What Happens When AI Gets a Lab?

0:00 Orangutan mothers arrange play dates for their infants in order to invest in their social development. You're turning the patient into a collaborator for training the neural network, for training the machine learning model. This is so good. They have a robotic arm that in 50 steps in 30 minutes can set up these lenses for this optical experiment, which we've talked about previously.

0:29 Hello, Internet. This is your captain speaking. Lester Nare joined, as always, by my co-host and our resident PhD, Krishna Chowdhury. We are back for a fun rundown episode this week with some special news. about our recent award-winning show as the recipients of a bronze signal award which we just got the news of this morning and so a big shout out to all of you in ffp nation for watching the show for giving us the inspiration a quick question for our resident phd how does it feel to be recognized for the fifth annual signal award in best podcast for science It feels great.

1:18 In the sixth annual Signal Award, I would like to get gold. Yes. So that's what I'm feeling. Gold is on the mind and we are going to get into three science stories today for a rundown. For those of you who joined us during the Nobel Prizes, welcome. Rundowns are where we do fun. a little bit more of a high-level overview of some science stories that we have found interesting, starting with an animal behavior story, the parenting techniques and sacrifices of orangutan.

1:48 Mothers followed by personalized medicine with AI using a chatbot to make epilepsy treatment more effective. And we're going to end with automated scientific discovery. The AI czars are telling us they're going to solve all human disease in 10 years. But what does that really mean? And there's an interesting story that actually came out earlier today.

2:10 It's going to help us look at the larger ecosystem of what is going to be required to have AI really impact clinical outcomes for humanity. As always, we're going to talk about the science from the ground up today because this is From First Principles. The headline here is Irresistible. Orangutan mothers arrange playdates...

2:47 for their infants in order to invest in their social development. I mean, as soon as I saw this, I was like, we got to cover it. We got to cover it. Because, I mean, it intrigues me in so many ways. As a new father, you know, I'm starting to think about the pros and cons of, like, putting my kid in daycare and, like, the social development that comes with it, but also the exposure to, like, you know, just getting sick and, like, all sorts of, like, nonsense that comes with that.

3:16 So, you know, there's a lot hiding in that headline. And when I was diving into the paper, there's a lot of rich science behind the methods and things like that that I thought would be perfect for this podcast. There's a fact that this seems to be a deliberate decision by the orangutan mothers. Like, they're not just, like, casually meeting up with other mothers specifically.

3:43 that have babies. This is a deliberate decision. There's a fact that they have to give up energy and time to do this. So there's a story where, you know, mothers are making sacrifices. It's so similar to how humans do things. 100%. And the science underneath that headline is really interesting. Like the way they collected years of data, analyzed that data, and then they can convincingly state that, you know, Orangutan mothers are consciously, or not consciously, but deliberately trying to do this.

4:17 I think it's really, really cool. Careful with the consciousness word with all the AI stuff going on. I know, right? It's up for debate right now. Exactly. I mean, if you ask me personally, my personal opinion is that, of course, orangutans are conscious. Like, just... Look at them and look at the way that they respond to their environment.

4:34 They respond to each other. They respond to stress. I mean, it's obvious to me, but there's going to be people that are like, you know, saying, yeah, you don't really know what consciousness is. Okay. Anyways, let's get into some history and background. Orangutans, they're the great apes. They're part of the great apes.

4:51 That includes gorillas, chimpanzees, bonobos, and us. Mm-hmm. The animals that are in this study are particularly the Bornean orangutans, specifically Pongo pygmaeus wormby, another taxonomical name that I'm struggling with. But these are the guys that live in Tuanan in central Georgia. kalimantan in the indonesian borneo so borneo is a giant island and there's indonesia has like a majority of that and i'll just know native uh for all of my nickelodeon fans the wild thorn berries is the first time i ever heard of the location of borneo because this family that traveled around is like wildlife videographers uh like that was kind of it was like i think it's literally in the theme song uh that and so it was really cool because they traveled the world but that was the first time i understood where the location was they must have been looking for orangutans in borneo because borneo is very famous for being like sort of a native

5:48 So the family had an ape, Darwin, that could speak to Eliza because Eliza could speak to animals. Oh, I see. That's the shtick of the show. That's the shtick. And they had a, you know, Darwin was around. It's a good name for an ape. Yeah. Clearly was done with intent. Yeah, yeah, yeah. So orangutans are a bit different from other apes, okay?

6:11 Because they are not that social. They are solitary, okay? In the sense that adult females, they occupy home regions that maybe overlap with neighbors. But in a real sense, like, they're very solitary. Solitary here means that the usual organization of travel and feeding, like, they don't go in groups.

6:34 They don't assemble in packs, you know? So they're like Americans, very individualistic. Very individualistic. In modern Americans, very much so, right? Modern society, very much so, in some sense. And maybe that tells us something that we can gather from the study, but we'll touch on that later on, right?

6:52 The idea is that rather than traveling every single day in a stable troop, they actually move as separate units. So the mothers will be separate, and if a mother has an offspring, they're going to move as separate units, okay? Some section of the forest that they like explore, but those are going to be mostly disjoint.

7:14 Now, there are good ecological reasons for that. It actually has to do with the diet of these orangutans and like where they live. The diet is mostly like fruit, large fruit eaters. Now, fruit is something that is kind of rare in the forest, and it's sporadic. It's not like, you know, at a regular cadence, you can be like, that tree is definitely going to have fruit, and then this tree is definitely going to have fruit.

7:41 So group living has its advantages, right? Because you can like share information, you can share protection, but the disadvantages are that There's a limited amount of food, and you're going to be competing for that limited amount of food, right? In a space where the density of food is very low. So that's why orangutans, I mean, from a biophysical perspective, there's a lot of...

8:07 theories out there for why they are solitary. It actually has to do with the physics of the environment itself. There's a resource constraint that they are conscious, not consciously, but actively operating around. Exactly. This is very different from other great apes. Chimpanzees, very famously, they are in troops, right?

8:29 There's a lot of resource abundance in the place that chimpanzees live, right? So they can actually administer in that. So now, in this space of solitary confinement of these orangutan individuals, the adults who are solitary in the forest, now let's consider the infant. Okay. Orangutans usually have one baby at a time.

8:52 And usually that baby happens once every seven to eight years. Oh, wow. So they don't have siblings. Yeah. In the traditional sense, right? It's not like a litter of kittens or puppies where, like, they have a bunch of people their same age that they can play around with. These are single babies. Yeah.

9:15 To a single mother, effectively, because the dad... Went to go get the milk. Yeah, yeah, yeah. You know, the dad is not involved. Let's say that. So this is a single mom with a single baby, no siblings, and you've got to introduce some bit of social, you know... Interaction. Interaction, yes. You have to introduce some bit of that.

9:40 You don't have group living. So what do you do? That's the challenge, right? That's the challenge. And that creates a developmental problem inside of the ecological problem. That's interesting. Yeah, there's layers. Right? The ecological problem is, well, because of the forest that I'm living in and because of the constraints that I have depending on the fruit and all of the choices that I have for food, I need to stay independent.

10:08 But if I have a kid... I need the kid to socialize in order to actually grow up and be sort of functionally accurate, right? Like, play is a very, very important part of upbringing. So... Play is actually very important. It's been studied for a very long time. Play is difficult science-wise because it's easy to recognize, but it's hard to quantify.

10:32 Yeah. This is a photo of two bear cubs playing. Now here, they're probably siblings, which is why they're close to one another, right? And they're playing amongst one another. I don't think bear cubs... really mess with other bear cubs because mama bears are very very cautious of other bears when they are nursing and when they're bringing up their kids, right?

10:57 So bears have this advantage of having siblings. Orangutans do not. Now, in orangutan research, social play means that you have non-aggressive interactions such as tagging, wrestling, chasing. It's the stuff that humans are pretty used to. And there's a long history behind why play evolved. I mean, one idea is that it rehearses adult skills.

11:21 So... Even if the adult is not meant to be in a social setting, play can still be important. If the adult is meant to be in a social setting, then of course play is important, right? Because you're getting used to socializing. But you could have the argument that like, well, orangutans, they're asocial.

11:40 Why do you need to play? Well, play gives you a kind of emotional regulation, and it gives you the ability to respond to something that... That is unexpected. Right. Right. If you're doing self-play, then it's kind of a self-feedback. You know what's coming. You know what's coming. When you're playing with someone else, you don't know what the other person is going to do.

12:01 And that ability to respond to an environment or to a stimulus that is unpredictable, that can have a lasting effect. It's beneficial to your own survival, ultimately, because you want to be able to react to unknown stimuli in a way that is beneficial to your own survival. But that is not also an overreaction unnecessarily.

12:22 Exactly. Yeah, so there is a benefit to play even if the population itself is asocial. That was the first thing that when I was reading the paper, I was like, that's actually... interesting to think about, right? Because it's trivial that play is important for social animals because you need to be social.

12:40 But even for asocial animals, there is a component of play that is important for your development. Any case, do orangutan youngsters play? Well, in 2024, Julia Kunz, who is a co-author on the current study that we're going to be talking about in depth, she published this article that showed 11,000 observation hours from 99 young orangutans at two sites, one in Borneo and one in Sumatra.

13:10 And she showed that social play was more frequent in the Sumatran site versus the Borneo site. But actually, if you controlled for the amount of times that... play was presented, it was the same. So if they're given the opportunity to play, youngster orangutans will play. So now let's try to hone in on the other part, which is how many times are they given the opportunity to play?

13:37 That's what this paper is about. Because there's a difference. Right. It's like if I'm not given the opportunity, then I can't do it, right? Right. And the new paper builds on that groundwork. This is the paper that came out in the Proceedings of the National Academy of Sciences in October 2026. I see you've avoided the acronym.

13:54 PNAS. That's what it is. That's what everyone says, okay? Everyone says, stop saying the acronym. No, no, no, no. That's a joke. Everyone says it. This is from Max Planck Institute of Animal Behavior. And there's a few authors from the University of Zurich. So the earlier work established that orangutan youngsters play and that the opportunities for that play varies, right?

14:20 And here what we're saying is that the maternal association could provide for that opportunity and the moms actually make a concerted effort in providing that opportunity. Okay? That's what's so cool about this. Okay. The methods are incredible, I would say. So since 2003, Van Nordwyck and her team have been following mother-infant pairs at Tuanan Research Station on the island of Borneo since 2003.

14:55 Over a decade of research, they've observed encounters between mothers that are rare. But when they do happen, the mothers are very cautious. But the kids, as soon as they see each other, they just like run at each other and start a blast. Yeah, they're having a blast. The mothers are very cautious. They've, you know, developed the whole adult thing of being solitary.

15:16 The kids, they're having a great time. they seem to come at a cost to the mothers. Oh, okay. The cost being it's a really long and hard day for them when they have to arrange these playdates. Are they even arranging these playdates? And if they arrange these playdates, are they doing it at an expense to their own benefit, right?

15:44 Because they could be doing other things. Resource gathering. Yeah, they could be eating. They could be resting. Instead, they're doing this, right? So is it a concerted effort? After 15 years, they've amassed 32,000 hours of observation from 31 mother-infant pairs. That's actually incredible, first of all, right?

16:04 Like 15 years of work, 31,000 hours. That's like on the order of like a year, a year. of just like continuous observing orangutans. It's incredible. The basic technique is something called a focal follow. This is a very famous photo of Jane Goodall doing a focal follow on chimpanzees. The idea is that you have researchers that choose an identified individual, and then you basically follow its behavior from sundown to sunset, or from sunrise to sunset.

16:41 where the person is. So you have like a GPS monitor and you're like, okay, I'm going to monitor where the orangutan is right now. I'm going to monitor what the animal is doing at that moment. So is the mother feeding? Is the youngster playing with whom? And roughly every 30 minutes you log a GPS location, right?

17:01 So you, you get a kind of GPS of, of where this person, where, where this animal is going and you get a, every two minutes, some kind of log of the activity over the entire day. That's a rich data set. Yeah. Incredible, right? Yeah. And it requires a human being just straight up following this orangutan or this mother-infant pair for the entire day.

17:27 Pretty incredible. For days and days and days. And it's the same individual that you have to follow for days and days because you want to have some type of longitudinal data on the same individual. Yeah. Now, the data set here includes 3,000 plus follow days, 2,000 of which were full nest to nest days, meaning like there's like mothers involved.

17:49 Only 6.3%. So only 6%. Did they record play? That's incredible. What we're saying is over the course of all this observation, a less than 10% proportion of it has this peer play concept, has the thing we're looking for. And that's why you need such a large data set because it's so infrequent. Yeah. And you need a large data set because if you have infrequent events, In order to do good statistics, you need a large data set, right?

18:21 Reminds me of the IceCube neutrino detector. 100%. It's the same thing. It's the same exact paradigm, right? Because you need a lot of events in order to hone down on a very rare, like, per chance. So now let's get into the figures. They've collected all this data. What do they actually see? Let's start actually with figure C here.

18:40 This is figure 1. In C, on the x-axis, you have the age difference between two offspring that are meeting up for a play date. And meeting up meaning, you know, they're playing with one another. This is obvious from a human that's like looking up and seeing them. On the y-axis is their play rate, which is some normalized sense of how often do these two individuals get together.

19:05 We're looking at C. Okay, yes. Right? We're looking at this part right here. Yes. And there, you can see there's a negative correlation, meaning that as the age difference gets bigger... the play rate diminishes. Yeah, that makes sense. Kind of makes sense, right? My little brother and I have a 13-year age gap.

19:22 Yeah. So our play rate was much lower than siblings who were much closer together. Exactly. The other cool thing is on the y-axis, it's actually exponential. That's a log scale. Okay. Because it goes like 10 to the minus 1, 10 to the minus 2, 10 to the minus 3. So this is exponential decay. It's a very strong effect that you're seeing, which is like, it seems that like the...

19:42 The closer you are to age, the longer you're going to play, right? Okay. On A and B, what we're actually seeing is the effect size based on different markers that the human beings that are on the ground are actually marking. For example, you can track the relatedness of two individuals, and you can track the fruit availability in the...

20:07 forest. Because you could imagine, right, fruit availability is a confound. If there's a lot of fruit, then of course the mothers are going to be foraging a lot, and so they're going to meet up a lot, right? You want to disentangle those and see what's the actual effect. Well, actually, fruit availability is not that high.

20:25 In effect, the correlation is not that big. On the other hand, relatedness, like whether you're related to the other kid, That's high for occurrence. But if you control for occurrence, the amount of play that you do per occurrence is not that high. If you're family, I'm not about that life. Yeah, like we'll meet up a lot.

20:50 Yeah, but. But I would play with you just as much as I would play with a friend. You see what I'm saying? Yeah, yep, yep. Like there's a difference here. Like maybe the occurrence is higher with family because like the moms know each other. But just because your family doesn't mean I play with you more.

21:05 Yeah, yeah. I play with you the same amount that I would with like a stranger's kid, right? The two big anti-correlations are, again, age difference. That's the very top. It's like if the age difference is very high, then the amount of play rate is very low and the occurrence is very low. So it seems to be that the mothers...

21:25 have honed in on who has the offspring that is about the same age as my kid. That's fascinating. Isn't that crazy? That's so good. Like, I mean, they're keeping track of who has the right age group of kid. Now, the other cool thing, the ranges overlap for moms. Because I told you, right, the moms have a...

21:52 They're solitary, which means that the range of forests that they forage might not overlap all that much. But overlap is higher for moms that have similar age group kids. Oh, that's really good. Right? That's really good. Because you have this data. Yeah. Over 15 years, you've been collecting GPS data.

22:10 Yeah. You can actually tell where the moms forage, and the overlap is higher, it's stronger, right? for two moms that have similar aged kids. Which is pointing to this idea of some intentionality of going to that place of overlap. Yes. Because I know the other mom that comes around here, their little Bobby and my little Timmy are about the same age.

22:31 About the same age, right? So they can play around. This entire paper is going into that intentionality. Yeah, that's really good. Can we prove intentionality? In orangutan mothers. That's the point. And the way they've structured these questions and the data points they're tracking in order to reach the conclusions about intentionality are very clever in the areas that they're trying to cover to remove other confounding variables.

23:01 Exactly. Exactly. Yeah, and that's, I've always find this fascinating because like, especially with animal behavior, it's very hard to think of all the confounds, but these guys are doing a pretty good job. It's quite nice. The next figure is figure two. Here we talk about what it costs for the mothers.

23:20 Right. Here, the researchers are comparing three different categories. There's days when the mother was alone with her offspring. There's days when another orangutan is nearby, but there's no play. And then there's days when there is infant play. Okay, those are the three sort of categories. The middle category is really important because here we're simply disentangling between, like, okay, if another orangutan is nearby, does that have an effect versus another orangutan is nearby and my infant is playing, right?

23:54 There's two different things that are happening. And here we're seeing an actual difference, right? On the left-hand side, this is the daily amount of path that the mother takes. When the kid is playing, the path length that the mother traversed in the forest is a lot longer. Almost 20% longer. Which means that she's exerting more effort to actually go from one place to another.

24:19 On the right-hand side, figure 2c, that is the proportion of time feeding. This is what's most important to me and most salient to me. The amount of time that the mother spent feeding herself was lower on days when her infant played with another friend. There's more caloric deficit from the amount of movement while having a lower amount of caloric intake at that stage where you're having this higher caloric deficit.

24:51 Yeah, yeah, yeah. It's a twofold... Yeah. Yeah. Effect. Yeah. And, you know, as human beings, part of this is going to feel like, oh, obviously. Yeah. We talk about childcare and all these things. Yeah. And you are experiencing this personally right now and going through that. Yes. Okay. With human beings, for sure.

25:12 Yeah. Yeah. Yeah. Yeah. But we're trying to empirically show this. in a different species that does not have all of the things that we have. And so we have to ask what may seem like basic questions. Yeah. Because we don't know unless we have the data. Right. And that, I think, is what's key about this paper, right?

25:31 It's like with human beings, I can just ask mothers, and they'll tell me, yeah, it's hard, right? But here, I think it's really cool that using data from behavior over 15 years is... we can empirically show that orangutan mothers are kind of feeling the same stresses that human mothers feel, right? The last figures that I'm going to talk about I think are really, really cool.

25:56 The next step is really asking whether mothers' movements change before and after play, okay? Like, is it intentional? You know when we talk about, like, intentionality? Yeah. Well, here... What we do is first we define each mother's reference core using the average of her GPS location from the preceding year.

26:16 365 days. So we've got a range, right? Because we need to be like, where was this mother before the foraging? Like, what is her control? And we're like, okay, this particular mother... Mindy goes in the green region of the forest. And another mother, Janak, goes in another region of the forest. And there's a little bit of overlap because they have similar aged infants.

26:44 Okay. We've already covered that part, right? Now, Let's talk about the day when the infant play happened. Okay? Let's look at the day before and the day after. The day before, you can see both mothers actually venturing into each other's zones, which means there's planning involved. Mm-hmm. They're scoping the scene out.

27:07 Yeah, they're like, let's try. Let's try to find the other woman who has the kid that's the similar age as mine. And then let's try to meet up. That's the yellow. You see Mindy going into Jannik's zone. And then the day of the play, they actually play around. And you see where the play starts and where the play ends.

27:30 And they're moving around in their territory. And the day after. Incredible. The day after Mindy's like, you know what? We're good. That was enough. We're good. We're good. We're going to hang out entirely in my own zone. This is incredibly, like, it's so visceral and visual, because you can see it visually.

27:47 Yeah. You know, and again, the pre, during, and post. This is why anytime you do data capture, like, you want the pre and post. You want controls. Of the event you're tracking for. And it's the distinction between the day before and the day after is so in your face. Yeah. And they have statistics to prove this because this is obviously an anecdotal thing, right?

28:12 Sure. One of the statistics is this directional bias. I mean, these are the statistics on figure 4e. They're actually showing the difference between like the correlation day before, day after – There is a huge difference in the behavior. Yes. The directional difference is what's really cool. In the day before, they're moving towards the other...

28:35 territory that's the yellow compared to the control on the left yes and the day after they're moving away yeah yeah and it's you know i i one of the things that as we've gone through so many episodes of our last year almost and a half now is really getting a strong understanding of the the types of data that are gathered and the type of analysis that's done on that data and understanding how those three component parts connect to conclusions you want to draw.

29:13 And with animal behavior, especially we've done a lot of these stories. It's a very complex, all biology is complicated. Now we're adding behavior on top of just the biological substrate. And when you see something like that, it's, it's very, it's very, I don't want to say the word convincing, but it is present.

29:36 Yeah. I mean, to me, it's like it's confirming a notion that I already had, which is that like, of course, these animals are like doing all these things. Right. But it's just it's pretty awesome that like with 15 years of data and all of this effort, we can like quantify these things. I think it's I think it's absolutely amazing.

29:59 This makes me think of the whales have culture story we did. And we were having this very emotive reaction about it because we were showing something that might feel obvious. Yeah. If you're just like, of course, whales have culture and things like that. But like, how do we know? Yeah. I think that's kind of cool.

30:17 And I think it's cool that we can like say this is how we know. Yeah. Like if you were to say, oh, the orangutan mothers are solitary, but they... The headline, like, just looked at, oh, how do you know that orangutan mothers are setting up playdates for their kids? Well, actually, here's the one, two, three, four.

30:36 Here's how we removed other elements from it. And it's well-constructed. The methods are well-constructed. Yeah, yeah. I really quite enjoyed this, yeah. The evidence here supports active maternal facilitation and is consistent with planning. as we were talking about just now. Now, the broader scientific contribution...

30:59 concerns paternal investment because in evolutionary biology, we think about like investment, meaning effort or resources devoted to an offspring that have a cost or trade-off for the parent, right? It doesn't require some kind of conscious calculation. We usually think about it in terms of nursing, caring, and protection.

31:19 Those are the obvious examples. In this case, as something that we need to talk about when we talk about paternal, parental investment, right? The fact is that like a mother can shape development by changing which other animals her youngster has a chance to meet. It's part of a growing movement. There's been a growing movement in the science of behavioral analysis these days where, um, The biologists are now starting to study solitary mammals, and they're saying actually play and socialization is important even if these solitary mammals are kind of doing their own thing.

32:11 Like you can't count out social... and social interactions just because, you know, by tradition, these animals are solitary. Now, there is a second implication, which I found quite intriguing. It was in the discussion in their paper. As an Indian, I found it quite intriguing. Here's why. So group living means that, you know, young animals...

32:43 they meet potential partners almost automatically. If you live in a group, you get introduced to potential partners pretty regularly. If you're like an orangutan and you're solitary and your mom's solitary, which means that you are solitary because you're tied to your mom, you're not going to meet a lot of the opposite sex on an organic basis, right?

33:08 Just from Brownian motion of like the orangutans. So perhaps in a dispersed society, someone has to bear the cost of making these encounters possible. Right. Kind of like an arranged marriage system. Yeah. Right? And this is in the discussion of their paper. They're saying, look, like, you know, in a solitary society, This is an evolutionary advantage.

33:38 Yeah. Right. Because the moms like, okay, if we get real nitty gritty down to it, right? The moms care about their genes being passed on. This is about evolutionary fitness, Darwinian fitness, so on and so forth. Right. That's going to require their kid in a solitary society to like kind of ingratiate themselves with.

33:58 individuals of the similar age right what if what if that's the motive yeah yeah yeah right right look look that's the question now we can ask that question now and and the more data we have we can ask that question and that's actually what's next right is whether the investment actually ultimately pays off In an increased fertility of the offspring.

34:22 We don't have the answer to that question yet. But as an Indian son, that kind of reminded me a little bit about some of the struggles that I've had. This is so, I love the connection. You know, it's funny. Yes, yes. I will make, and this is a big spoiler alert for anyone who's not watched the new Netflix TV show below.

34:46 So, spoiler alert. Fast forward like a minute or two. What's funny about what you just said is... In this show below, which is about this sort of local town in Newfoundland that gets moved around because some rich entities want access to resources where they live. The main character got moved when he was a kid and now he's an adult in an island next to the island he grew up on.

35:17 in this Bay Area. And one of the plot lines in the show is that they thought where they were moved from was abandoned, but there was actually a family, a family and a few other individuals that had stayed there that were solitary. And there was a daughter that the dad was struggling to find a partner for and there was this sort of scheme where the main character's son got whisked away to the solitary island to be the potential partner and this idea of the the parent having this intent around this evolutionary event like one thing there it was it was and it was a core tension point in the show that had anyway but like it I'm just bringing a popular culture reference to something that you're talking about here because it's literally narrative storylines.

36:11 Yeah. And like the number two show on Netflix right now. Yeah, yeah. And that's what the authors are proposing here, right? In their discussion, they're saying like, look, like as a solitary society – The moms are also thinking, look, I kind of want grandkids. I would love that. And even these orangutan moms are thinking the same thing.

36:28 They're like, look, I want orangutan grandkids, which means that you're going to need to meet people of similar age. It kind of explains the negative correlation between age difference and meet up. Very good point. It kind of explains the intentionality behind all of this. I think – It's quite nice. So if – I like the theory, you know?

36:51 And of course, like there's going to be a lot more data that's going to be involved because what you're going to have to do as scientists, you're going to have to now follow the offspring. Yeah, right. And see what their reproductive success is over several decades, right? This is a long, longitudinal study.

37:07 Yes. A long, long, longitudinal study. Exactly. But I'm willing to bet that I think that's what's happening. very good very good i just want to know because usually in the context of below the mother was dead which is why the dad was doing the job and so that's part of the gender difference there but like obviously here it's like all moms it's the moms yeah it's here it's generally speaking generally speaking general speaking look like we we know how it is okay even in humans it's it's usually the mom that's how it goes yeah Fantastic first story.

37:43 The animal behavior stories are always so interesting because it's kind of not obvious for a layman to understand, like, how would you even go about asking the question? Yeah. And they're so clever at it. We are going to transition into our next story, which is going to be about AI chatbots and epilepsy.

38:01 But before we do so, I'm going to do a quick... Quick commercial break here for some housekeeping. Welcome for those of you who are joining us. This is From First Principles, the best science show on the planet. I have one very amazing announcement about how you can watch the show. If you're an audio listener or have been an audio listener, we do a lot of overlays and visuals in our show to kind of talk through the fundamental science, the methods.

38:27 It really helps you get an understanding of why these stories are so important. And I'm very proud to announce that finally, after weeks of promo, that the podcast is now available in video on Apple Podcasts. So for those of you who have iPhones or listen previously on Apple, you can now watch the show.

38:50 It's also available in video on Spotify and YouTube. On social, we have so many clips that are a way that so many of you listen to us. If you've come from an Instagram clip... and are now watching the full podcast, leave a comment below letting us know. You've come from TikTok, Facebook, etc. We'd love to know that you've made the transition to the full show.

39:11 It's super helpful. You can always get all of the research papers that we talk about at our website, ffppod.com. And we are now going to move into our second story of the rundown, AI, chatbots, epilepsy. How do they all connect? Well, we've covered a lot of AI doom and gloom. Yes. Because, well, there's a lot of AI doom and gloom.

39:38 Yes. But this one, I think, is genuinely uplifting because it's the promise of AI that got me excited about it in the first place all those years ago when we got the first chat GPT moment. Everyone was like, oh, this is going to change the world. The future. You know, it's going to be the future. This isn't the apocalypse.

39:56 This is actually great. Now it seems like, oh, I don't know. Yeah. In any case, this particular story, it's about personalized medicine. This time it's for epilepsy patients. The story comes out of the University of Pennsylvania. The headline reads, AI platform allows real-time conversation between patients and their medical devices.

40:19 I think it's really, really cool. This is interesting. Now, in this case, when I say personalized medicine, I'm talking about really a personalized machine for medicine. Okay? We've covered things like this in the past. We have pacemakers that regulate your heart. We have cochlear implants that digitize sound.

40:36 And we've got deep brain stimulators that suppress the tremors of Parkinson's disease and epilepsy. Okay. Now, these are marvels of modern engineering, but they're all entirely silent to the patient that is housing these machines, right? Like, if you get a medical implant today, the device records your physiological data, and then it stores it in some black box, effectively, right?

41:03 Some kind of USB drive or something. And then months go by, weeks go by, depending on what the treatment is. You go to the clinic. the physician downloads whatever the data is and then only the physician really understands what's going on and then he or she or they are going to be going through and understanding and you're essentially locked out of that loop that's not actually really good i never thought about that yeah like the patient who is creating the medical data is out of the loop, right?

41:33 The physician kind of just like uses the patient for the data and then tries to understand the patient from the data. But there's no real feedback that's happening, right? Nowhere is this disconnect even more critical than with the treatment of epilepsy. Globally, around 70 million people live with epilepsy.

41:52 Epilepsy, a third of them get seizures that simply do not respond to medication. Right. A seizure is basically a temporary episode of abnormal, excessive activity in the brain. It's like synchronized neural networks. It can sometimes make a person experience convolutions, brief loss of awareness, unusual sensations, disturbance in memory.

42:17 There's a myriad of different things that a seizure can do. And a single seizure does not automatically mean epilepsy, but if you're prone to this recurrence, over time, then it's reason to diagnose with epilepsy. Now, sometimes drugs work, but then there are drug-resistant patients. As with almost every brain disease, there are drug-resistant patients.

42:44 In this case, one of the ways that we try to combat it is with implantable neuromodulation devices. This is where... We stick an electrode into the brain that has the ability to read the electrical signals and then pump electricity in. So what this thing is going to do is read brainwaves, and in principle it can detect seizures using the brainwave profile, and then it can also provide current into the brain to mess with the neural circuits and then perhaps stop the seizure.

43:18 And the history of these types of devices has a rich, rich history. You've got the earliest FDA-approved devices in 1997, something like the vagus nerve stimulation. This is purely open loop. This is something where you actually stimulate the vagus nerve that's down near your neck. Mm-hmm. Not actually in the brain.

43:44 And what you're actually doing, you can think of it as like a lawnmower. Not a lawnmower. A lawn sprinkler. Okay. Okay? That just like turns on and then it like sprinkles. It's doing that but with electricity. This is what we would call open loop. Okay. Because there's no feedback coming from the brain.

44:04 And there's no, it's just, it's on a timer. Yeah. And it's just like going. It's just stimulating the vagus nerve at regular intervals. It doesn't know if the grass is growing. It doesn't know what patches of the grass are growing more than others. Any of that kind of stuff. Any of that kind of stuff is just blindly implementing.

44:23 I know it needs water. Yeah. And, and, and. It had its effects. Sure. And it was great. 2013, there was a massive leap into closed-loop systems. So this is the RNS system, which is the responsive neurostimulation, meaning I've got some kind of read mechanism that is going to maybe identify that a seizure is happening and then only stimulate when the seizure is happening, right?

44:53 So this is where the responsive thing comes in. Mm-hmm. Now, instead of open loop, we've got a closed loop kind of thing, right? The device knows when the seizure is coming. But at the same time, it's not closed in the sense that the patient is aware of what is going on, right? You can't tell the patient, maybe you need to sit down.

45:15 Maybe you need to pull the car over. Alert, alert. Like, yeah, alert alert. You're about to have a seizure. Right. Which would be obviously great. That would be great. And that's where the breakthrough currently comes in hand. The patient was never involved in any of the communication cycle of this. Obviously, we weren't there yet.

45:34 Yeah. But... Because it required like a human being with a trained experience to understand all of these things. Now, with the advent of machine learning, you can put the patient in the loop of his or her own medical degree. Destiny. Yeah. It's pretty incredible. This is a really big deal because I think it also – there's this whole era of like quantified self and all this medical information from all these devices we have around us.

46:03 It's going to help you improve your well-being. But part of – and that's a different thing than medical and clinical treatments and solutions. I'm not saying they're the same thing. But there's been this big promise with the rise of technology as bringing us closer to understanding our own bodies in real time.

46:22 And this is a perfect example of how that can have a hugely important impact for epilepsy. Yeah. Yeah. And I think it's not only limited to epilepsy. This particular case is taking epilepsy as kind of like this microcosm because it's such a rich data set in terms of like, you know, you get intracranial stimulation.

46:49 There's a lot of data that's going on. It requires a lot of expertise that perhaps a machine learning model can elucidate. But it's absolutely an incredible paper. It's a team out of the University of Pennsylvania. And what they've done is fundamentally re-imagine the patient-device relationship by bridging this EEG that comes with the patients of epilepsy with agentic large language models, right?

47:18 That's the key. You're tying in the brainwaves from inside the brain of these epileptic patients with now agentic models that use large language models. This is the promise that has been, again, large language models are not just a tool in the context of a chat bot. Yes. Is kind of the point here. Yeah, they've built a kind of interface that allows...

47:43 the patient to maybe interface with the chatbot, but the chatbot now has access to all of the data from within the patient's brain itself. Which is... It's incredible, dude. Yeah. Now, I actually, like, you know, from a personal standpoint, my PhD had to do with... electrophysiology, right? Where, you know, you're taking electrodes, you're sticking it inside the rodent brain while the rodent is freely behaving inside mazes, inside, you know, behavioral tasks.

48:18 One of the... The bane of my existence during my PhD was deciphering between real neural signal and all of the artifacts that come with behavior. Because if the rodent is chewing, that's going to cause electrical signal, right? If the rodent is going into like a micro nap, that's going to cause electrical signal.

48:43 This problem is real even for epilepsy patients, right? Because if you've got electrodes that are going into the brain, there's a giant signal-to-noise problem. Because brainwaves are incredibly chaotic and highly variable. Here we're seeing brainwaves where chewing artifacts come in. All of those jagged things, that's when the person is chewing.

49:08 That's crazy. If you have a really terrible machine learning classifier that just thinks you're having a seizure every time you chew, that's not going to be great for two reasons. There's this thing called alarm fatigue, which is basically the boy who cried wolf, right? If every time the thing just says, hey, you're going to have a seizure and you're not having a seizure, there's so many false positives.

49:36 Well, then you're going to just disregard the thing, right? And two, if you're kind of someone who has seizures and has anxiety... That's going to add to your anxiety and it's going to like be detrimental to your wellbeing at the end of the day. Like the whole point is to get better and like having these alerts is making it worse.

49:56 Right. You're constantly put in a state of a heightened, exactly. You know, um, Defense mode, anyway. Right. And the reason why this is so hard is because historically, tuning these algorithms requires an expert neurologist to manually go through all of the data from your brain and be like, yep, that was you chewing, that was you not chewing, that was an actual seizure right there.

50:20 Right? You got to have some MD be like, that was an actual seizure. We got to tag that, not tag the other thing. Now... You flip the script entirely, right? You're turning the patient into a collaborator for training the neural network, for training the machine learning model. This is so good. Dude, I think it's so cool.

50:42 I mean, there are things that are involved here that they were very careful about. You've got to be HIPAA compliant. And these guys at UPenn, they made a system that features a HIPAA-compliant instance of the GPT-40 mini. So this is one of those mini models that OpenAI has, right? Which means that you can, like, put it on a local thing.

51:04 Right. It doesn't need to go to the cloud. It doesn't. Yeah. Yeah. So now patient data is not going to the cloud. And, like... you know the the the levant al poge and buck master uh a problem nightmare where right where it's like they if you guys don't know the navier stokes episode right buck master was talking to open ai codex about his solution for navier stokes and now lo and behold lo and behold a week later yeah open ai came out with navier stokes right like i I don't know.

51:34 I don't know how that happened. Well, in this case, they've got a GPT-40 mini that is local, right? So there's no way that HIPAA... That the data from the patient is going into the cloud. And it's not trivial to be able to execute something like this on one of these smaller mini models. Like that's been part of the challenge.

51:55 Exactly. Generally is you want the capability. Yep. As you go to smaller parameter models, you lose some of the capability you're looking for in an LLM. So to be able to work within that constraint and still get the solution is, I think, an interesting part of the solution. the success here. Exactly.

52:13 And one thing that I want to note here is the LLM is not actually doing the medical diagnosing. Okay? They've got a specific model, a neural network that they've trained for earlier purposes that is doing the medical diagnosis. Like, there's a separate model that is, like, taking in the brainwaves and being like, hey, is this epilepsy?

52:34 Is this not? Blah, blah, blah. The LLM is being used to interface with the patient because that's the whole point of the language, right? It's like, as a patient, I want to use natural language. To interact with whatever data modality. This is so crazy. Right? So it's really agent-based. Like, there's the agent that's the LLM.

52:54 Then there's the agent that's, like, has nothing to do with English. Right. But just trying to classify brainwaves. Mate. Yeah, I think it's really cool. So you can imagine, right? Like the deep learning model detects anomalous brain synchrony. The AI then texts the patient, hey, did you just experience a seizure?

53:18 This is so crazy. And then the patient is like, no, that was just me like jumping up and down. Or actually that felt weird. That might've been a seizure. And now it's, again, you're taking the cycle of training goes from, Patient collects data over X period of time. It goes to the physician. The physician spends X period of time with some assistant or someone else that's tagging manually.

53:43 The physician doesn't even know the personal experience of the patient who's experiencing the seizure. So they're just looking at it. And based on our past things, this is what normally – It actually – it has both the combination of bringing the patient into their own – giving them a sense of control in their own experience but it also has a very functional benefit to improving the detection because the part of the challenge is always patient translating their experience to the physician and then like that back and forth which there's it's a very noisy signal the noise there is also not very right right right right I mean, it's absolutely incredible.

54:24 I think we're standing at a new dawn of medical technology. We're moving away from passive hardware that treats patients as biological substrates and towards semi-autonomous conversational partners that help us navigate our own biology. Like, imagine you have agency over... your own kind of outcome in some sense.

54:48 And I want to make a small note here. There's, I know there's a lot of people have a lot of feelings about specific providers of foundational models. Yeah. I think part of the point here is, is, This could be done on an open source mini model. It's not open AI mini specific. It happens to be what they used in this case.

55:09 The large language model is doing a very small amount. Right. Honestly. Right. Right. But I just wanted to make that distinction that now that we understand it can be done, this can be extrapolated to whatever provider you prefer. Yes, exactly. Yeah. It's incredible. I want to take a brief note about how they actually did it because even though this is a rundown episode, I want to get into some of the— We got to hit two hours.

55:35 You know how it is. Yeah. Yeah. I'm sorry, guys. But these Penn researchers, they recruited 13 adults in the epilepsy monitoring unit. The setting here matters because it's an epilepsy monitoring unit, which is one of these things where you go in— High quality recording. There's clinical supervision.

55:53 There's detailed event review. And also sometimes they even take reduced medication to induce more and more epilepsy and more and more seizures to actually get more and more data. So that matters a little bit. Right? Sure. The platform, actually, there's three parts. There's a stream of EEG and clinical information.

56:13 And then there's software that extracts whatever the useful measurements are. And then that goes into a phone interface that then connects to the patients. And the patients can be just in the epilepsy monitoring unit with their phone talking to this agent. That is also talking to like the models that are interpreting data from his brain.

56:33 This is so good. Dude, it's absolutely incredible. It's clever too. Yeah. It's so obvious. Like it's so obvious in retrospect. Yeah. And there's so many multiple modalities here. Like there's a system that estimates sleep stages so that that is not confounding the algorithm. There's a system that estimates medication because there's medication information that in the patient's notes, right?

56:58 That'll tell you, okay, the patient took the medication at this time, and then there's a model that tells you how the medication is deteriorating with time, and that's going to inform this ensemble of machine learning Models. We have all these inputs into the model that gets to decide, do I set the alarm on or off?

57:20 And you also now have the feedback of the patient, which is arguably one of the most important inputs into that, let's call it detection system. And to be able to have them all connected and streaming real time. Yeah, it's absolutely very, very cool. And the result is a structured collection of measurements, right?

57:39 You can have these measurements be queried, plotted, and discussed. And the chatbot and the seizure detector are different models that are doing different jobs, but they're all coming together. for the seamless sort of loop. The patient has a brain interface. It goes through all of the stuff, but the patient only interacts with it using natural language, which means that, you know, it's kind of a democratized system.

58:05 The patient doesn't have to understand what the machine learning nonsense is happening, right? The chatbot is going to go ahead and ask it questions and get the feedback, right? This is something we talked about at the end of our Navier Stokes episode, which is when we talk about these concepts of AI, we have to think about them as systems, not as individual component parts.

58:25 Because this is a perfect example of why that framework of how to think about these things matters. Because they can be individual modules with discrete functions that can work in collaboration towards an end goal. And they're also not just LLMs. There's a whole variety. The word AI has a whole ecosystem of meaning and capability.

58:46 Yeah. And this is an example of combining two things, the machine learning and the LLM. Exactly. For an outcome that has clinical relevance. Yeah. And actually, to piggyback back off of that, you can ask, like, what exactly is this machine? There's a bunch of agents, right? What exactly is happening?

59:05 Well, let's suppose a patient asks how much sleep they got last night and how good it was. Well, they've published kind of a flowchart of how things happen. You receive the patient message. Now, if the message is unsafe, then you don't do anything. You maybe alert the physician who's in charge and things like that.

59:25 But if it's safe, then you have a message router. Yes. And the router asks, okay, is this a general chat? Or do I need to actually do a data query? Yep. Right? If it's a data query... Then it actually instantiates a SQL query. Yeah. Like there's an agent that goes, okay, I'm going to go into the SQL database of all of the data on this patient and like instantiate, get that bit of EEG or whatever brainwave.

59:56 And then now instantiate another thing that asks maybe a clarifying question that asks... What are the tools that I need to understand this data? And then you go back and send the AI message through the chat bot to the phone. Right? It's able to actually look at the raw data, analyze the raw data, and...

1:00:17 Kind of understand what to do next. And from this flowchart I'm seeing there's actually also two safety gates. One at the front end to moderate the patient message if it's unsafe. And then once all of the processing is done, like you just mentioned, it also moderates the AI response if it's unsafe prior to getting to the patient.

1:00:36 So, again, they've thought about both. the front and back end from a safety perspective. And, you know, medical researchers, clinical researchers are some of the most anal retentive about details like this in ways that maybe people wish the frontier labs would be more... conscientious. 100% dude 100% and one last thing that I want to talk about is the fact that the system can ask questions back so it's not just that the patient asks questions for example suppose it has detected a possible seizure, it sends the patient an alert asking hey did you consciously experience a seizure sometimes you have these kinds of mini seizures where you don't really experience anything and sometimes you do In any case, the patient answers yes, no, or uncertain.

1:01:32 And in machine learning, that means that you've established a label to the EEG that the model had triggered as something that possibly could use a label. Now, usually you only want high-quality labels, but actually, even in the case of these low-quality, weak labels, you can do a bit of learning. called distillation.

1:02:00 Here's the idea. You've got a pre-trained model, right? Let's say you have a bunch of data on epilepsy, and you've got EEG recordings, and you know when the epileptic seizures happened. So you've got this model that is pre-trained on thousands of hours of epileptic data. And now you want to sort of fine-tune it to this specific patient, right?

1:02:21 Because this specific patient's brain is different, and the way that they experience seizures is different, and so on and so forth. You can use this technique called distillation. That's going to do two things. One, it's going to try to get it right for that specific patient, but at the same time, it's going to penalize changing up what you already know, right?

1:02:43 There's a two-fold principle. thing that you want to do. You want to gear towards what the patient specifically is experiencing, but you don't want to forget all of the stuff that you've learned from general epilepsy data. And that's this idea of model distillation. It's a penalty for changing too much on the basis of limited evidence, because you have all this hours of evidence, and now you're getting, what, maybe a few hours of evidence on this patient.

1:03:12 You can't totally upend your world model based on everything that you've sort of realized beforehand, right? And now the question is, do the models actually do better? In fact, they do. The models do better from this closed loop arrangement between the patient and the training, right? On the left-hand side, what we're seeing is from baseline, that's called the area under the PRC curve, the area under the precision recall curve.

1:03:46 It's a... It's effectively like how well is the machine learning model doing in terms of precision and recall. You can see that the baseline model is highlighted in green and the fine tuning steadily increases. The F1 score, again, another metric for how well precision and recall is working, is increasing.

1:04:08 The false alarm per hour is decreasing. And that's crucially very important. We are on the foothills of continual learning in clinical application. And continual learning is a very specific thing. I know that this is not literally continual learning. But that's why I said foothills. Because part of the challenge with all of these model architectures is they're all pre-trained.

1:04:36 And so they can't learn from new information is the argument that's said a lot. And even in this case, right, they're scaffolding off of pre-trained models. The LLM is the ChatGPT4O Mini. That's a pre-trained model. The seizure detector is a pre-trained model. But especially because of the complexity of biological variation and individualized medicine, you are going to kind of necessarily need this foothills of continual learning capability, which is can we allow this distillation to happen on a per-patient basis in a way that takes the benefit of all of the massive amounts of data but can correctly weight how much to take the influence of the individual patient in future decision-making.

1:05:27 It's a very narrow thread to pull, but the... The efficacy that we've seen in this study in particular is very encouraging in, again, being able to now kind of take something out of the box and make it evolutionary in tandem with a patient. Mm-hmm. Yeah, I think it's a really, really cool study. It's a small study.

1:05:56 It's short, and it was conducted at only one center. The participants, they need sufficient English, cognitive capacity, and smartphone ability, I guess. So you could think of that as a kind of limitation, but I think a convincing next study would be to measure the outcomes that matter to patients For example, like the seizure burden, the quality of life, anxiety, the amount of emergency visits, the clinical workload.

1:06:27 I think the fact that we're seeing this, though, is very encouraging to me because, you know, we hear a lot of AI doom and gloom. You know, this is some good news. This is what was promised to me. okay yep when when when ai came in they were like we're gonna make medicine better we're gonna make your health better that's what they said this is what was promised this is what was promised right so we're starting to see it slowly come out and it's not just the frontier lab companies that are using these technologies um and a fascinating second story and not to burden you with another ai related story but unfortunately uh When we were in our 20s, the phrase was software is eating the world.

1:07:16 And now in our 30s, AI is eating the world. And so there are just a multitude of areas, especially in science research, where it comes up. And we are going to now move into our last story, which is going to be a little bit of a zoom in, but also a zoom out of what we mean by this whole idea of automated science research or AI going end to end from hypothesis to clinical outcome.

1:07:45 And I think this will be an interesting, there's a story that came out today that I think will give us a nice on-ramp to discuss what is the current state of play in what is a pipe dream at this point around AI scientific discovery. So earlier today, we'd planned this episode and then this story dropped.

1:08:04 I was like, oh, this will be interesting. And so this is a last minute edition about the other frontier lab that we did not talk about in our last story, which is OpenAI. This is an anthropic story. about scorpion binders and the path from AI to real-world science. And so what are we talking about? We've seen all these stories, like we mentioned every week, about AI is going to change science and make all these new discoveries.

1:08:34 And today, Anthropic published this research article, How Claude Designed Protein Binders Against Scorpion Neurotoxins. And it was a collaboration with Citrini Research and the laboratories AdaptiBio and Twist Bioscience. The idea is they're trying to design proteins for these dangerous venoms that are And scorpion venom potentially leading to a treatment.

1:09:02 So interestingly, there are some antivenoms that exist, but they're very specific to different species. They're species specific. And even with any individual species, there's also some regional things. And so they're very complex, as all the biology we talk about. It's very complicated. And so one of the holy grails is this idea of discovering a universal antivenom.

1:09:28 Uh, obviously. Yeah. Yeah. Yeah. Cause then you currently see the anti-venoms that have been created are not in underserved areas or areas with less of this clinical and medical infrastructure. Yep. Um, and so, you know, the challenge becomes when you don't have these anti-venoms for, for scorpion, stings, the current clinical solution is this idea of aggressive, supportive stabilization, right?

1:09:58 You're basically waiting for the body to naturally metabolize, but you have to keep the person alive while that's like happening, right? That's the stabilization part. You just like keep them alive. Right. And so that can be done... By supporting, for example, artificial air, like if they can't breathe and things like that.

1:10:16 And they also have this idea of utilizing secondary drugs to counteract the specific damage that the venom is causing, which in some cases can work and it impacts the nervous system directly. So this story was immediately all over the place. Okay. Because, again, marketing. Marketing. And so the headline read they tested 211 designs in the lab and 17 met their binding criteria for four toxins.

1:10:43 So like what is – I'm reading this. I'm like what is this? What does this actually mean? Crucially, they've not shown that this can actually neutralize the toxin. protect an animal or actually help a human patient. And they explicitly made that clear that they did not create an, an anti-venom, but this is sort of this incremental stepping stone process that we've talked about previously.

1:11:11 Another small detail point is none of the candidates also could bind to multiple toxins. So in the journey for a universal anti-venom, That's not – it's still not – but it was still an interesting design challenge where they were able to find the ability for the protein to bind, which is still interesting and impressive for the system to have hypothesized and designed on its own.

1:11:37 So the first thing I want to look at here is like what did this experiment actually establish? And so they have this video overview we'll walk through really quick. So this was the scorpion from a Mexican scorpion, a scorpion from Mexico, that the CH2 is the toxin. And so we've talked about sodium ion channels that allow things to transition.

1:12:02 And so this toxin comes in and activates so that it causes the ion channels to become overactive. Okay. Right. And so that's obviously a problem. So there are current antibodies. This LR antibody is actually a lab-engineered antibody that we know can bind to it to prevent this from happening. So Claude designed this CN2BI protein that is able to bind to the venom, but it conflicts, what it looks like conflicts with where the antibody is able to connect.

1:12:37 But not all antibodies, this RU1 antibody, is able to connect. the anthropic designed one does not have a confliction with. And what's interesting is what they were basically trying to identify is with LR, the initial one we saw, can we bind to the same location for a known antibody that we know is is able to combat the venom itself.

1:12:57 So we're going to repeat this one more time, and I want to come back to this really quick. So they use two antibodies, the LR and the RU1. They bind to different locations on this venom itself. This binding test was trying to figure out, can we design a protein that can bind to the same location as a known working antibody?

1:13:19 And so it was a competition experiment here. So now, again, we see the claw-designed protein is able to connect. When the LR antibody is there, it's saying, It's in conflict. Yeah, it's in conflict. It's not able to. So it's kind of in the same location. Yeah. But when R1 is there, it's like it's a little bit, but it's only at 16%.

1:13:40 Yeah, yeah, yeah, yeah, yeah. And so the idea, what they're trying to say here is that, look, we can kind of have a binding in a similar location to something we know that works, but that does not mean it neutralizes. The toxin. We're just saying that it can bind. Right? Am I doing okay so far? Yeah.

1:13:57 Yeah. I think that makes sense. So LR came from an earlier antivenom research from UNAM in Mexico. They engineered a human-derived antibody fragment that showed that it protected in mice. And it's a laboratory reference. A lot of times with these anti-venoms, what they end up doing is they'll inject a horse with the venom, then take the antibodies from the horse, and then do some voodoo magic.

1:14:20 Oh, yeah, yeah, yeah. And then give it to the human. Okay, fair enough. And then the RU1 is similar that just attaches and binds to a different location. So they wanted to see how it worked in both cases. The idea here is like neutralization is the idea of capping the ability of the venom to— create that overactivation.

1:14:41 Yeah. And ultimately, like, that's a solution here. So the headlines and the analysis are always related. So what we're saying that they've been able to do is they designed a bunch of options for what could potentially be successful at binding to a similar location as the LR antibody, 211 of them. They found, as a result of that, that they were able to get 17 to meet the binding criteria across four different toxin types.

1:15:14 So that's interesting. Yeah, it's like generalized across four different... Toxins, that's pretty good. That's interesting. Okay. And so – but this is still not helping a patient, right? Because we've not even gotten to the point of being able to actually confirm that it neutralizes the toxin. So I wanted to take a step back and say what – are the steps between this binding mechanism, which was the result of this and a medicine to a patient.

1:15:47 Yeah. And, uh, so Trini actually put this roadmap out that we can look at. That's kind of how they sort of define it. So step one, Right. Is does it neutralize the toxin? We haven't even gotten to step one yet. We found a binding mechanism. And so now step one is making sure that it can actually neutralize the toxin.

1:16:05 Because it could bind, but not necessarily neutralize it. Oh, interesting. You know what I'm saying? I see. Yeah. And so there's a difference between those two things. We know it can connect, but we don't know that it protects the actual nerve function itself. So then number two, does it cover the relevant venoms that we're trying to address?

1:16:24 Mm-hmm. Number three, does it work in a body, in this case in animal testing? Then we get to stage number four, which is, is it ready from a safety perspective for human testing? And then five, does it actually help patients once it gets through that developmental process? They estimate that whole pipeline for this specific use case to be about four to seven years plus.

1:16:47 Yeah. And all we're talking about is not even yet having solved step one. But this is the pipeline that all these AI is going to cure X, Y, and Z. This is the pipeline they kind of have to go through. And they've identified a few things in their chart about where they think AI is going to be the most impactful.

1:17:08 So improving the neutralization of the venom itself, they're like, I think we think we'll have moderate impact. in being able to then cover, once they know it can neutralize, they think it has the highest impact of being able to then generalize it across all these different classes of venom, this universal anti-venom.

1:17:27 They also think that in that first stage of animal testing that it can have high impact as it relates to animal testing. But when you get to preparing for human trials and the trials and acceptance process, it goes to moderate to limited, which is obviously very understandable. I think the point here that is sort of being brought up is So there's a real world aspect to this stuff that is just naturally going to be slower unless and until some of the processes, the regulatory environment and the way in which we do those downstream processes fundamentally changes.

1:17:58 So that's kind of part one of what this OpenAI announcement is. And it's interesting because it dovetails with the wet lab story we talked about a few weeks ago where everyone's like, oh, Anthropic has a wet lab. Yeah. You know, but it was where the AI was designing and the human scientists were in the lab actually conducting the entire process.

1:18:24 Yeah. Itself. So I want to sort of briefly talk about the timing. Because the promise here is that this is going to accelerate. It's going to happen very quickly. So this next chart kind of looks at how long did it take them to create the design candidates? And how long did it take them to test those candidates in a laboratory with the two partners we mentioned earlier?

1:18:49 They were able to get the first 180 designs in 64 hours, which is quite quick. That's quite quick. Quite fast. Yeah. The easiest part of the process, maybe. But it's still quite fast. Yeah, two and a half days, though, is quite fast for 180 designs. It's quite quick. Yeah. The first binding results arrived after 39 days, which is from the beginning of when they started the design process, not after they got the design process.

1:19:16 So about five and a half weeks to get to this point where we have the designs done, we've tested them in the lab, and we have these... There's candidates in between there. There's synthesis. There's protein production. There's getting the reagent, scheduling the work, actually measuring what happens.

1:19:33 There's quite a bit. Yeah, there's quite a bit that hasn't been optimized for... An AI lab ecosystem. Let's say that. That's exactly right. Like this is – there's so many parts of this that are still sort of not even optimized, I would say. Just like catering to human timescales, like PhD student timescales.

1:19:58 Right. Right. And I think part of what we're trying to talk about in this story is looking at each of the stages of this process and where is – AI, but also the combination of automation and robotics, because ultimately you need to interact with the real world. That's right. And so this is where sort of the AI and robotics together fit in that process.

1:20:24 So I wanted to, again, I was trying to. help myself understand what are we looking at here and so i created this map of how ai is going to like meet the physical lab and where the current state of play is and some of the sort of this is not a uh complete map of every player in the space but it's just to give a kind of representative example so it kind of breaks down into these three stages there's the design and prioritize so we just talked about that in this anthropic story with the scorpion anti-venom so i'm The AI...

1:20:58 Is going to design and prioritize. Correct. In this model. We may still define the question. Right. Scorpion, anti-venom. Right. Be like, give me an anti-venom to all of Scorpions. Right. And I think the original prompt for this was about 35 pages. Oh, wow. That they fed. They had, I think, 300 GPUs.

1:21:19 It cost a total of about $40,000 to $60,000 to run this design and prioritize process. Mm-hmm. Which is not crazy. No. But it's a little expensive, but it's, you know, not crazy. No, it's not the usual stuff that I ask Claude and ChatGPT to do. Right. My prompts are not 35 pages. Right. And if one of those ultimately became something that was able to get to the end of the process, that would be fantastic.

1:21:45 Okay, so a variety of the frontier labs are doing that design and prioritize, but we're talking about this anthropic story in combination with Citrini. So then sort of stage two is this idea of physical automation. Not necessarily AI, anything. We have a couple of players in here, MIT, Emerald, Cloud Lab, Stratios, and Ginkgo.

1:22:06 They're trying to now take the physical process that humans do in a lab and have it be run by a robotic automation. Obviously, it's a connective tissue because you can take part one and then send it to part two. Now, there are some companies that are trying to do... The full closed loop cycle, going back to our previous story, of the learn and repeat.

1:22:30 This is Lila Science Madeira, where it is the combination of both the AI designing, prioritizing, the robotic lab executing, getting the measurements, feeding that back directly into the AI to then decide what to change and then retry. Human oversight, but – Kind of a closed-loop cycle. These processes are also using a variety of shared infrastructure.

1:22:56 For example, Carnegie Mellon is providing some of this shared infrastructure as well as Slack, the national lab. Stanford Linear Accelerator Lab. Yep, is giving some of the infrastructure around these pieces. So that's where we're trying to go is where Lila Sciences is effectively saying we are right now.

1:23:19 And so the first step that the frontier labs think they can take is design and prioritize because it does not require this connection to the downstream robotics and automation piece, which is not their bread and butter. So now let's transition. And this is something we've talked about a lot with the methods is this equipment and measurement.

1:23:44 We talk a lot about tools on the pod. We talk a lot about being precise in our measurement and our methods. And MIT actually gives us a great example of this automating step. This is not necessarily new as in new in the last couple of years. It's been happening for some time. But understanding how this now connects to these AI workflows is a different way to start to think about these steps.

1:24:08 So researchers at MIT working in collaboration with Nokia Bell Labs and Arizona State built a robotic system that can assemble and align models. an optical experiment. And this was actually highlighted recently in September. This research, always talk about the funders, supported by National Science Foundation, Army Research Office, and some other funders, private funders, including Shell.

1:24:32 And this was one of their demonstration videos where they have a robotic arm that in 50 steps in 30 minutes can set up these lenses for this optical experiment, which we've talked about previously. And not only can it set up the lenses, because it's constantly taking measurements, it can also make a change to fix something when a misalignment arises as well.

1:25:02 So it's both things. This is crazy. Human-designed setup. Yeah. But it can set up and iterate when there's a misalignment. I mean, I mean... So already this is giving me flashbacks to my PhD days. Like, for example, the table that they're on, you see all those holes, those little screws that you can put in and you can fasten like mirrors and lenses.

1:25:28 Um, Any type of tabletop AMO experiment, meaning atomic molecular optical, like if you want to do molecular traps, a lot of quantum computing experiments are literally just this, which is establishing lenses everywhere and then... You know, aligning them to the point where you can suspend atoms in a laser or you can, you know, isolate a single atom to do some type of spectroscopy measurement.

1:26:00 This reminds me of the MIT double slit experiment episode. Yeah, this is all that is. Like a lot of these AMO experiments are literally just this. And just so, you know, I want to ask you briefly, like, the sensitivity and complexity. Because someone who's watching this might be like, oh, this looks pretty simple.

1:26:21 No, this is incredible. Like, if they're able to align a laser and, like, make sure that the transition frequency is exactly aligned to whatever, you know, system that they're trying to control. Like, that's pretty incredible. Like, that's... 50 minutes, 30 minutes? That's, like, half the work of these AMO PhD students.

1:26:46 Like, my roommates, you know, like Pratik Puri. Yeah, yeah, yeah. I don't mean to call you out, but this is half your job. Like, you know? Like, so... um elliot borah another friend of the pod yes like this is something that he does a lot of i mean obviously there's a lot more that's in his head sure right there's a lot more stuff that they could do perhaps if they didn't have to spend their time doing that aligning lenses and lasers And this was a demonstration going back to our three stages of mainly this idea of execution and measurement.

1:27:24 Because AI can design all these things, but we still have to be able to execute and measure. And this is for optical experiments, not biological experiments. experiments it's a different category but i'm just trying to give the listeners a sense of where these component parts are that people are ultimately trying to connect in order to get this closed loop end-to-end uh ai-based research cycle um you know but that step might can close the equipment control loop yeah as one component part of that Now, the next video that – this is sort of a promo video so that there's some history about how we've done these things for quite some time.

1:28:07 So this is from Emerald Cloud Lab and Stratios. And they kind of address a slightly different part of this execution, which is making physical laboratories – accessible through software, through the cloud, remotely. So I'm in one location. I need to do X, Y, or Z in another location. And the entirety of the process that I need to be done is remotely executed by automation and robotics.

1:28:38 But I have a software interface that I can interact with in real time. to execute. So like a lot of folks might go to the doctor and then they need to get, you know, a blood test and then they send it to Quest Diagnostics, right? So this idea that you have one person doing the question, one person doing the execution, and then it comes back to your physician to give you the result.

1:28:59 The difference here, again, with still some elements of human involvement for some of the component parts, many of the components are now directly This is, I think, a big area. This is crazy. I think this is very big. No, this is crazy. This is absolutely... And you get how long it takes to run each.

1:29:23 You get the real-time data back in the exact format you want. It's a wet lab that you can access through the cloud. Isn't that crazy? I mean, no, they're doing mass spec, I mean, or something, you know? This seems like stuff that would require specialized equipment, but there you go, like, right? No, I mean, that's actually incredible, right?

1:29:46 Because for the longest time, we've seen machine learning researchers do this. You just access GPUs from the cloud, AWS, and then you're doing machine learning experiments. Well, now you're actually doing real biological experiments. You've got pipettes that are automated. You've got... I think I saw a centrifuge that was automated, right?

1:30:09 And they did it correctly, unlike the Ryan Gosling movie. Look, it was one of the only flaws that I brought up from Hail Mary. If you were going to say it or miss it. But, like, they did it right. They did it right. Like, the AI knows to balance the centrifuge. Unlike Ryan Gosling. But no, this is incredible.

1:30:35 And what, again, this is not a promo for any particular entity. There are a variety of companies that are trying to do so. No, I'm just saying this is a new era in scientific. Correct. Yeah, so, I mean, one thing I want to understand is like, is Emerald Cloud Lab going to be a co-author? I wonder if...

1:30:57 Like, as a former researcher in, like, a PhD or a postdoc setting, it's like, I'm always like, who's going to be co-author on the paper if I were to use this data, right? Do they have some kind of thing in their contract being like, we got to be... Acknowledgements are fine. Co-author, it's a bit, you know, like, okay.

1:31:16 Maybe a footnote. Yeah. Maybe a footnote. And... One of the things that this brings up that I was kind of thinking about that's not necessarily as obvious because like the idea is like, oh, you know, a lot of the thoughts around these AI implementations is it's going to allow the rich to get richer. Which is true.

1:31:35 Yeah. Probably. In a lot of contexts, that's true. But what it does also do, like what – type of research teams that might not have access to these types of resources who are in underserved environments, what kind of questions can they ask? And it's the idea of AWS and all this cloud computing allowed people to build websites and do video.

1:32:01 And we can do this show because of cloud computing and distributed networking. This wouldn't exist without that. So it does bring up the question like what kind of research questions and what access that teams that otherwise would not have been able to access these kind of questions. high fidelity resources, particularly in my family who's worked in, you know, global health and neglected diseases, it becomes a very interesting democratization of the access to the types of tooling when you have this cloud accessibility.

1:32:32 And that was a non-obvious conclusion after kind of digging into this is like, it is a different paradigm that allows us to ask some different questions. Yeah. One last sort of note I want to bring up going to the third stage. Because we talked about design and how to prioritize what questions to ask.

1:32:51 We then looked at automation at two different levels, right? One where the humans are still involved in the lab, one that's cloud accessible, and So Lila Sciences is trying to connect the full loop, right? They had this catalyst study. The AI helped select compositions, automated the equipment, made the test materials, and the measurements informed subsequent choices.

1:33:15 So it was a very narrow question. Yeah, but it's closed loop. But it's closed loop in some sense. Okay. They evaluated almost 3,000 compositions related to hydrogen production. And this diagram here kind of shows how they look at that cycle and what they're ultimately trying to do in terms of this closed loop catalyst discovery.

1:33:36 And so – The goal here was trying to give a baseline lay of the land. There are more than just the frontier companies doing this. People are attacking different component parts of the process. Some people are trying to do the whole thing. Some people are just trying to do a module of the thing. And there is sort of this shared infrastructure.

1:34:02 So Carnegie Mellon has AI Science Foundry, which is trying to connect these cloud labs and scientific infrastructure to like a shared – this sort of shared distributed research infrastructure to make sure it's – evenly distributed and accessible for many people. Slack has this new Spire project, which is trying to connect their robotic sample handling, beam control, and real-time analysis.

1:34:24 And we actually have a reference video here. This is from... from Slack's Spire project. Now, it's just part of the execution and measurement infrastructure. It's not the completed future system. But again, I think that there are multiple national labs, universities, private companies that are all moving in this direction.

1:34:47 And when we go back to one of the reasons we started the show, the thesis was we're going to be in a Cambrian explosion of a whole variety of breakthroughs. And we want to be here having a front seat and bringing you along with us for what will happen. Again, there are very complex social, economic, political, and other issues as it relates to, quote, AI issues.

1:35:12 But on this show, we talk about the researchers who are using this in very specific ways that are ultimately looking to create opportunities to benefit us collectively. And it's important not to just say AI is bad in its entirety when there are really interesting and valuable use cases available. that are coming up.

1:35:36 Now, I will just end on this note of all the Doom scenarios expect there to be automated wet labs. Yeah. And then an accidental release. Yes. Obviously, none of the things we discussed are related to pathogens that are extremely deadly. But there are certainly a lot of consequences as it relates to the progress we've already made with...

1:36:02 cloud wet labs yeah and dude i mean it's it's uh it's the best of times it is the worst of times as charles dickens said right um and it really feels that way there's if we if we do nothing it's going to be both right and um the scales only need to tip one way or the other it's kind of like the chemistry noble prize right You only need a slight asymmetry, and then all of a sudden, it's very much not a great thing, right?

1:36:35 Or all of a sudden, it's a very good thing, right? So we want our homochirality to be positive. Positive in our direction. In our direction, right? Not in the A.I.S. direction. So I thought you'd find this, because we kind of touch on aspects of this, but as we continue in future episodes, this is going to continue to come up.

1:36:58 um and uh i successfully talked through a live molecular diagram yeah that was pretty good it's not bad yeah i'll get better i'm learning guys i'm trying to get better um but i i thought that was that was pretty cool i mean um the the laser stuff is it was crazy to me because i've been in those labs right um in the physics department at ucla like those tables are the bread and butter of every tabletop experiment.

1:37:34 So that was crazy. There was just a robot that could just hang out and do that, which was like half the job of the PhD. 30 minutes, 50 moves. Right. That's crazy. Yeah. It's not bad. Yeah. And no breaks. No. It could just work overnight kind of thing. You know? Yeah. It could adjust when there's an issue.

1:37:53 Yeah. Yeah. Yeah. It's not bad. That's pretty good. That's not bad. We are going to continue to bring you the frontier of science research globally, not just here in the U.S., but from all around the world. My name is Lester Nare. Joined as always by my co-host and our resident Ph.D. And I also just want to remind everybody.

1:38:17 Signal award-winning podcast. Yeah. Bronze. We're going for gold. We're going for gold, but it'll go like my point. There it is. Krishna got the point. And for those of you watching on Apple Podcasts on video for the first time, welcome. We know it's been a long time coming. We will see you all next week.