1. Podcasts
  2. The Neuron: AI Explained
  3. Can AI Help Cure Genetic Disease?
The Neuron: AI Explained

8 Oct 2026 · 57m · 17 chapters

Episode artwork

Can AI Help Cure Genetic Disease?

In short

Whether AI can accelerate CRISPR-based genetic medicine toward durable, potentially one-time cures, and what data/clinical-trial bottlenecks still limit progress.

Guests

Dr. Trevor Martin, co-founder and CEO of Mammoth Biosciences. Mammoth was co-founded with Nobel Laureate Jennifer Doudna, Janice Chen, and Lucas Harrington. Mammoth builds next-generation in vivo CRISPR systems aimed at hard-to-reach tissues (muscle, brain) beyond liver/blood.

Key claims

With sustained effort, CRISPR could eliminate liver or blood genetic diseases within 5–10 years; AI can help design smaller Cas-like proteins and improve editing efficiency; the biggest AI gap is translating preclinical results to clinical outcomes due to missing public datasets linking both.

Notable examples

Sickle cell is the first approved CRISPR therapy (ex vivo editing). Mammoth’s lead concept includes cutting the APOC3 gene (high triglycerides) using in vivo lipid nanoparticles (IV infusion; DNA change persists after drug clears). AI approach includes searching metagenomic databases (~34B proteins) and using LLMs to generate novel Cas-like proteins.

Written by AI. May contain mistakes.

Chapters

0:00

The Promise of CRISPR and Genetic Medicine

Explore the potential of CRISPR technology to eliminate genetic diseases.

“And like, sure, it's only for a couple of years for now, but I think like there's line of sight there.”
1:40

Mammoth Biosciences and Genetic Medicine Goals

Discussion on Mammoth's mission to advance CRISPR for various diseases.

“We'll hear more from them here in just a little bit.”
4:04

Ex vivo vs. In vivo CRISPR Therapies

Differences between therapies that edit genes inside and outside the body.

“The alternative to that is in vivo, so in the body.”
6:50

Patient Experience of In vivo Therapy

What a patient can expect during an in vivo gene editing treatment.

“And if you wanted to test for it later, you would get vanishingly small amounts over time.”
8:55

Understanding CRISPR Technology

A deep dive into how CRISPR works and its applications in gene editing.

“So which version of CRISPR are we using today and how has that evolved over time?”
13:20

AI Integration in CRISPR Research

How AI is utilized to enhance CRISPR systems and protein searches.

“But whenever it's repaired, kind of in this like minorly incorrect fashion, that will interrupt the gene and cause it to not have any function anymore.”
14:00

Understanding CRISPR Systems

Learn how CRISPR systems are identified and utilized in genetic editing.

“And in addition, it allows you to then kind of build all these other techniques on top of it that we can get into.”
15:00

AI Evolution in Genetic Research

Explore the evolution of AI's role in advancing genetic research and CRISPR technology.

“Yeah, so it's definitely evolved over the years of the company.”
22:01

The Protein Focus of AI

Examine why AI in biology emphasizes proteins and the implications for drug development.

“So how do – why is it, for starters, that so much AI work really zeroes in on proteins?”
24:29

Challenges in Drug Development Data

Discuss the difficulties in linking preclinical and clinical data in drug development.

“That's after a massive winnowing of stuff that didn't work before that.”
Show all 17 chapters
28:00

Episode Discussion

“It's all about a good batting average, yeah.”
42:05

Transformative Potential of Permanent Treatments

Discussion on the implications of gene editing as a one-time treatment rather than ongoing medication.

“larger populations where maybe we don't have to take drugs for the rest of our life.”
44:30

AI and Recursive Self-Improvement in Biology

Exploring the concept of AI's recursive self-improvement and its potential application in biological research.

“And another question people often call this up with is like, oh, does that mean like pharma companies make less money?”
47:08

Challenges in Biological Data Collection

Evaluating the difficulties in integrating wet lab data into AI systems for better clinical outcomes.

“But yeah, I think that'll be more gated by the coding, I guess, escape velocity.”
49:51

The Role of Organoids in Drug Development

Discussing the utility and limitations of organoids in drug testing and the future of tissue engineering.

“And I think that creates really interesting problems of even just training loops because you're used to being able to verify answers and do RHL and stuff relatively quickly.”
53:07

Improving AI Models for Scientific Experimentation

Identifying desired capabilities in AI models to enhance their understanding of experimental failures.

“other approaches in terms of tissue growing, well, there's a great movie called The Island.”
56:16

Engagement and Newsletter Invitation

Discover ways to engage with the podcast and join the newsletter.

“We love this stuff and find it just so fascinating to listen to.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On The Neuron: AI Explained, what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00Well, in the CRISPR space, I think that, you know, if we all worked really hard and like poured effort into the space, like we could potentially live in a world that has no like liver or blood genetic disease and like definitely five to 10 years, which is a crazy statement. And like, sure, it's only for a couple of years for now, but I think like there's line of sight there. There's a lot of work to be done, both in terms of like turning through clinical trials and also some additional work for trickier targets about how you can address it. But it's something that, yeah, if we had a sustained effort, we could do.

0:28You can ask it to just generate wholesale new proteins that are Cas-like, but you're not necessarily basing it on mutations in a known Cas protein. I mean, I've always been a believer in you are what you eat, and I think that's true for models especially, and they just haven't eaten a lot of the key things they need to yet. I would like a model that could translate preclinical data to clinical. I think that's what everyone wants, right? So that would be like the simple kind of maybe trivial answer that doesn't exist, but it would be transformative.

0:57Corey: Welcome humans to the Neuron AI Explained. I'm Corey Knowles here with Grant Harvey. How's it going, man?

1:04Grant: Doing good. I'm excited to have this chat. How about you?

1:06Corey: Same, same. I think it's going to be a good one. Today, we're talking about one of the places where AI starts touching something much more fundamental than chatbots, and that's human biology.

1:18Grant: That's right. And our guest is Dr. Trevor Martin, co-founder and CEO of Mammoth Biosciences, which he co-founded with Nobel Laureate Jennifer Doudna, Janice Chen, and Lucas Harrington. Mammoth is building next-generation CRISPR systems with the goal of turning some diseases we managed for years into durable, potentially one-time treatments.

1:39Corey: Before we get started, I'd like to give a big shout-out to the sponsor of today's video, Mercore, a leader in AI data production, evaluation, and model training. We'll hear more from them here in just a little bit. Dr. Martin, welcome to The Neuron. It's great to have you. Yeah, thanks for having me on. I guess to get started, who's Mammoth? What are you looking to change? Yeah, so the company was started about eight years ago, spun out of Jennifer Daldin's lab, as you mentioned, really with the goal of delivering on what we see as the promise of genetic medicine. Maybe like any new technology, anytime something comes out, whether it's AI or CRISPR or maybe even semiconductors or the steam engine or whatever.

2:16You kind of always think everything's going to happen in the first couple of years. And I think inevitably people are somewhat disappointed. But for the best technologies and the ones that are truly transformative over the next five to 10 years, you really do deliver on that promise. And I think that that's where CRISPR is right in the middle of that cycle right now. Because it kind of came out, say, 15 years ago now, some of the seminal papers, I forget the exact dates. But in biology, that's not a long time. I know in AI, that's, you know, forever or whatever, like 15 days. But 15 years is pretty short in biology.

2:52Maybe that'll change. Hopefully. Yeah, hopefully. And that means that I think it's one of those things where, you know, it's really kind of slow, almost like a logarithmic phase, and then everything all at once. Where I think you're going to see there's now one approved therapy using CRISPR for a disease called sickle cell disease. And there's many more in the pipeline that in the next couple of years could get approval for different liver diseases. And I think you're going to see an explosion of these kind of one-time therapies for different diseases over the next five to ten years. And I think that is, at Mammoth, what we really want to be a part of.

3:28And in particular, there's certain classes of diseases that people found it very challenging to use CRISPR to go after. So to give you an example, the first approved therapy is for sickle cell disease. And as you might be familiar, sickle cell disease is a disease of the blood. So it's one of the rare diseases where you can actually take the cells outside of the body, do the CRISPR editing you need to do, and then put them back in. So obviously that's a very kind of complex procedure in terms of taking the cells out, but it actually simplifies the CRISPR part of it very substantially. The alternative to that is what we call, and that's called ex vivo editing, because it's ex the body.

4:03Outside the body, yeah. The alternative to that is in vivo, so in the body. and that's much more difficult because now you need to somehow get the CRISPR technology kind of to the cells you want it to go to but not to other cells. It needs to be delivered efficaciously enough that you're editing enough of them. And so there's some tissues that there's a relatively large number of companies working on diseases for including us like the liver. And that's because the liver you can think of it as kind of the waste fodder treatment plant for your body. So like pretty much anything you're ingesting is going to hit the liver on its way out.

4:37But there's others that are way more difficult, like the brain, famously the blood-brain barrier and things like that, or really kind of anything else, muscle. But when you think of genetic disease, unfortunately, you probably think of diseases that are kind of more relevant to those tissues, right? You probably think of like Alzheimer's or Parkinson's or ALS or like a muscular dystrophy or something. so a big part of why Mammoth was founded is that we wanted to develop novel CRISPR technologies that actually allowed us to reach into the muscle and into the brain in addition to the liver and the blood and not the other places and that was really a big part of our motivation

5:15Grant: that's really cool so what does a one time gene editing treatment actually look like I've always wondered this do I get an IV, do I get a shot is the procedure, you explained sickle cell but what would actually happen if I was getting a therapy like this for my muscles or my brain or my liver? Yeah, so it's very different for an ex vivo or an in vivo therapy. So we only work on the in vivo type. So it's quite simple from the patient's perspective. Like a typical liver-directed in vivo therapy, for example, would be a systemic IV. So you go in, kind of sit in like one of those big comfy chairs, I assume.

5:51You get an IV. You know, you're there for like hours. you have that technology then kind of in your body going to the cells it needs to go to. You don't really feel anything necessarily. Kind of infusion-like? Hmm?

6:08Corey: Infusion-like kind of? Yeah, yeah. Like almost like you're at a dialysis clinic or something. Yeah. But less time. Yeah. So that's like what you physically kind of experience. Maybe the other interesting thing to talk about would be the timescales that this happens on. So basically, the whole point of this infusion is that you're trying to potentially permanently change the DNA in your cells. So that happens over a matter of days, basically. You basically probably get 90 % plus of the editing you're going to get relatively quickly. And then over the course of weeks, all of the drug is leaving your system.

6:54And if you wanted to test for it later, you would get vanishingly small amounts over time. So what remains, though, is that kind of permanent change to your DNA. So that means that even though you're only having that drug in your system for a matter of weeks, and it's only really maybe doing the editing for a matter of days, there's some nuance to how it's delivered that I'm happy to go into. This is using a thing called lipid nanoparticles for the liver, which was also used for things like the COVID vaccine, but it's a very different kind of application. But it's similar technology. And basically, it's unique in terms of medicines because then for the rest of your life, though, you would never, in theory, if it works appropriately, have to go back to the doctor.

7:36So you're getting this like very transient kind of medicine, but it's leaving this kind of permanent change that is really potentially curing you so that it is fundamentally addressing the cause of the disease, not like a symptom or anything. Yeah, totally.

7:53Corey: So AI models are getting better fast, but how much better are they at doing actual professional work? That's what Mercore's AI Productivity Index, or APEX, is built to measure. Mercore is a leader in AI data production, evaluation, and model training, and APEX tracks how frontier models perform on long-horizon tasks across fields like corporate law, consulting, investment banking, accounting, medicine, coding, and more. The progress is pretty wild. In January 2026, frontier models could complete just 25 % of Mercore's agentic tasks. Today, that number is nearly 70%. And Apex updates with every major Frontier model release, so you can see how the newest models stack up, compare token costs, and explore results from third-party benchmarks like Browse Comp and Terminal Bench.

8:43Corey: See where the Frontier actually stands. Visit mercor.com slash Apex, that's M-E-R-C-O-R dot com slash A-P-E-X, and explore the rankings today. Now back to our interview. So which version of CRISPR are we using today and how has that evolved over time?

9:03Grant: Yeah, maybe you can explain CRISPR for people who don't know it. That's a good call. Yeah, so CRISPR, I don't know if anyone wants to guess. It's actually an acronym. So that's why I usually see it in all caps. But it stands for clustered.

9:16Corey: I'm not going to try to guess the acronym. No, you got it. Just go straight to that. Yeah, it would be unlikely to guess it. It stands for Clustered Regulatory Interspace Short Palindromic Repeats. So the reason why that is the name for it, which is a little bit hard to say, is that the CRISPR technology comes from what we call the adaptive immune system of bacteria and other microorganisms. So similar to how we all have immune systems to protect against viruses, for example, bacteria have immune systems essentially to protect against viruses as well. And those are a little bit different, though, and in some ways more simple and in other ways, like, you know, of course, very complex.

9:59And the basic idea, though, is that when a bacteria has, like, seen a virus before, it stores a little snippet of the virus sequence into its genome in these, like, clustered interspaced sequences. And then it's actually kind of constantly expressing these little chunks of the sequence and combining them with this thing called a Cas protein. So actually the protein, it's kind of funny, we always call it a CRISPR technology, but somewhat ironically, and this is a fight that I think everyone gave up on long ago, CRISPR refers to kind of the more adaptive immune part of the bacteria, not to the protein that is really what people are usually referencing that's kind of ripped out and then tons of engineering is done by us and others to turn into something that's useful.

10:41But anyway, so the Cas protein, C-A-S, is what then kind of takes these little short sequences, which get formed into a thing called a guide RNA. And that guide RNA tells the Cas protein, hey, I want you to bind to whatever matches that sequence. In this case, the viral sequence that it's like, please do not let this invade me and kill me and turn me into a factory for more viruses that then kills my friend. Right. So that's like the fundamental thing. And then if it there's actually some really other complex cast related systems where when it sees a new virus, it stores the sequence. And then so it can look for that.

11:15But the fundamental thing that happens when it sees the virus and it binds to it is that it then kind of chops it up. That's how it defends against it, because if the viral sequence, it can't really do anything to the bacteria because it's, you know, in little fragments. And what was really cool was that Jennifer and her lab realized that this can be a more generalizable technology that you can then take out of bacteria and use for things like human therapies or even editing plants. It really allows you to treat biology like you treat a computer and like engineering. Like you can actually change the sequence of life, which is DNA.

11:58so now when we use it for like a human therapy instead of having these clustered regulatory interspace short palindromic repeats that are having like viral sequences or whatever you design a guide RNA that's specific to the location in the human genome that you want to edit so in our case for our lead drug which is for high triglycerides that's a gene called APO C3 so it's this one location

12:19Corey: it's going to be right there Yeah, so out of all the billions of base pairs in the human genome, this is the one spot, and I really don't want you to go anywhere else, and that's very important, because then you might edit something you don't want to edit. So you really want you to go to this one. And then when it goes to that spot, it will cut the DNA. And there's other ways of doing editing that we and others have developed as well, where you don't cut, or you do epigenetic marks, I'm happy to get into this. But the simplest version -

12:44Grant: Yeah, I heard of one that was like a light switch where you can kind of turn a genome Yeah, or like a knob, you can kind of turn it on. Yeah, yeah. Um, yeah, these are all technologies that are built on top of this kind of fundamental principle. But the, the basic version of it that does work very well is if you're just trying to turn off a gene is that then it goes to that one location. It physically kind of like cleaves, like physically cleaves the DNA. Um, and then in the repair of that, um, there's some sort of mistake that's entered or there's like some sort of gap that's created. And then the, in the repair of that, it's never repaired exactly right.

13:18If it's repaired exactly right, it'll just be cut again. But whenever it's repaired, kind of in this like minorly incorrect fashion, that will interrupt the gene and cause it to not have any function anymore. It will kind of like garble it, like you're kind of scrambling the letters. And that turns off the gene so that it's no longer active in that cell. So that's the fundamental kind of principle of CRISPR. And some of the revolutionary aspects of it are that it's very programmable. So you can like just give it a different guide RNA and you can have the same protein. Because before, there were ways of doing this where you kind of redesign a protein every single time.

13:53You want to go after a different target. And not only is that just a very painful and, like, very long process, but also, like, you never could quite get maybe the specificity that you want compared to having, like, a sequence matching and a protein matching. And in addition, it allows you to then kind of build all these other techniques on top of it that we can get into. But that's how our lead drug works, actually, is that it simply cuts the APOC3 locus and then turns off that gene. And then there's a whole universe we can get into as well of how do you decide what gene and how do you know it's safe or not safe.

14:31That's the fundamental principle.

14:34Grant: Well, that kind of leads naturally into our next question, which is how you're using AI as part of this process. And in particular, your team told us that Mammoth can search a database of roughly 34 billion proteins for promising CRISPR systems and, I guess, these pairs that you're trying to cut. How do you search that? How are you using AI as part of that process and finding the information that you need? Yeah, so it's definitely evolved over the years of the company. And we apply machine learning AI. When I was doing my PhD, we just called it machine learning. Sure. It's changed now. We did applied linear algebra to many different parts of the process.

15:17And one of the original ones that we were doing extensively, even when the company was founded, was exactly that. So one of the things that we really innovated on is that Jennifer and Janice and Lucas and the team there realized that you could build up these giant, giant databases of just kind of what we call metagenomic sequencing data from all over the world. It could be anything. It could be like a sample from a river, a swipe of a table, a plant like a root or something. Just bacteria and microorganisms are all around us. So you just build up this giant diversity of data. And all these microorganisms have kind of different flavors of CRISPR because they have different evolutionary pressures that are causing these microorganisms to evolve in different ways.

16:01They're encountering different viruses, et cetera. So you can imagine there's a ton of diversity encoded in these somewhat random sampling data sets. So one of the first things we did was we assembled the largest possible data set at the time that we could of different metagenomic data and really kind of structured it well. And then in the early days, what you would do is you'd just kind of use something like a hidden Markov model. And you'd just kind of try and match. Cas9 was the original protein that Jennifer's lab discovered. And so we had that as a reference point. and you could kind of look and see, okay, what sequences kind of rhyme with CAS9, essentially, and pull those out and then test them.

16:39And we have a whole part of our lab here at San Francisco that has these high-throughput automated liquid handling systems that you might have seen. There's often like B-roll footage when labs want to theme.

16:50Grant: Is that where it's like the little dropper thing where the robot arms come in? Plates that are getting dropped around and moved and usually break. Oh, yes. So we have a bunch of those machines that can run just tens and tens of thousands of experiments at a time of saying, okay, you've pulled this out of nature. Does it have any hope of work? Because when you pull these things out of nature, they usually don't work at all because the bacteria has all these other components that are like helping it. And like the context of biology is often like, as you might guess, very complex. So then you have to figure out how do I reconstitute this and re-engineer it so it works like by itself and all these other things.

17:23And also usually they're like not very effective necessarily. You need to get them like way, way better. So why do all this though, even at that level? Well, one of the original, there's several problems we wanted to solve. And one of the original ones was what I mentioned about not being able to like really get these technologies to all the different cells in the body, like the brain and muscle and other places. And one of the big reasons why that was difficult is that Cas9 and the other original CRISPR proteins are quite large when it comes to proteins. So of course, like if you were to hold them all in your hand, you couldn't tell any difference.

17:54You wouldn't see them at all. But on the scale of proteins, they're pretty big, like 1 ,400 amino acids, let's say. And that means that it was really tough to package them in certain delivery vehicles that we knew could work for delivery outside of the liver and the blood. Like things like adenom-associated virus, for example, AAV. It's a popular delivery method that's been around a long time. So people had been, like, you know, figuring out, like, some cool tricks to maybe barely squeeze these in. But we were wondering, would it be possible to just have systems that are half the size of CAS9 or something?

18:30So that was one of the key things that we started looking at initially. We found that you can find systems that are a third or less the size, actually. And they have really cool ways that you can accomplish that. But that was one of the original ones. And then the other one is that you might want to do different types of edits. So going back to this other point around, do you want the knob that tunes and things? that's actually done by combining the calf's protein with other proteins and we can get into that but often you'll want special properties in the calf's protein itself in addition to having it be small because now you have all these other proteins that you've literally glued to it.

19:05Biology can be very simple in some ways as well. Just literally put it together. What's called a linker but yeah think of it as like you just like slam them together and you're like well if I localize this near this then like I hope I get this. You also might want specific properties, but the size was like the big one that we became famous for originally. So that was like gen one of using machine learning. I wouldn't even call it AI for sure to kind of do this. And then gen two is that you can try and use machine learning algorithms to say, okay, I've ripped this out of nature. I've done like some engineering.

19:43I need to get it 10x more efficient at editing in a human cell. What are the edits? What are the changes, the mutations I should make in that protein, they're most likely to achieve it. Because there's basically infinite combinations. Because this is a protein that has many thousands of locations that you could mutate of any reasonable degree. And you can do any combination. You could mutate all 1 ,000, or you can mutate any 1 ,000 choose five. So these are instantly trillions of combinations you could never run. It's not just even changing one location. So it's a combinatorial problem that you'll never search the space.

20:17So you need good heuristics for what actually might work. So that was like kind of Gen 2. And now Gen 3, which is enabled by these more kind of recent advances with large language models and things like that, is that you can actually then take those giant databases, shove them in the LLM and say, okay, forget trying to mutate specific spots, although you can still ask it that. It can still accomplish that work. Maybe you can ask it to just generate wholesale new proteins that are Cas-like, but you're not necessarily basing it on mutations in a known CASP protein. And you'll get really cool results.

20:53Like sometimes it'll look very similar, other times it'll look very different. Often they don't work at all, but sometimes they'll work okay. And then you need a little bit of the combination of art and science to know which ones you should really bet on and then spend a lot more time in lab cycles to then kind of work with the LLM even to say, okay, I want to improve this and like, what are the edits I should make in this brand new thing you gave me? So that's, I think, the evolution from, like, kind of mimicking nature to really generating nature on each end of the books. And, yeah, Mammoth has kind of been along that entire journey over our history.

21:26Corey: So your agents can talk to each other, but they still can't think together. That's the gap Outshift by Cisco is closing. OutShift is Cisco's incubation engine, building the internet of cognition, shared intent, shared memory, and guardrails that allow agents to work together from anywhere. Open source protocols and code, a system for intelligence that scales. Read the paper and experience the demo at outshift.com. That's O-U-T-S-H-I-F-T dot com. Check out the links in the description. Now back to our video. So how do – why is it, for starters, that so much AI work really zeroes in on proteins?

22:09Corey: What makes them extra susceptible to that? And how do they interact specifically with genes in this case? Well, my simple answer to why it's so zeroed in on proteins is it's the street lamp problem. And I don't know if you all are familiar with that. But let's say you drop – Could you tell us? Yeah. Yeah, so let's say you drop your Stumble Einheim drunk from the bar and you drop your keys. Where do you look for them? Well, you look for them where there's light. That's where you can see. And unfortunately, that's where the data is. There's two places we have massive amounts of data in biology. One of them, well, with one caveat, really where we have massive amounts of data in biology is sequencing.

22:47We have so much DNA and RNA sequencing data because we have technologies like Illumina sequencers that have been around forever that are very, very good at generating amazingly high throughput amounts of this, like trillions of base pairs. And the other area where we have less data, but it's extraordinarily well structured because of the work of thousands of scientists over decades, is protein folding. Because it's kind of a problem, and we have mass spec that's very expensive, but people have grinded it out for decades. And alpha fold and all of the amazing data. that they've done. On the protein structure data because there was a fair amount of it.

23:24And more importantly, it was like very well structured and like annotated. And then on the other hand, now you have people really throwing LMS at the sequencing data in bulk because there's just so much of it. So those are the two areas where there's a street lamp, sequencing data, protein folding. So of course, that's where people go. That doesn't necessarily mean it's the most productive area. I think that we are lucky and they're actually quite productive. But to be honest about why those are the areas, it's because that's where the data is. And why is the data there? Because we happen to have technologies that are good at measuring it.

23:55Corey: Where do you wish there was a street lamp? Yeah, actually I haven't really thought about that. I think the trivial answer to this that is not interesting is that if we really had full kind of human phenotype data from like drug trials that are like, because maybe this gets a little bit ahead of like, okay, where should AI really be applied? And like, where is it less useful? But maybe most people are applying it. I think that really all the AI work is preclinical right now, and we just don't have good data sets that link, that are public and well-structured and available that link preclinical to clinical outcomes.

24:29Because everyone loves to throw out these statistics that are true and very sad that one in 100 drugs from the preclinical and even one in 10 drugs maybe that reach phase one are going to fail, and maybe AI can change these numbers. right yeah um so i think that's going to be very challenging until there's data sets that actually allow you to link that because i mean already let's assume ai is as good as like the smartest drug hunter at doing pre-clinical work let's just assume that i don't think it's quite there but let's give it the benefit of the doubt okay the best drug hunters are doing the one in ten like you think you were just dosing people with whatever like only the best so like if that's what the ai can achieve great i mean you can fire the drug hunters i guess or whatever but like You're still doing 1 in 10.

25:12That's after a massive winnowing of stuff that didn't work before that. So I guess the optimistic view is that that'll be created at some point. Maybe it's 50 years from now. Maybe it's five months from now. I don't know. But whenever that's available, I am optimistic that AI can make progress in ways that maybe has been challenging for humans to do. But that's not really where people are focused. Everything's about, oh, I have a new sequencing LLM, or I have a new protein folding LLM, or I have a new thing that's going to pump out a hundred more preclinical drugs that I don't know who's going to test or whatever.

25:52Corey: The sexier ideas with all the curb appeal. And I get why. I can't say, oh, you should do this other thing, because I don't think the data exists. How do we get that data?

26:03Grant: Do we have to increase the number of trials that we do and then just record everything? How do we actually do that? Yeah. Increasing number of trials would help, but that is like a societal risk trade-off. Like on the one hand, this gets into the geopolitical things. So like China is doing like very fast, some would say somewhat loose clinical trials. Yeah, and they've had pushback about that recently. Well, more recently there's been news about deaths that were maybe not reported in the way they should have been in those trials. But you can argue either genuinely either side of it, right? Like patients are getting access to drugs that could benefit them that they wouldn't have access to on those timelines.

26:39On the other hand, are they really informed in the right way about the risk? And then some of them are dying. You know, that's a question not for any company to answer, but for the NIH. The whole system. Yeah. So that's something that society can decide. But wherever we tune that knob, I doubt it'll be all the way at China. And maybe it'll be slightly more than it is now. But OK, we'll tune that wherever it goes. I think the only institutions that have it right now are the major pharma companies. I mean, I think they actually have quite, who knows how well they've actually structured internally.

27:12I'm sure it's all over the place. But I'm sure some of them have pretty well-structured data on how preclinical models have failed across many hundreds to thousands of clinical trials. But for better or for worse, I don't think they're stupid enough to make that data public. Right? Because if that's the secret sauce, they should train on it. And I think that's how much money. Yeah, are they? They should be if they're not. Probably very quietly are. Yeah, biotech is slow, but we're not dumb. So, you know, or maybe like some of these AI labs will like buy a pharma company for that data. That's like a common rumor that I don't necessarily know is true.

27:49Grant: Like if Anthropic gets a big enough market cap, maybe they buy like Merck or something. Yeah, I think obviously it's not a market cap. The market caps of AI companies are so high now, like kind of in a crazy way, they could ingest all the pharma and like it wouldn't be a majority of their market cap i think wow the interesting question for AI companies is do you want to take on well a that type of business model it's like as if they went into the movie just because they have a good model at making movies do they want to get into the movie business it's very similar it's a hits business like 90 of your movies fail 10 of them become blockbusters yep like do you want to be providing the tools for people to make movies or you want to make movies you can argue it either way, but it's a similar...

28:26Corey: It's all about a good batting average, yeah. Right, but it's a similar kind of idea, but no one's like, oh, OpenAI's going to get it in the movies, but maybe.

28:35Grant: They sort of tried, and then they're like, maybe we'll do that later. And then I think the other side, it's a double-edged sword, because I think we all realize AI's getting a lot of pushback. I live in San Francisco, we all love AI, but if you go out, like I'm from Georgia, I don't have the accent, but I'm from Georgia originally, people hate ai like yeah and we can talk about why but like they hate it like it's just a visceral and you'll see it politically like people don't want data center like and they come up with like no one cares about the water they just don't like ai and it's the one way that they can fight back

29:08Corey: it's the one way they fight back yeah it's the one way it's kind of the proxy yeah it's a complete proxy so i think a lot of people like oh like no but actually it doesn't use much water no one who gives a fuck like yeah that's not the point like truly they they it's the one area where they feel a locus of control so they're gonna use it so i say all that to say i think drug development curing cancer like things dario talks about like that's the one area that i think people actually love ai like that even in georgia it's the one area where they'd be like okay fine if ai can cure cancer then i guess like that's the one area so like i think the reason to do it would be well A, benefit humanity and like truly, but B, like, I think it's like, it could turn the tide of maybe how AI is perceived if there's real delivering on these promises.

29:57But on the flip side, trials go wrong. You can have people die. And I don't know if an AI company is wanting to take on actual clinical trial risk. I think that's something you have to decide pretty rigorous, maybe post-IVO. I don't know. Yeah.

30:11Corey: Yeah. Fair. But it's not a thing I'd considered before either. Yeah. I mean, there's downside.

30:18Grant: Yeah. You hit on like three or four different questions we have in here. So I'm going to branch into something. But you brought up this one that we kind of had for last, but I think we should dive into it, which is, like you said, a lot of people see AI's disruption and risks more clearly than they see the benefits in their own lives. So what true medical breakthrough do you think would change that? And how close are we really to major progress on these things that people care about most? For example, curing cancer or slowing aging. Are we actually pretty close on some of these things in terms of the information that we have now, and we just need to scale some of the solutions through trials?

30:54Grant: Or is this something that we desperately need AI to help us with through better data and modeling? What's your take there? Yeah, I think the optimistic view is both in terms of AI and not AI, I think we're very close on a lot of this stuff. For example, I guess when we started on genetic medicine, I think we're about to see an explosion of that. And actually, that kind of genetic medicines in CRISPR is perfect for AI. Because if AI can churn out new... Because the hardest part now is we're really good at editing cells, which is kind of funny to think about. Because I actually always thought it would be the opposite way.

31:24I thought we would have great bioinformatics that would be like, okay, this is the exact location to edit. And oh, man, it's just so hard to edit there. How do we do it? It's the flip. Actually, we're getting better at editing cells in the human genome. But we still haven't really figured out, for example, for schizophrenia. incredibly heritable. Like if your parents have schizophrenia, you are so much more likely, even if you're kind of separated at birth, having schizophrenia. But as far as we tell you, there's maybe 100 loci, like which one, like where to edit. It's not like some one gene that maybe is causing it.

31:54And there's like holy wars about, do we need more data to figure it out? Is it because it's polygenic? So it's just so platropic that maybe no amount of data will tell you, even if you sequenced everyone on Earth. Anyway, who knows what the answer is. That's something where I think AI could be extraordinarily effective at parsing it out and giving you some sort of better answer than we have.

32:13Corey: Finding patterns and correlations. Yeah. And like figuring out, okay, this is the location that you should change. Because I think the big risk in drug development is usually you don't know if actually your mechanism works. Like you have a molecule and like, okay, like this pathway seems to be perturbed in the disease. I'll give the example, Alzheimer's plaque, there's like plaque formation in the brain. So people are like, oh, I guess if we had a molecule that reduced plaque in the brain, then maybe that would cure Alzheimer's. And, you know, now it's like, oh crap, maybe this is just a secondary indication.

32:44And like the plaque is just kind of downstream of whatever's causing us. We've spent a billion dollars in 10 years building plaque drugs and maybe none of it works. But with genetic medicines, you actually, it's beautiful. It's like you literally have a cause of the disease and you're changing the cause. So you kind of remove, not, nothing's ever completely true in biology, but like 99 % of the time, you can be pretty confident that you're not going to have nearly as much mechanism risk as you do for like any other approach that you're not changing the literal causal DNA for the disease. So I think it's actually quite attractive for AI based approaches.

33:20So I think that's one. The area where most people are working is definitely like antibodies because we've gotten very good and there's a lot of companies doing really great work at doing antibody design. I think that could be very promising for a variety of diseases, including cancer. I think where things will change is when something actually hits. Because unfortunately, we had a first wave of truly AI bio companies 10 years ago plus. Unfortunately, those drugs haven't really worked out in the way maybe we all hoped they would. Companies are still working hard, but it's been a while and it hasn't maybe gone as fast as we wanted.

34:00So I think that if there's another version of that where these companies go to crazy high valuations and then we're sitting here 10 years from now and it's like, none of them have approved drugs, I think that's rough. I think then it starts to lessen people's confidence in these kind of grand proclamations. So that's why I think that it will be important for everyone in biology, I think increasingly, to really think hard about like, yeah, what are the low-hanging fruit that maybe AI could help us pick in ways that were not possible before? Whether that's new targets for genetic medicine that are really good, that we can very quickly, in months, build a new CRISPR guide for and actually test.

34:39Or that's antibodies that there's some really hard binding pocket, but we know it'd really help us with some nasty cancer. Just slam AI against it, see if it can figure it out. But I think those wins are going to need to come in the next couple years

34:52Corey: or else I think, especially given the level of proclamation. Start trickling out. That's what we need. We need even a steady trickle of exciting news. Do you think simulated early testing by AI is a thing that could really speed that up? This is not mathematical based at all, but my assumption is that in terms of simulated testing of a broad swath, an AI is going to do in a day what a human is going to do in their life. and be able to really narrow in on things that are more promising, I would hope. Yeah, I think eventually it'll get there. You get what you measure, and unfortunately, I think a lot of the data still isn't quite there.

35:35Like, we're really good at folding protein, let's say, but then maybe the next step would be, okay, does that protein interact with this other protein? And people are spending a lot of time and effort generating those data sets, and it'll get better and better. But then the problem is, okay, you maybe get really good at simulating that, But all the data sets are these very clean data sets where the only two proteins in that tiny well across a thousand were the two proteins. But surprise, biology calls a hundred thousand other proteins that are active at the same time. Maybe one of them preferentially binds to that pocket over your protein.

36:05So congratulations, your drug doesn't work, even though it beautifully works in your both simulated data and your actual data. So it's nastier than you think. So I think that that's going to be the thing that just keeps happening over and over. is like every time you get a little bit better, biology will surprise you by how complex it is on the back end. But that doesn't mean it won't ever work. I think as we work our way up the chain, yeah, the goal would be not going from 10 % of drugs to interphase one to 100%. It might be going from 10 % to even 20%. That would be double the drugs. That would be insane.

36:41Which would change so much. How many lives would be saved by that? So I think it's the classic like people overestimate maybe the capabilities but underestimate how even real improvements can have just massive impact on both patients and financially. That's twice the market cap.

36:58Corey: Even a third less time as far as would go a long way. There's so many factors in there. This is such a fascinating subject. It's one that Grant and I talk about a lot. And we're always really interested in this end of things. what's what's your pie-in-the-sky hope that you'll see within five years um well in the Christopher space I think that you know if we all worked really hard and like poured effort into the space like we could potentially live in a world that has no like liver or blood genetic disease and like definitely five to ten years which is a crazy statement and like sure it's only use for now but yeah I think like there's line of sight there there's a lot of work to be done both in terms of like turning through clinical trials and also some additional work for trickier targets about how you can address it.

37:50But it's something that, yeah, if like we had like a sustained effort, we could do, which is pretty cool. Like that would change many, many people's lives. We need a whole industry of people cleaning data

38:02Corey: is what we need. If you've been displaced by AI, you should go work in data. But yeah, and then beyond that, like from like 10 to 20 years, doing the same for muscle and brain and other tissues. That'll take longer for sure. But yeah, it's the classic. People thought, well, when we sequenced the human genome back in the year 2000, people thought we'd cure all genetic disease immediately. And then the whole industry went through this giant, it was like dot-com bubble, but for biotech basically. But now I think we're finally, 26 years later, I think we finally really have line of sight to that type of promise and that type of opportunity.

38:45Grant: That's so amazing. How far are we from truly personalized medicine? So I don't know if you saw, there was a story that came out a week or two ago at the time of recording, where someone was, they used like AI as part of the process for an mRNA thing to make like a personalized, what's it called? Melanoma cure? The Moderna drug. Yeah, they were able to reduce some of it. How long really, like you mentioned, five to ten years for blood and liver until you can go to a doctor, get diagnosed, and maybe even sequence however it works, and receive a therapeutic design around your biology that can solve it.

39:23Grant: Is that even further out than we're talking, or how likely is that? It depends on the disease, but for a genetic disease, you could do that now if you had enough money, just being honest. So it's just scaling, right? It's more of how do you make that available and make it make any sense for the healthcare system. And I think that, I mean, that's another thing that's changing. Historically, so gene editing is a subset of a more broader field called gene therapy that's been around for many decades. And gene therapy is kind of notoriously this, like, expensive, like, you know, only for rich Western countries type of thing.

39:57But many genetic diseases, like sickle cell, for example, are actually most commonly found in, like, Africa, you know, other places around there. that would never really have an opportunity for gene therapy probably for the foreseeable future. So I think that scaling these types of technologies, not just for going to more developing countries, but also for how do you not bankrupt the healthcare systems in the Western world? We're seeing this with GLP-1s as an example, where basically everyone's going to be on a GLP-1 now. And how do you make that make any sense? or whenever there's like a really effective Alzheimer's drug, everyone's going to be on that.

40:36So there's more and more medicines that are for more kind of mass market. Like our lead drug is for high triglycerides. So it's like, you know, how many millions of people could benefit from that? So I think one thing that we have been excited about is in part because of actually things like COVID and using lipid nanoparticles that were part of that technology, that got scaled to more doses than any drug in the history of humanity. So like people learned how to manufacture this really, really effectively. and at much lower cost than ever would have happened without that. So trying to leverage what's already happened in the world to be able to create genetic medicines that are actually things that could be scaled to millions of people, not just like tens of thousands.

41:15And that can have trickle down effects of like, oh, okay, if now you have drugs that are much more cost effective, maybe for diseases that are even like really rare, like one person, maybe it makes sense actually to start going after these if you can do it in the right way. So it opens up kind of surprising avenues. And a lot of genetic medicines are ultra rare. There's like only 10 people that have that specific mutation. But then there's hundreds of thousands of people that have each of these. So it's like these giant populations, but you have to kind of redesign a drug for each one. So there's a lot of interesting work going on, including at Mammoth, about thinking about like how you can actually address this with a combination of maybe regulatory changes and also, you know, applying AI to design and manufacturing scaling.

41:57But I think in general, the goal would be for gene editing in particular, like, yeah, how does it kind of move beyond like really, really niche applications and go into much larger populations where maybe we don't have to take drugs for the rest of our life. And Glyph1s aren't necessarily the best example for various reasons, but like imagine if you could have like a permanent Glyph1. Yeah, who wouldn't do that? for some people. Side effects, things like that, you might want to engineer it a little differently. But what if you had that as an option? That could be really transformative. Also for people that don't have access to healthcare systems all the time.

42:38But it changes the business model of it too, right?

42:40Grant: Because I think Perna has for many years been used to like, oh, we're going to sell this drug to you every month or however long the interval is. And they're used to that continuous profits. So how would you price it if it's a once-a-life treatment?

42:54Corey: What's ZepBound going to cost if it's a once-a-life treatment? Is it going to be a quarter-million dollar house? You could probably price it under what a lifetime of the drug would be. But then the question becomes, let's say the lifetime of the drug is$2 million. And you price it at$500K even. Throwing out random numbers, right? Well, still, I think a lot of people might balk at, I have to pay$500K now versus$2 million over 30 years. Those are very different numbers. So there's a lot of interesting work going on about like, okay, and also another, the U.S. health, most drugs make most of their money in the United States, followed by Europe.

43:31So looking at the U.S. healthcare system as an example, we bounce around our health insurance like on average, like every three years or something, because it's tied to our jobs. So like, it's not like you, we don't have single payer kind of like socialized medicine. um so how to like why would an insurance company want to maybe get unlucky that they are the ones paying this bill and then everyone else gets the benefit or maybe it evens out and they should

43:54Corey: accept it or maybe if you stop taking it what happens if you like have to not or maybe for some people like for whatever reason something happens and they need to get redosed or you know like who knows so then maybe you can have a model where it's like oh actually um you still just pay over the 30 years, but you pay the gene editing company installments.

44:14Grant: Yeah, it's literally, I was thinking it's like a down payment on a house or something. Not just a down payment, a mortgage for your health. Yeah, and you only keep paying as long as the patient is cured. Yeah. And it's even a check on like, you know, okay, did it actually work? Anyway, so there's lots of models. That's interesting. But they are kind of new models that have to be figured out. Yeah.

44:36Corey: Yeah. And another question people often call this up with is like, oh, does that mean like pharma companies make less money? Maybe for that particular drug, but like if patients want a one-time cure, they want a one-time cure and you actually might make more money in the first 10 years, but then like you've cured people over time. But I actually think that for better or for worse, I know people like don't like pharma almost as much as they don't like AI. I think that it'll allow you to go after way more diseases. So the TAM of pharma, I think, could get gigantic. I mean, you start to be able to go after more common diseases, and you start to be able to go after way more different examples of rare disease.

45:13So I think pharma and biotech could end up making way more money even as you're curing people. I don't think it's necessarily some sort of like, oh, pharma would make less money. I think it's just a different structure.

45:24Corey: Like the conspiracy idea that money's all in there. Yeah, yeah, that it's like something suppressed. Yeah. They don't go for the cure.

45:31Grant: They just want the cure. I don't get invited to the games.

45:37Grant: One other idea I had here, which is kind of sci-fi, so bear with me. There's a big conversation around an AI about better models helping build even better models. This is called recursive self-improvement. Is there an equivalent process potentially in biology where better models could design better experiments? They could deploy these experiments into autonomous labs that run them. And then the data from those labs helps train better models and the loop compounds. could this create something like a biological intelligence explosion?

46:04Corey: Even into human wetlands.

46:05Grant: Yeah. What are your thoughts on that? I mean, biology at its core is that on some level. Biology doesn't always optimize for intelligence, maybe, but it's done a pretty good job with humans. So, yeah, I mean, I guess there's some proof of existence there. But, yeah, the question of, like, RSI, my guess is that probably is going to come more just, I mean, I have some background in statistics when I was a PhD. I actually worked on some of the early LLM stuff. Maybe I should have stayed in that field.

46:39I think that that's going to come more from probably lots of coding data sets. And then the big question in my mind is, you get it to do RSI on its own code, and it becomes better and better just as a generalist AI. And then when you plug that into a wet lab, is it generalizable? I think that's a bit of an open question right now. And would it be able to say, oh, even I need these data sets and spin up a wet lab and have the humans do whatever, or robots or whatever, do whatever it needs. But yeah, I think that'll be more gated by the coding, I guess, escape velocity. And then it'll be really interesting if and whenever that happens to see how generalizable it is.

47:19and can you take that kind of just general really good RSI model and then have it kind of plug into a biolab and make intelligent decisions about what it's doing. I think it will probably suck at first, but if it's truly RSI, it will improve very quickly. So once the coding part is solved, I think, yeah, it will start to hit other fields pretty fast, and it will be more a question of getting the right APIs for it to figure that out. Totally.

47:45Corey: Have you noticed any as models have progressed And as more and more medical models come out, models that are various specialties, have you noticed any, like, waves of, like, oh, we've got this new wave from this model that came out three months ago or something? We're starting to see new stuff. And maybe watched it improve over time for what you do? Yeah, I mean, it definitely improves. I don't think I've ever been shocked by, like, a step function the way people on Twitter often are. It's very hypey over there. My own inability to notice things. But I don't think that's a bad thing. I mean, yeah, they're definitely generally improving.

48:27Like the answers are... Sometimes I just think they're longer than necessarily better. But in general, I'd say the trend is very positive. I think, honestly, I notice the personality changes in the models more than I notice even the information changes sometimes. I think a good trend has been the models have become much less sycophantic over time. I think for science, that's a really horrible trait.

48:49Corey: Long horizon tasks have come a long way too as far as like, can you just go work on this for nine days? Yeah. So like the rise of agents I think is huge for biology because biology in particular is usually the sequencing together of like thousands or even just dozens of like kind of relatively simple but important tasks like a bioinformatics pipeline or something. You need to download the sequencing data. You need to clean it. You need to run it through this particular analysis method. You then need to cross-reference a bunch of papers. So I think biology is actually a really great test case of long-horizon agent tasks compared to, I don't know, insurance or something.

49:26And then the trick with biology, though, is how do you then get the wet lab? Because you can do a lot of just pure in-silico biological agents work, and there's people doing that. And I think they're probably rapidly improving because it's coding, but for biology. Right? So it'll have the same. I think the problem, though, is you're going to cap out on that pretty fast, maybe in the next 6 to 12 months, if I had to guess. So then you're going to need to somehow start to get new data, like wet lab data, into the system to actually figure out kind of how you're doing and honestly to start to get things that are more useful for clinical drug development.

50:02And I think that creates really interesting problems of even just training loops because you're used to being able to verify answers and do RHL and stuff relatively quickly. but for this it could be like weeks of like okay the system's paused waiting for an answer and yeah how to like even productively allocate chips when like your agents are sometimes waiting for weeks for something so it's just like an infrastructure problem but ignoring that yeah more generally yeah how does that end up kind of being figured out because you can really only accelerate biology so much and on the clinical side not at all like for human data generation um so yeah that's going to be a tricky one where in a few years yeah i think that's why a lot of these companies are also doing kind of like material science because the cycles are a bit faster so they'll be like oh we're gonna like do science biology and then you look and it's like material science which totally makes sense it's like then like in seconds often you can get a measurement off your metal or whatever and be like okay it's a room temperature semiconductor or not right um yeah yeah yeah i don't know that's an open problem really interesting i guess one of

51:14Grant: the questions where this inevitably goes to is in order to get some of that data we've heard stories of you know a company there is a specific company now that is like running ai models on like human brain cells so they're like growing brain cells in a lab and they're running an ai model on the human brain um does this at some point lead to in order to get more data growing like specific organs in a lab, like a synthetic liver, a synthetic what have you, and using that to run some of these experiments on. So it's not a human, but it's the next best thing. And then I have another question after that.

51:50Grant: Let's start with that. Does this lead there? Organoids have been around for a long time now, actually. And they are an incredibly useful tool. I think there's still not enough body of data of drugs that have been tested in organoids that then have human clinical readouts to make it some sort of definitive statement. But when I was a grad student, I found organoids one of the coolest things because like you grow these cells in addition in a very special way. And like you actually get if you grow like neurons, you get like a prefrontal cortex looking thing and a hippocampus looking thing. It's kind of scary.

52:18That's wild. I can't even imagine that. That's so cool. So you don't let it grow too far. And organoids is a great word. Yeah. Yeah. It's got like a cool name. So it was a really hot field for a minute. And I think it'll probably come back with the rise of AI because it's like an obvious like what if we generated data in organoids? unfortunately, I think it's like, yeah, it's good, but it's like, is it better or worse than a mouse? I don't know. It's got its pluses and minuses. It's not, and I hope I'm wrong. It's not going to be like, oh yeah, mice didn't translate, but organoids do. No, I really, really doubt that.

52:52I think it'll be another piece of data along with mice, along with just human cell lines in a dish, along with monkeys, but there's no reason to think it'll be like 10x more translatable. I hope I'm wrong, but that's my genuine guess. And then in terms of other approaches in terms of tissue growing, well, there's a great movie called The Island. I don't know if you'll remember.

53:17Grant: Yes, I know exactly what movie you're talking about. Where they try and do that and then they decide it's easier just to have humans. They make clones of people and then the main character. You can imagine, why not test a personalized medicine on your clone before you test it on you. Right, exactly. It's very sensible in a lot of ways. Sure, it's very tempting. Dubiously ethical for the clone.

53:42Corey: Freak the world out.

53:45Grant: What capabilities do you want out of an AI model to help you most here? If you had a perfectly generalizable biology model, could it be able to take all of those stray data points from the mouse, from the organoid, from the human trials and be able to solve some of these problems if it was truly general? Or what do you look for out of these models that you don't have yet that you wish you had? Yeah, I think I'll give two answers. Based on my experience, I think these models, I mean, I've always been a believer in you are what you eat. And I think that's true for models especially. And they just haven't eaten a lot of the key things they need to yet.

54:28I would like a model that could translate preclinical data to clinical. I think that's what everyone wants, right? So that would be the simple, maybe trivial answer. That doesn't exist, but it would be transformative. Other than that, in terms of maybe more day-to-day, I think that the models still struggle to interface with the physical world. If you were to just take a picture of a wet lab bench, I think a model would struggle to understand why an experiment went wrong. or what would be the next step or something. I think that could be pretty interesting. Because there's a lot of data analysis.

55:09And I think models have gotten very good at that, honestly, even creating figures and things. But when it comes to one of the most painful things is an experiment didn't work. In fact, maybe most experiments don't work. And then you have to sit there and think about, OK, why didn't it work? Like what went wrong? And that's a huge part of the time component of science, even, I think, is just debugging physical world stuff. So if the models could be really good at that, I think that would actually be more transformative than another bioinformatics.

55:40Corey: Dr. Martin, where can people go to learn more about you, your work with Mammoth? I think a lot of people will have an interest in circling up on this. Yeah, our website's just mammoth.bio. So pretty straightforward. Well, thank you very much for joining us. We appreciate you taking the time. It's been a wonderful discussion, man. Yeah, thanks for having me. Thanks again to Mercore for sponsoring today's episode. If you haven't yet, please take just a moment to like and subscribe. Bring the bell up there and it really helps us continue to bring you some of the best guests we can find who are either in the AI space or seeing other industries transformed by it.

56:16Corey: We love this stuff and find it just so fascinating to listen to. Also, if you haven't yet, please pop by the Neuron.ai and sign up for our newsletter. We'd love to count you as one of our subscribers there, too. But that's it we have for today. So, farewell for now, humans. проблема in yourmagacul...

56:42EHHE the to

56:59way to

From the publisher

Corey and Grant sit down with Mammoth Biosciences co-founder and CEO Dr. Trevor Martin to explain how CRISPR is moving from a laboratory tool toward one-time medicines.


Trevor breaks down how Mammoth searches massive biological datasets for new CRISPR systems, why AI is useful but still constrained by missing clinical data, and what it would take to build a feedback loop between AI models, wet labs, and human trials. The conversation also covers personalized medicine, the economics of one-time cures, organoids, and why a medical breakthrough may be the clearest way for AI to prove its value to the public.


Mammoth Biosciences - https://mammoth.bio


Subscribe to The Neuron newsletter: https://theneuron.ai


Sponsored by Mercor: Check out the AI Productivity Index (APEX) here: https://www.mercor.com/apex/?utm_source=neuron&utm_medium=podcast&utm_campaign=apex&utm_content=10_7


Sponsored by Outshift:“Scaling Out Superintelligence” Vijoy Pandey, January 2026. The technical whitepaper detailing the Internet of Cognition architecture, three-layer stack, and cognition state protocols. https://outshift.cisco.com/internet-of-cognition/whitepaper?utm_campaign=fy27q1_outshift_ww_paid_ioc-neuron-wp_podcast&utm_channel=podcast&utm_source=podcast


Internet of Cognition Interactive Demo Clickable walkthrough showing per-agent activity, intent, context, and collective reasoning across a multi-agent SRE system.https://outshift.cisco.com/internet-of-cognition/explore?utm_campaign=fy27q1_outshift_ww_paid_ioc-neuron-explore_podcast&utm_channel=podcast&utm_source=podcast

More from The Neuron: AI Explained

All episodes