re in. The business is the wrong word.

In this case, it’s about patients’ lives. Where is the problem in the clinic? Where did the patients suffer? Where do the drugs not work?

Where are the problems that are limiting? Why are they limited? And the. Deeper you dig into that, the more insights you will generate where it fits.

And to be clear, it’s also important to say this. You don’t start with the tool. You start with the problem and then it will naturally emerge which tools are best to use. Like try to imagine you want to build a house.

You want to have a house. You want to have a roof over your head and you want to feel really comfy. And now you’re going to start planning the house. How, what do you want to have?

What’s the output now? You, and now naturally you’re going to find the right tools to build that thing. Right. And you don’t, you don’t start with a screwdriver and say, what could I build with that?

Exactly. You start, you start with, Hey, I really would like to have a house. That’s really beautiful. Let me design one.

Okay. So now I need some materials. I need some tools and now I’m going to get to work and I need some skills. I need really skilled people who know what they’re doing so that house doesn’t fall on me when it’s finished.

So all of these same principles apply. This. Is. All.

Very, very. Normal, rational, standard. Thinking. It’s just put your, you know, use, put your own brain to work and it will automatically emerge.

Which is, and I think like it’s, it’s actually very nice summary from common sense. It’s what I mean. You say this is slightly, but like if, you know, when I talk to many executives and I talk to many boards and the question is, how do we deal with AI? Now, if I just.

Tell them common sense, Cyriac says common sense, then it’s. Yeah. I don’t want to know. I don’t, I want to be clear.

I don’t want to come across like condescending in any form. No, It’s like, it is. Far from what I, what I, what I’m trying to say is let it naturally emerge from the, from the problem because you’re facing certain problems. And now all of a sudden you’ve got this really powerful tool that you didn’t have before in your toolbox.

Now you can apply that tool because it’s so good at so many things. That you weren’t able to solve before. Meaning the core part here is actually, and you said it very nicely, Cyriac. Understand the problem of a space.

So when I, I just recently, like as many as of you, like of many of the listeners may know that I publish at Forbes, right? So I wrote this article about what do boardroom members need to know about AI and essentially it reads like common sense, because what I say, the, the underlying. Focus doesn’t change. You focus on your business.

You focus on your patients. You focus on the problem at hand. What has changed is the tools and you need to understand the tools. Yeah.

It’s basically all of a sudden you’ve got a bazooka in your toolbox. Totally. You’ve got this. You still need to aim.

You still need to pull the trigger. You need to know exactly what to do with it, but there’s this really powerful, smart tool sitting there that you didn’t have before. And that’s the change. I could spend.

More hours with you and I should actually do, we should have another coffee at one point, but, um, Cyriac, thank you so much for being here. Like to all the people who, those questions that I did not ask, I’m very sorry for it, but we got about 50% of the questions done. Um, Cyriac, it was a pleasure to have you here. Thank you so much.

It was really fun. Thank you so much Lutz and hope everybody got a little bit out of it. Appreciate it. Thank you.

Thank you. Hopefully that will feed off some questions that will later will feed off