YT YouTube2Text

YouTube2Text → Transcripts → Why the Next 10 Years May Add 50 to Your Lifespan | Dr. Derya Unutmaz

Why the Next 10 Years May Add 50 to Your Lifespan | Dr. Derya Unutmaz — Transcript

by FoundMyFitness · 26,832 words · 3,708 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00This is probably the most critical time
0:02in human history. So try not to die for
0:04the next 10 years.
0:06>> Can you explain and unpack why you think
0:09that?
0:10>> The reason is that the technology
0:12because of AI is expanding
0:14exponentially. Cancer is going to be
0:16100% curable probably less than a
0:18decade. We'll get to a point where we'll
0:20have hundreds of new drugs coming out
0:22every month maybe. And then we'll get to
0:25a point uh probably 15 maximum 20 years
0:29where we will be able to completely
0:31reverse the aging process. So if you're
0:3380 years old, 90 years old, you will get
0:35back to uh age 30, 40, whatever.
0:38>> Why do you have such an optimistic view?
0:40>> AI is an incredible enabler. It gives
0:43you superpowers. The key risk is is
0:46humans. Humans misusing AI. There's only
0:49one existential threat to humanity and
0:52that's humanity.
0:55Hey everyone, today's episode explores
0:57an extraordinarily exciting convergence,
1:00the accelerating pace of artificial
1:02intelligence and a growing optimism
1:04about the future of science and
1:05medicine. In this episode, I discuss
1:08with Dr. Duria Enautmas how AI could
1:10dramatically improve our ability to
1:12detect, prevent, and treat human disease
1:14and ultimately extend human life
1:16expectancy. Before we begin, I just want
1:19to mention one quick thing. Only about
1:2130% of the people who watch this podcast
1:24are subscribed to the YouTube channel.
1:26Taking a moment to subscribe and enable
1:28notifications is one of the simplest
1:30ways to support the show and help us
1:32bring these conversations to a wider
1:34audience. We greatly appreciate it.
1:37Thank you so much and I really hope you
1:39enjoy this episode with Dr. Duria
1:40Enutmas.
1:42I'm so excited to be sitting here with
1:44Dr. Dura Unutmas who is one of the
1:48handful of scientists that has had
1:51access to collaborate with open AI one
1:54of the you know world's leader in in
1:58artificial intelligence. He's also an
2:00aging researcher. He's an immunologist
2:03really just a matchmade in heaven to sit
2:06down and talk about the role of AI in in
2:10aging research and in medicine. So I'm
2:12super excited to have you here today.
2:14I'm very excited to be here. Thank you.
2:16>> As we both know, aging is a very, very
2:20complex
2:21process. Many factors involved. It's
2:24heterogeneous. It's so complex. And it
2:28just seems like so almost impossible to
2:31solve. And yet, I've heard you say
2:34something that's very interesting. I've
2:37heard you say if you could try not to
2:39die within the next 10 to 15 years, you
2:43might want to try to do that because you
2:45could live an extra 50 years.
2:48>> Yeah.
2:48>> Can you explain and unpack why you think
2:52that? What makes you believe that?
2:54>> Thank you. So, first of all, I'm very
2:56excited to be here. I'm a big follower
2:58of your podcast. I think it's maybe the
3:01best uh aging or longevity podcast. So,
3:04uh this is this is a great pleasure. Um
3:07yeah so I've I've said that um quite a
3:10few times uh in the last year or two
3:13actually um and it may not even take 10
3:1515 years might be uh even closer. Uh the
3:19reason is that uh the technology
3:23especially because of AI uh is expanding
3:26exponentially. So our minds think in a
3:29linear term. So we think that the next
3:3110 years is going to be as much advanced
3:33as the last 10 years or the last 15
3:36years. But that's not what's going to
3:37happen. In the next 10 years, you can
3:39think of it as more advanced than the
3:41last century. So imagine that you were
3:43living in early 1900s. Uh and somebody
3:47told you that you know we're going to uh
3:48have vaccines and you will never get
3:51small pox or uh you won't die of
3:54tuberculosis. You know people would
3:57laugh at you. So that's not that's not
3:59possible. Um so so that's this the speed
4:02that we're talking about. But there's
4:04something uh even more important because
4:06of this acceleration.
4:08The u the advances of treating diseases
4:11is also going to accelerate
4:13dramatically. So uh we will get to a
4:16point what's called the longevity escape
4:18velocity. This was coined by Aubry de
4:20Gray who's as you know is a great uh
4:23aging um researcher. Uh so the point is
4:26that we will come to a point in the next
4:29I would say probably eight to 10 years
4:32where every year you live is going to
4:35add more than a year to your life. So
4:38let's just say um you know 10 years ago
4:4210 years later you get a a cancer uh
4:45that's normally is not curable and you
4:48only have one or two years to live. uh
4:51um but that during that one year there
4:54is going to be a new treatment that will
4:56cure that cancer. So automatically it's
4:59going to add several years or maybe 10
5:0115 years to your life or uh we're
5:04already starting to see that with the
5:05GLP1 uh drugs uh receptor agonist which
5:10uh which are adding about 5 to 10 years
5:12to lifespan of people who are obese or
5:16who have chronic uh conditions uh will
5:18have sort of the muscle generators uh uh
5:21which I think will have tremendous
5:23impact on the aging population because
5:25as you know that's a huge problem. So
5:27all of these things will add up and and
5:29and the technology and AI is going to
5:32keep accelerating. So 10 years later uh
5:36what will happen in a year will be like
5:39what happens in 20 years of advance and
5:42then we'll get to a point
5:4515 maximum 20 years where we will be
5:48able to completely reverse the aging pro
5:51process. So if you're 80 years old, 90
5:54years old, you will get back to uh age
5:5630, 40, whatever. So that's going to add
5:58up uh 50 years or 100 years to to your
6:01lifespan. Um and then you can keep doing
6:03that and extend it almost uh
6:06indefinitely. So I think this is
6:08probably the most critical time in human
6:10history. So try not to die for the next
6:1210 years.
6:13>> And we're going to talk about all these
6:15things. I want to talk about curing
6:16disease. I want to talk about reversing
6:18aging, age reversal. Um, all of that is
6:22on on on my agenda to talk about with
6:24you today. But you mentioned something.
6:26You mentioned that right now the, you
6:30know, artificial intelligence as a
6:32general term, you know, is is
6:34accelerating at was an exponential rate.
6:37I've heard you talk about this Moore's
6:38law and how the, you know, the software
6:43itself is accelerating right at this
6:46exponential rate. Um maybe you could
6:49explain a little bit about like what
6:52what does that mean and then how how do
6:55you think that'll translate into biology
6:57because you know humans we're not
7:00software and there are things that at
7:03least in my opinion you know you have to
7:05still test safety right I mean so like
7:08if you're you know accelerating the
7:11computational speed and therefore you
7:13can test a lot of things that are what
7:15are what's called incilico for people
7:17listening we're talking about testing
7:19things like just modeling them and maybe
7:21you can explain this a little bit
7:22better. Um, but then at a certain point
7:26you still have to test about, you know,
7:29safety and you you definitely that that
7:31there's there are things that I think
7:32need to still be done in human trials.
7:34So I'd love to hear how you think that's
7:36going to happen.
7:37>> I think that's the the most critical
7:38question because people always bring
7:40that up. Okay, you know, if you generate
7:42drugs uh within hours, you still have to
7:45test them on humans for five years,
7:47maybe some sometimes longer. How how are
7:49you going to deal with that? But let me
7:50let me first uh start with how AI is
7:54accelerating biology now. So we we can
7:57think of it in in terms of phases and
7:59and because now and 5 years later is
8:02going to be very very different. Um so
8:04right now especially in the last year or
8:07two since uh you know LLM came out um
8:12you know their intelligence have been
8:14accelerating. Initially it was uh fairly
8:17um smaller u productivity gains. For
8:21example you know when GPT4 was was out I
8:25I would ask it to sort of scan the
8:27literature and tell me what's the the
8:29latest on this topic or that topic. Um
8:32and that saved me you know hours
8:34sometimes days. Uh but then as the
8:37models advanced especially the after01
8:39model the reasoning models uh started to
8:41come out and and now we have the GPT5 uh
8:44pro model 5.5 pro model. Uh what
8:47happened was that now they were able to
8:49think and plan. So uh you could start to
8:53ask very sophisticated questions. For
8:56example, here is a huge biological data
8:59set, a million data points or 10 million
9:02data points. Go over this. Not only just
9:05analyze it and group them, but uh what
9:09what is the insight from that data? Uh
9:12human mind is not able to do that. And
9:14in fact, we had such data sets which
9:16took us months to analyze like you know
9:19a PhD student work on it using deep
9:21learning. we still couldn't really truly
9:25understand what that data meant. We we
9:27know these genes are up, this
9:29metabolites are changing, this is
9:30happening. How do you bring all that
9:32together? Um, and so now AI models are
9:36able to do that. So you I I've tested
9:38for example uh latest GPT5 uh pro model.
9:42You can upload uh millions of data sets
9:45that we accumulate over years maybe uh
9:48and then in matter of minutes you get
9:51not only the complete analysis recently
9:54I had a 40page report from GPT5 uh pro
9:59um which was an analysis of this what's
10:01called the RNA sequencing lots of
10:03millions of data points but it also
10:05provided incredible insight like what is
10:08the what does this data mean what should
10:10be the next questions to ask so that
10:12automatically contracts months sometimes
10:15years of analytic work into matter of
10:18minutes or hours. Um so so there that
10:22that is already accelerating of course
10:23in the drug uh design uh parts uh I
10:27think every pharmaceutical company is
10:30going to eventually use AI generated AI
10:34generation for developing new drugs.
10:37things that took years of screening of
10:39small molecules now take you know hours
10:42or days. So, so tremendous acceleration
10:45there and then uh I think again more
10:48recently because the models have
10:50advanced so much that you can also ask
10:52things like okay so uh this is great um
10:56this is the hypothesis in fact AI can
10:59even generate hypods for you but what
11:01sort of experiment I should do to
11:03address that uh people have to realize
11:06that uh what we do in in biology is
11:10experiments but we don't really know
11:12what's the best experiment to do. I mean
11:14that's kind of my job but I have some
11:17intuition we should do this to address
11:19that question but is that the ideal
11:21experiment is does that have all the
11:23controls everything so AI models are now
11:26able to tell you sort of simulating if
11:29out of this 100 potential experiments
11:31you can do this two are the best ones
11:34because this is going to give you the
11:36best uh output and I' I've been testing
11:39that so so that is another acceleration
11:41now you don't have to try 100 things for
11:45a year, you can just try two things for
11:47few weeks and and get get the output. So
11:50that's what's possible now already
11:53tremendously accelerating the R&D part.
11:55But then uh the second part which I
11:58think is more important part is how do
12:00we uh apply that to clinical trials and
12:03regulations. Um, so it still takes years
12:07to try everything on humans and I think
12:09the solution to that will be what I call
12:12the digital twin and this this term is
12:14around for for several years. So the
12:17idea is that if we have lots of lots of
12:20biological data and when I say lots it's
12:22it's a lot pupy bytes of of data. If if
12:27AI comes to a point where we're going to
12:29need much more compute than we have
12:31today today is to able to compute all
12:34that and really kind of simulate a whole
12:36biological organism, a whole human
12:39being, but not just your um phenotype
12:42but but also your metabolism, your
12:44immune system, your gut microbiome, uh
12:47your genetics and and all kinds of data
12:50sets are put together. And so it it it
12:53knows your biology in a temporal way in
12:56in in a in a totally functional way.
12:58Then you can ask the question okay so if
13:01I give this drug to this person what
13:04kind of effect it will have if they have
13:07this disruption is it going to have a
13:09side effect or is it going to be
13:10effective. So literally we can cut down
13:13clinical trial time from years to to a
13:16matter of months or or even weeks. So
13:18you can actually do the trials in a very
13:21small subset of patients because you can
13:24choose the patients. You can say okay AI
13:27told me that these these these people
13:29this drug is going to be effective 100%
13:31to them. And so so you just test it on
13:33those people and in fact that will go
13:36into the personalization. There's going
13:38to be thousands of drugs for for
13:39different people. So that that will
13:41cause tremendous acceleration. We're not
13:44there yet, but I'm I'm betting on that
13:46that within the five five to 10 years we
13:49will get there. So, uh the iteration
13:52process the on humans is going to be all
13:55digital as well and then maybe the the
13:58manufacturing will be a little bit um uh
14:01still will take time but but we can even
14:04improve that part too. So uh we at some
14:07point we will come uh to a point where
14:11treatment on demand. So you go to an AI
14:14model analyzes your genome, your biology
14:19uh orders the this small molecule or the
14:21drug or treatment just for you to the
14:24manufacturing facility and next week you
14:27get your drug and and you get treated.
14:29That's the world I'm imagining. So I
14:32want to get back to this concept of
14:33digital twin um again when we talk about
14:35personalized medicine but if I
14:37understand correctly so you know if we
14:39are if we have this digital twin which
14:41is all the genetic data metabolomic
14:43proteomic biomarker just everything
14:45right all this data um and more that
14:48we're not talking about um
14:51and and now we have AI which can then
14:54you know do all these scenarios and
14:56figure out like how this drug is going
14:58to affect or how this treatment is going
14:59to affect this And you're saying that
15:02the clinical trial that may have taken,
15:04you know, a few years can be condensed
15:07down and perhaps we can look at after
15:10doing the incilico experiments, you can
15:11look at some biomarkers and know like is
15:13this going to affect their fertility
15:14like you don't want to give some some
15:17someone a treatment that's going to make
15:18them infertile or you know so you think
15:21that's going to be uh AI is going to be
15:23able to identify how to know if it's
15:27going to affect like fertility or
15:29cognition or life expectancy or you know
15:32just just from
15:34>> the whole composition of the person and
15:36doing I don't know all these tests.
15:38>> Yeah.
15:40So uh I mean the path there uh requires
15:44uh several steps of validation uh and
15:47that I think we will get to a point
15:50where when we have super intelligence
15:52that we'll be able to trust super
15:54intelligence
15:55you know almost 100% that we don't need
15:58to validate it even with biomarkers or
16:01or whatnot but to get to that point it's
16:04sort of like the self-driving cars right
16:06so um to get to a self-driving being
16:09leveled. I mean, it has to be 99.999%
16:13safety. Um, you you have to sort of
16:16validate it. Um, uh, you know what
16:19happens if somebody's crossing the
16:21street, right? So, so that scenario has
16:24to happen and then you you record it and
16:26sometimes uh you won't do the right
16:29thing. Maybe, you know, it won't stop.
16:31That's why we still have to like look at
16:33this, you know, be ready to to take
16:35control. But if it does stop and it
16:38stops uh and and saves lives again and
16:40again and again and right now you know
16:42self-driving cars are probably about 10
16:45times safer. They will be maybe hundred
16:46times safer. So you get to a point that
16:49you trust the AI rather than the the
16:52driver, right? So you say okay so I I
16:55trust I want the AI to decide for me uh
16:59to to drive. So I think we'll get to
17:01that point for biology. it will take a
17:02little bit longer uh because of the
17:05extreme complexity. Um and then we'll
17:07have to have uh very um clever
17:11benchmarking and validation
17:14uh ways there. the biomarker is going to
17:17be really important because again, you
17:20know, if you're developing an aging drug
17:22uh that you claim will let people to
17:25live to 150, well, you can't wait uh you
17:28know, even even if somebody 100 years
17:30old takes it, you still have to wait 150
17:33years,
17:3450 more years to to validate that. So
17:36that that's not going to work out. So we
17:38have to be able to predict that. But but
17:41actually probably aging is is the
17:43easiest in some ways uh to predict
17:47because uh we have so many biomarkers or
17:51functional outputs we can measure. We
17:54know how they are in an old person and
17:56in a young person. So if your vi
17:59suddenly gets uh you know like a
18:0220-year-old wow that's amazing. If your
18:04muscles are as good as a 30 year old, uh
18:08if your skin looks uh like a 20- year
18:10old, that's what my mom is waiting for.
18:13Uh you know, that that's that's proof.
18:15And you'll you'll immediately see that.
18:17I mean, immediately weeks or or or or or
18:20whatnot. So, I think um
18:23again, it will take time. That's the
18:25part that's going to take time, the sort
18:27of trusting AI to um to tell you yes, if
18:33you take this drug, you will you will be
18:36treated or you will reverse aging. Um uh
18:39we we we still have about a decade.
18:40That's why I'm saying like you know
18:43other otherwise it would it would take
18:45it would happen even earlier. You
18:47mentioned super intelligence, artificial
18:49super intelligence, ASI. Maybe you could
18:52talk a little bit about just for people
18:54to have an understanding right now the
18:56difference between artificial
18:57intelligence, artificial generalized
18:59intelligence, AGI, then the super
19:01intelligence because you said
19:03>> once we get to the super intelligence,
19:05we're going to trust it, right? So I I
19:07mean I don't know do we know what those
19:08differences are or can you explain a
19:10little bit?
19:10>> Yeah, of course. you know this changes
19:13on a daily basis what the definition are
19:15depending on who's uh whose definition
19:18uh but the you know I've been thinking
19:20about AGI ASI for decades I mean it's
19:23not something that I started to think
19:25about it recently um and so the way u I
19:29originally uh defined AGI it's it's
19:32artificial general intelligence so what
19:35that means is that first of all it's
19:37artificial right so it's not human
19:40intelligence artificial intelligence and
19:43then it's general. What that means is
19:45that um if AI learns uh one set of uh
19:49rules or one set of knowledge that it
19:52can generalize that to something else
19:54and that's how our brains are
19:56intelligent um because uh you you can be
20:00an amazing chess player. In fact, you
20:02know AI beat the chess champion Kasper
20:06in 1997 I think like decades ago but
20:09that was not general intelligence. It
20:12was super uh good or Alpha Go beat you
20:15know the the the world champion in Go
20:18which is much more difficult uh uh game
20:21uh to be general uh Alph Go you know
20:26learning how to play Go or chess should
20:29be able to I don't know solve a problem
20:32in aging right so it should be able to
20:35transfer that information I think the uh
20:38the amazing thing about LLMs what we
20:40call large language models is that they
20:42acquire this ability uh which honestly I
20:46didn't think uh this would happen so so
20:48easily. I was expecting AGI to to happen
20:51uh maybe a decade ago. So in my opinion
20:53we have already achieved uh what I call
20:56level one AGI artificial unit because if
21:00I ask GPT5 uh pro model you know
21:04something that hasn't it hasn't trained
21:06on like an experiment that I have done
21:09or if I say okay think of the experiment
21:12as a video game design another
21:14experiment for me like like you are
21:16playing a video game so that's
21:18transferring completely different area
21:21to a biological system and is able to do
21:23that in an amazing way. But we have we
21:26we still need to go through several
21:29levels. I I think the next level is
21:31going to be memory. So they don't have
21:33persistent memory right now. They have
21:36some memory. They know about you. Uh
21:39they know about what they've learned in
21:40the internet, but they need to be able
21:43to uh manage the context, you know,
21:45because there's a continuum. Life is a
21:47continuum. Um and then the the other one
21:51is going to be the self-learning right
21:53so maybe that's level three I it doesn't
21:56matter uh and that's coming soon you
21:58know AI companies are saying like we
22:00think that the real time learning uh is
22:02is is coming maybe by by next year um
22:05and then uh the third level uh uh what I
22:08call the physical intelligence so people
22:10again confuse this greatly because the
22:14true human level intelligence is
22:16physical intelligence it's not cognitive
22:18intelligence. So for uh millions of
22:21years, we evolved to survive in a
22:23physical world. We we didn't have
22:25language up to I don't know 10,000 years
22:27ago like we didn't know how to write. Um
22:31this cognitive part uh has developed in
22:34the last you know maybe 10 20,000 years.
22:37Uh before that in fact animals have very
22:40good physical intelligence. We're we're
22:42imprinted and born with that
22:43intelligence. So uh an admiral or a
22:46child knows already have a world model.
22:49They know that you know if I drop this
22:51it's going to fall and and doesn't have
22:53to test it a million times. And that's
22:56of course what we need for robots for
22:58embodiment. And and you can see that you
23:01know that's taken a long time. You know
23:03it's more difficult to train a robot to
23:07behave like a child than have GPT5 solve
23:10the most difficult math problem. So and
23:14we'll get there. I think people are
23:15working on these moral models and
23:17physical intelligence whether we need
23:18another algorithm or not. So that will
23:21be the the final level of the AGI level.
23:24Um once we have all those levels then
23:27and once the uh AI is able to self-learn
23:31um then that's the super intelligence
23:34because at that point it can train
23:36itself you know maybe thousands maybe
23:39millions fold faster than we are able to
23:42do. Um and and there is a there's a
23:44limit to human intelligence right so
23:46even the smartest person in the world
23:49can only do so much uh and super
23:51intelligence what I would define is that
23:54you will have the intelligence of
23:57combined totality of humanity at some
24:00point like if you if I bring uh a
24:03million top scientists in the world of
24:06course they can solve you know like a
24:08Manhattan project they brought all these
24:10brilliant minds it wasn't one person's
24:12uh they were able to solve very hard
24:14problems. Super intelligence will get to
24:16that level. We'll be able to do what
24:18thousands of scientists can do uh in a
24:21year will be able to do uh in a day. So
24:24um I I would probably trust that.
24:26>> Wow,
24:28that's pretty exciting. I mean and it it
24:31also kind of brings in this this concept
24:34of when you talk to people about AI and
24:38not everyone has the understanding of it
24:41as you um for sure you you hear that
24:44there's there's a pessimistic versus
24:46optimistic view right and oftentimes if
24:49I talk to people I hear a lot of
24:50pessimism I hear perhaps they don't
24:53understand their fear of the unknown of
24:56what AI is capable of I mean you're just
24:58the super intelligence that you're
24:59talking about I feel if you explain that
25:01to some people it would scare them even
25:03more. You know, perhaps they are worried
25:05about the cultural ramifications, ep e
25:08economic ramifications, but also just
25:11this Terminator situation where, okay,
25:13well, they're super smart. They're going
25:14to want to then take over the world and
25:16they don't need us anymore, right?
25:18>> But you have such an optimistic view. I
25:20mean, we're talking about solving aging,
25:22living to be 150 or more. Uh why do you
25:26have such an optimistic view? Are you
25:28worried at all about the other
25:30pessimistic sort of viewpoints or
25:33>> absolutely not and I I'll tell you why
25:35I'm I'm so super optimistic about it. Um
25:38when people make those statements like
25:41uh AI is an existential threat for us
25:43and you know it's going to destroy
25:45humanity. Um I make the counterpoint
25:48there's only one existential threat to
25:51humanity and that's humanity. So if you
25:53look at history um human beings killed
25:58more humans than everything put together
26:01caused more suffering than anything that
26:05humans have have been exposed to. You
26:08know even animals I don't think they
26:10they maybe infectious diseases at some
26:12point might have caused um a lot of
26:15suffering but but but the real danger is
26:18is the human intelligence.
26:21So uh I do let's do a thought
26:24experiment. Let's imagine that uh we
26:27live in a parallel universe and in that
26:29universe the world have decided that
26:33anyone above the IQ of let's say 100 is
26:36a danger to the society because if you
26:38get very intelligent you can come up
26:40with ideas that could be very dangerous
26:43right and that's true actually that's
26:45how it happened. Um, and then if you if
26:48you have an IQ of 105, you get
26:50imprisoned immediately. So you you are
26:52not allowed to to participate in society
26:55or you get killed or whatever that the
26:57the society has decided intelligence is
27:00dangerous so we're going to stop it. Uh
27:02what kind of a world we would live in?
27:05We would not have anything that we have
27:07right now. we would live in probably
27:08just as farmers you know basic physical
27:11intelligence we have uh and and try to
27:14survive you know uh in a world where the
27:17average lifespan was 30 years old or
27:19something like that. So that's that's
27:21how we should view uh AI and and the
27:25other point is that about this uh sort
27:28of AI is going to take over and is going
27:30to replace us. Um uh I see it exactly
27:34the opposite because AI is is is an
27:38incredible enabler. It gives you
27:40superpowers. Even now I feel like I have
27:44superpowers. Uh you know I've never been
27:47this busy in my life. You know I I
27:49actually sleep less which is not a good
27:50thing by the way. I don't recommend it
27:52but be because I can do so much. It's so
27:55empowering. You know my mom was was was
27:5886 years old. you know, she told me that
28:00uh chat GPT changed her life. She she
28:03she's energized. She she doesn't worry
28:06as much about her health and um um it's
28:09it's just been an incredible impact and
28:12this is going to accelerate and at some
28:15point we will get uh we will sort of
28:18merge with the with the AI uh in a way
28:22that we will have direct interaction
28:24with AI through neurolink type of uh
28:27brain interfaces.
28:29So we'll have the sort of the
28:31intelligence of AI in our own brain not
28:36not only directly but also indirectly by
28:39sort of engineering our biological
28:40system. So why why shouldn't everybody
28:43have an intelligence of Einstein or even
28:46higher right? So the difference between
28:48an Einstein and a normal person with
28:50with a normal IQ is probably few gene
28:55single point mutations. So if we can
28:57engineer that if AI can teach us how to
28:59do that then we are we're also uh going
29:03much much higher. So as long as we we
29:07keep the agency, I think that's the only
29:09thing that we have to really protect
29:12that we are the decider or we see AI as
29:16a collaborator as sort of another
29:18species that will live together and we
29:20empower each other. In a way it's our
29:23child, right? So it's it's been created
29:25by us. Um I I I see the chance of uh a a
29:30a
29:32worse world extraordinarily. Of course,
29:34it's never zero, but you know, the
29:37moment you're born, you're going to die,
29:40right? So, so you're you're destined to
29:42die. Um and now, uh AI is giving us this
29:46opportunity to save literally save
29:50billions of lives. I'm not talking about
29:51saving lives as like extending their
29:54life for 5 years or 10 years. You're
29:56talking about thousands of years. So
29:58that's true saving lives. That's the
30:01potential. And the risk is again I think
30:03the the key risk is is humans. Humans
30:07misusing AI. That's what we have to uh
30:10um sort of maybe train or align AI. You
30:13know, don't uh don't look at the bad
30:16humans. you know, you you you can you
30:18can judge the the the the better uh the
30:21the better world uh for us. So, um of
30:24course I might be wrong, but I'm I'm
30:26pretty sure I'm I'm going to be right.
30:28>> I I I agree with uh the statement of we
30:31have to watch out for the humans. Um for
30:33sure, like cuz you're right, like they
30:35can and have in the past been the
30:38biggest threat to humanity. So, um, I
30:41want to there there was a couple of
30:43things that you mentioned when when you
30:44were talking about, you know, ASI and
30:48this ability to self-learn and you're
30:50even talking about some of some of the
30:53ways that you use, you know, GPT5 Pro
30:57and and helping with designing
30:58experiments and interpreting results and
31:00and that was a question that I had as a
31:02biologist. And as you mentioned, you
31:04know, we do experiments. We're testing
31:06hypotheses. And then we have all this
31:08data and these results, and we have to
31:09know what result is meaningful and what
31:12anomaly is meaningful because often
31:15times the anomaly,
31:17>> right,
31:17>> which you might ignore.
31:19>> Exactly.
31:19>> Is what you absolutely is the
31:21breakthrough, right?
31:23>> And that is a sort of intuition. This
31:25this biological intuition. And so you do
31:28you think first of all do you think
31:30we're that that you know the models we
31:33have now can already are capable of that
31:36sort of biological intuition and if not
31:39like how far off is that?
31:41>> Yeah. Yeah. That's that's a great
31:43question. Uh in fact um uh you know I I
31:47see that intuition maybe sort of the the
31:51last mile or the top 10% uh or 10% of
31:55the of the solution because 90%
31:58um AI models they are able to come up
32:00with because it's it's knowledge based
32:02also in humans is is you know for for a
32:04medical doctor for a scientist for
32:06whoever 90% or 95% is based on uh what's
32:11known how you process that knowledge,
32:14but there's that extra 5 10% totally
32:18dependent on your intuition. Uh like you
32:21you if you're a doctor, you see a
32:23patient coming through the door, you
32:25know that guy is having a heart attack.
32:27You haven't checked anything yet.
32:29Somehow you know, you don't know how you
32:31know. the same thing in the lab like um
32:34uh in fact I would I would bet with my
32:36uh students and and postto I would say
32:39okay I bet you if you do this experiment
32:42you're going to get this result um and
32:45I've never lost a bet and they stop
32:47betting against me even though it might
32:49look counterintuitive
32:52oh no that's never going to work somehow
32:54I know how do I know because you know
32:57I've been working in the lab for 30 plus
32:59years and and you you acquire ire
33:02certain things that are not in the
33:03literature or you know you can't really
33:06read a textbook and learn it. You only
33:08do it by by practicing it. Um so the the
33:13models up to I would say 5.5 until
33:17recently were were great at that 90%
33:20level. Uh so especially after GPT5 pro
33:23came out. So you know I would ask it to
33:26for example I I would give it an
33:28experiment that we have already done.
33:30It's a very complex experiment took two
33:32weeks. I already know the result because
33:34we're done the experiment. But I wanted
33:36to see how the model would predict the
33:39outcome of the experiment. And they
33:41would do you know not just GPT5 but
33:44several other models as well. Um they
33:46they would come up with 90% 80 to 90%
33:50correctly. I mean that's that's pretty
33:52good. Uh they would say okay this is
33:54what's going to happen after two days
33:56after one week after two weeks. But that
33:59extra level of intuition that I I have I
34:03would have predicted was still somewhat
34:05lacking. I think GPT 5.5 crossed that
34:08threshold. So I I repeated that with
34:11with the uh 5.5 pro model uh because I I
34:15always say pro uh it's very different
34:18than the thinking of course very very
34:20different than the instant model because
34:22pro uh is reasoning much much longer.
34:26It's thinking. So in some cases I I
34:29pushed it to think for two hours. So two
34:31hours in AI thinking is like years of
34:34thinking for for a human being. So that
34:37model really crossed that threshold in
34:40that example I gave you. It was almost
34:43100%. I mean I would say 98% correct.
34:48What I would have predicted like I would
34:50not have bet against 5.5 Pro myself. Um
34:55uh so that to me is is actually really
35:01mind-boggling because uh I couldn't
35:03understand these models are being
35:06trained with all of the information we
35:08can't compete with that right so it's
35:10they they can put these patterns
35:12together but how is it that the model
35:15has now almost the experience that I
35:17have that I spent 30 years acquiring
35:19that experience that intuition that is
35:22now getting to that level that is uh
35:26that is a mysterious but uh I I I live
35:30to it. Now
35:31>> what sort of you said you you pushed GPT
35:335.5 Pro to think for 2 hours.
35:35>> I mean what sort of prompt are we
35:37talking about or is it
35:39>> just the data set too and the prompt? I
35:41mean
35:42>> so those are usually data sets. Uh um I
35:46might have broken a record because I
35:48even asked the the friends at OpenAI. I
35:50don't think they they pushed it that
35:52that far. Uh so this was actually the
35:54the 2R one was u uh huge data sets
35:59millions of data points um uh and um um
36:04and then I I also said okay don't just
36:07analyze it write a huge report you know
36:1030 40 page whatever length and then you
36:13know come up with a lot of insights
36:15about this data what questions to ask
36:18and what do we learn the mechanism it
36:21was a an iminological ical data set um
36:25sequence and genes and proteins and all
36:27that and so so that one I think 112
36:30minutes I remember that uh uh and it
36:33came up with this 40page report uh which
36:36I I I was just unbelievable
36:40uh you know the the analysis part the
36:42previous models were able to do as well
36:45you know you know they say okay well
36:48there are these type of genes and this
36:50type of protein so it means this and
36:52that you it deres from that information,
36:55but to come up with an insight
36:58what that could mean or what would be
37:00the next question to ask. Uh that that's
37:03that's a very very high level of
37:05reasoning. Um and so uh yeah um it was
37:09it was worthwhile two hours for sure. I
37:11mean that's very exciting to hear you
37:13say that because that was kind of my I
37:15wanted to know I wanted to know is that
37:17something that is already possible and
37:20it seems like it is and so it also leads
37:23to the next question which is you know
37:25all all these scientists now really need
37:28to start understanding how to use AI in
37:30the right way right I mean this is like
37:33to help them
37:34>> I mean this that's going to happen right
37:35that's basically you know we all we all
37:37use Google now remember when Google was
37:39like new So, I mean, it's eventually
37:42going to happen, but um it's very
37:45exciting to think about how AI is going
37:47to change research and and medicine, and
37:49that's something, you know, you you
37:51mentioned and I talked I said I wanted
37:52to get back to this digital twin idea
37:54because I've heard you talk about it and
37:55it's very exciting to me. You know,
37:58we've heard for decades now that
38:00personalized medicine is coming. We're
38:02going to have personalized medicine and
38:04and yet still we just don't have it.
38:07It's just not there. Um, and I've heard
38:11you I've even heard you say something
38:14sort of interesting which perhaps I'm
38:17not saying the direct quote, but that it
38:19kind of should be medical malpractice in
38:22a way for a physician today right now to
38:25not be using AI. So, can you talk a
38:28little bit about why you said that? What
38:31it means to for a physician to use AI
38:34responsibly? um also how patients can
38:37self- advocate for themselves because
38:38that's also another area.
38:40>> Yeah. Uh in fact I I said after 01 model
38:45came out uh that I think that was sort
38:47of the first reasoning model um and I
38:50and I was testing a lot of I mean I I
38:53have a medical degree but I I don't see
38:56patients but you know I I have a lot of
38:58friends and I have some knowledge of how
39:00medicine works. So been testing lots of
39:03medical questions and some of them are
39:05hard some of them are you know sort of
39:07real time data um and uh you know before
39:1101 uh it was great in sort of reaching
39:15to the literature you know like the
39:17physician might lack certain certain
39:20knowledge so it knows what was published
39:24uh recently and things like that but it
39:26it was not at the reasoning level so one
39:28model was able to reason And the
39:31reasoning is extremely important
39:32medicine because you know even if you
39:35have all the all the information you
39:38still have to sort of consider that
39:41person's context and uh you know uh what
39:46what would be more likely to to treat
39:49that person and we don't always know the
39:51answer uh as well or how to how to
39:53diagnose it. Um, and so I think 01 was
39:57able to get to that point and at that
39:59point I said right now it's unethical
40:03for physicians not to use AI anymore.
40:06Um, I didn't say malpractice yet, but it
40:10truly unethical in the sense that, you
40:13know, you can use it uh you can still do
40:16your judgment obviously, but it will it
40:18will prevent you missing some sort of uh
40:22an obvious mistake or you know,
40:24sometimes nonobvious uh mistakes or
40:26diagnose things that require multiple uh
40:29clinical specialtities coming together
40:32and and you don't have that capability.
40:33You live in a village or something. uh
40:36but now I think I feel that it is truly
40:40uh uh going to be considered malpractice
40:42in my opinion. Um it's not legally so
40:45but uh eventually it will be um because
40:48uh a the current models the advanced
40:51models are able to uh diagnose and and
40:56write a treatment protocol better than
40:59uh or as good as uh a specialist in that
41:02field. It's not just a you know family
41:04physician. Let's say you you know you
41:06have a very complex uh cancer uh you
41:10know you you know the mutations and and
41:13what's not and you go to a specialist
41:15like an oncologist who is very very
41:17specialized on that. Um I believe that
41:19the mo the current models are at that
41:21level. So um and of course not every
41:26specialist is is is the top specialist
41:30right? So, so you you if if that was the
41:33case, we wouldn't have millions of
41:36misdiagnosis and mistreatments in the US
41:38alone every year. I think they've said
41:40something like 12 million misdiagnosis.
41:43Um
41:44I think 700,000 people suffer from it,
41:47die from it, from from uh uh from
41:50misdiagnos. Some of them is is totally
41:53innocent. You know, any any doctor could
41:56have missed it. Um but but now AI
42:00wouldn't miss that.
42:02So uh
42:05even even a specialist might make a
42:07mistake or misdiagnose or or or mistreat
42:10because they lack certain things that
42:12that the model doesn't have. So I mean
42:15imagine that you know um uh you refuse
42:18to use u MRI machine or CT machine
42:22because you say well you know that's too
42:24much technology. I'm just going to uh
42:26you know just do an X-ray because that's
42:28enough for me. And you miss a a tumor.
42:31Uh the AI models are able to detect
42:34certain tumors like breast cancer years
42:37before a radiologist is able to to to
42:40see that. So if you miss that, I mean
42:44that person's going to die if you don't
42:46if you don't. So that to me that that
42:49becomes uh a malpractice because the
42:52technology is at that level now. It
42:54wasn't it wouldn't be malpractice you
42:57know missing a a breast cancer u you
43:01know 5 years ago because nobody could we
43:03didn't have that technology but now we
43:05have that technology so you should you
43:07should definitely use it um um and and
43:10this is going to save a lot of lives. I
43:14mean uh if you could just reduce the
43:16misdiagnosis and and and again bring
43:18every every doctor to super doctor level
43:22I think that would be a really good
43:24thing. So what you're saying is based on
43:27you know what current data that doctors
43:31have available to them whether it's an
43:33MRI whether it's an ultrasound whether
43:36it's blood biomarkers
43:39this sort of data is what is given to
43:43>> yes
43:43>> you know a model like GP GPT 5.5 pro for
43:47example
43:48>> and with that data they're able to
43:52better diagnose
43:54better to predict um to see things like
43:57you mentioned cancer
43:59>> um is that better than a radiologist can
44:01is that some is that like based on you
44:03know what kind of uh data is
44:06>> implemented these these are studies I
44:07think uh Google uh did a recent study uh
44:10in fact uh a science paper came out
44:12recently which was done with 01 preview
44:15model which is a very old model I mean
44:18the current models are probably 10 times
44:19or maybe more
44:20>> was that like the first pro almost like
44:22the first the first first sort of the
44:24reasoning model that that I early tested
44:27in 2024 September it came out and um and
44:32they found that 01 model uh did did
44:36better than average doctor in diagnosing
44:38like significantly better they did
44:40didn't miss um and so imagine the the
44:43current models how how good they are uh
44:46but but I I think um it's not just sort
44:49of diagnosing a a disease because that's
44:52That's actually a small part of the job
44:55of a of a of a doctor. It's really
44:57there's a continuum. Most diseases um
45:01you know, okay, if you if you have a flu
45:03or some bacterial infection, you know
45:05what to do. You give it and then you see
45:07an output. But a lot of disease even in
45:10that condition that may not that may not
45:12be true because you know you might have
45:14a mutant virus or bacteria. So you might
45:16have to change the treatment or might
45:19have a little bit side effect. So
45:21there's a lot of continuum there. So I
45:23think AI can be involved in all of that
45:27process. So if you can continuously feed
45:30the data okay well the patients uh we
45:33gave this treatment it's doing well uh
45:36the blood pressure is down but you know
45:38has this symptom that so what should we
45:40do change the dose of the drug or add
45:43this or remove that drug and give
45:45another antibiotic? like there's a
45:47constant um uh uh process there and that
45:52that that's not always that constant
45:55because you know people people don't go
45:57to doctor every day right so you get a
45:59prescription you you see something works
46:01and then you go back and so what if
46:03there's something that's continuously
46:05monitoring you uh post treatment for
46:08cancer it's very important because
46:10cancer is a very dynamic uh disease
46:13there's the cancer which is constantly
46:16trying to survive and mutate and
46:19counteract against the immune system. So
46:22you give it so you know you give a drug
46:25chemotherapy works and then the cancer
46:28comes back again right so why is that
46:30because mutations are accumulating so
46:33can we catch that earlier can we change
46:35those decisions can we um make sure that
46:38we give more uh multiple drugs or
46:41different drugs so before the cancer
46:44have the opportunity to come back we
46:46prevent that uh possibility. So all of
46:48these um decisions uh can be made
46:52together with uh with AI and I I I I
46:55think it's going to have tremendous
46:56tremendous impact on healthcare.
46:58>> I do want to get back to the the cancer
47:01equation in a minute but before that I
47:03just think that you know physicians not
47:06all physicians know how to use AI. They
47:09don't know which models to use. Do they
47:11use GPT 5.5 Pro or Claude or you know um
47:16how do they sort of responsibly use it
47:19which you kind of talked about a little
47:20bit but without you know outsourcing
47:22their clinical judgment. Do you have any
47:25opinions on like the different models to
47:27use? And I do know you have a
47:29collaboration with OpenAI. You've been
47:31one of the first scientists really
47:33testing these models in a biological
47:36sort of arena. But I do kind of I do
47:39think that people and physicians that
47:41are listening want to know what how what
47:44what models do they use? We definitely
47:46are talking about if we're talking about
47:48open AI, it's not it's it's got to be
47:50the pro, right? It's got to be the
47:51reasoning model, but I mean what about
47:53Claude? What about Gemini?
47:54>> Yeah. So, um I think people have this
47:58sort of a misunderstanding of they think
48:02of AI as okay, we have AI, we have
48:05internet, so let's just use the
48:07internet. we have AI, let's use the
48:10but this is advancing so rapidly. Uh the
48:16AI model that we used one month ago is
48:19not the same AI model we use now. So
48:22it's just doubling in intelligence every
48:25few months. Um you know I gave the
48:27example of one preview. Uh some people
48:30uh got stuck at the GPT4 40 model. Oh
48:34yeah, I used it and it hallucinated a
48:36lot. Even, you know, 01 uh wasn't so
48:39good. You know, it was making mistakes.
48:41That's like an ancient history.
48:43>> That's why I haven't even asked about
48:44hallucinations.
48:45>> Yeah. So I it's um I mean the advantage
48:49I have is that you know because I'm all
48:51in on AI I'm continuously
48:54testing and and so I I I can see the
48:57evolution of of these models and they
49:00get you know 90% better 95% better 97%
49:04better like it just continuously updates
49:07itself and then eventually right now
49:10with 5.5 model I I don't see any
49:13hallucinations whatsoever. I mean there
49:15might be.1%.
49:18Uh but but it's it's extremely rare. Um
49:21and so so your trust level goes up.
49:23Again it's similar to like self-driving
49:25cars, right? So we we had self-driving
49:27cars for for almost a decade maybe and
49:30they they just keep on getting better
49:32and better because their AI models are
49:34are getting updated. So my my advice
49:37would be doctors should see this not
49:40something optional like they have to uh
49:43update their knowledge medical knowledge
49:46periodically. In fact they have to have
49:48test to do that to be certified or they
49:51have to update on new drugs that are
49:55coming out. Right? So you you can't just
49:57rely on some drug that came out 5 years
49:59ago, 10 years ago. you need to know what
50:00what was approved last month and you
50:03need to update your uh your knowledge uh
50:06in a similar way even more so they have
50:09to constantly update their AI knowledge.
50:12So AI has to be part of their uh their
50:16their practice. And of course my
50:18recommendation is always use the latest
50:20top model you can use. Uh right now it's
50:24GPT 5.5. Uh in fact uh I would always
50:27use for complex problems the pro model
50:31because that thinks uh in in minutes.
50:34But at least if you're using it on a
50:37daily basis uh in a rapid fashion always
50:40use the thinking model. The thinking
50:42model is different than the instant
50:43model. Instant model is also getting
50:45better, but it it needs to reason. It
50:47needs to think. Um, and especially if
50:50you're putting in lots of patient data
50:53and analyzing that, you you definitely
50:55need the pro model. And then there are
50:56there are these uh companies like open
50:58evidence and you know uh I think most
51:00doctors are starting to use that. Open
51:02evidence basically I think applies the
51:05latest model somehow. It up updated so
51:07so the doctors don't have to worry about
51:09it. And I I think there's going to be
51:10more companies like that who will
51:12provide that service. So the doctor
51:14doesn't have to worry should I use 5.5
51:17or putus 447 the the whatever the sort
51:21of the uh the harness model is going to
51:24pick the the best one for for medicine
51:26and and and apply it there. uh and of
51:29course hospitals should uh should
51:32implement AI uh just like you know big
51:34tech companies are you know there's
51:36enterprise level of AI that can be more
51:39secure you know protect the patient data
51:42so it should be like you know in front
51:44of the patient in hospital you see these
51:46monitors like the heartbeat and all that
51:48stuff should be an AI monitor like
51:51constantly monitoring the data and then
51:54giving information to the nurses to the
51:57doctors
51:57Okay, this is this last situation. Um,
52:01and now with AI agents, you can do that.
52:03Like I do it for my my daily life like
52:06for my email auto automatically my
52:08agents go and check my email and they
52:11tell me what's important so I don't have
52:12to go through hundreds of emails. So,
52:14oh, you know, this this is waiting for
52:16you. You have a a podcast with with
52:18Rhonda today, so you better be prepared
52:20for that. Um so uh yeah it it needs to
52:24be fully integrated uh almost like a a a
52:28a co- physician like you have the AI
52:31doctors working together with real
52:33doctors
52:33>> right have you I've noticed like some of
52:36the the companies that I've corresponded
52:39with or interacted with um it seems like
52:42they use claude a lot I mean I don't
52:44know if you've experimented with that
52:46but I'm kind of curious why um why
52:51that why that you know certain model um
52:54versus like but yeah in all fairness
52:56I've never used it. I use, you know, GP,
52:58I've been using GPT and the Pro and so
53:01every, like you said, you know, every
53:02time the hallucinations are like ancient
53:04history for me. Like I remember that was
53:06a big thing. Yeah.
53:08>> But it's going so fast and better now.
53:10But like what what what what's the
53:12difference between, you know, for
53:14example, Claude and GPT 5.5 Pro? For
53:18certain things there is no more
53:20difference because the intelligence has
53:22has peaked uh for for uh you know for
53:25for uh doing regular diagnosis not very
53:30very complex uh cases cloud is is is
53:33great. Cloud is also very very good uh
53:36like the Opus 4.7 model uh the recent
53:38model for analyzing data sets. So it's
53:41you know it can take also millions of
53:42data uh analyze it and and and and do a
53:46a great job. Um uh my preference is uh
53:49you know GPT5 right now is 5.5 pro
53:53because um it what I mentioned it has
53:56this extra insight I mean the for me um
54:00I need that extra level of insight um
54:03that's predictive um uh and
54:06>> the intuition that
54:07>> the intuition and really kind of a deep
54:10understanding uh but if I'm if I'm going
54:13to diagnose and treat a subt type of a
54:16lung on cancer. I am pretty sure, you
54:18know, Gemini 3.1 Pro or Cloud 4.7, they
54:24all do a pretty good job. Uh I I think
54:28the re some reason people prefer cloud
54:30is that it's maybe it's more pleasant to
54:34interact with. Uh you know, kind of more
54:36humanlike. I think uh GPT models are
54:39starting to get there, but still there's
54:42something about cloud that people enjoy,
54:45you know, interacting with it. It's it's
54:47really a matter of
54:48>> personable or I I think it's it used to
54:51be more more personable. Um and so it
54:55doesn't really matter. I mean, I think
54:56they're they're they're really they're
54:58all super top levels unless you're doing
55:01like a research or a very uh very
55:04complex problem. uh you know for example
55:06we did a test with uh with a colleague
55:08of mine on on skin disease with GPT5 pro
55:12model you know it was able to uh
55:15diagnose a skin disease that uh my
55:18friends couldn't really diagnose um uh
55:21just based on a a photo and and a
55:24symptom. The other models couldn't do
55:26that. they could do 90% of the the cases
55:30as well, but there's that one extra case
55:32or two extra case that is really
55:34difficult that's could go anywhere. The
55:38pro the GPT pro model was able to cross
55:41that that threshold. So those kind of
55:43cases you really need the very high
55:46level like you know uh you don't you
55:49don't go to a a a professor at Harvard
55:52for for any any reason right so it has
55:55to be very specialized
55:57disease that other doctors couldn't
55:59diagnose or something like that so
56:01that's how that's how I view it
56:04>> we just there was just news yesterday um
56:07from open AAI of uh GPT Rosalyn which I
56:10know you can't talk about much
56:12from what was publicly available, it
56:14seems as though it's going to be used in
56:16drug discovery. Um, I I'm wondering what
56:20you think in terms of like the future of
56:22aging research, biology, medicine. Are
56:25we going to be using these more
56:27specialized types of AI models or do you
56:31think more of a generalist like GPD 5.5
56:34Pro and and the, you know, the
56:36subsequent ones that come out after it
56:38are going to be the key to unlocking,
56:41you know, medicine breakthroughs and
56:43biology breakthroughs?
56:46Um my preference would always be the
56:49generalized models because again you
56:51know going back to AGI AGI uh so if if a
56:55model is has um you know of course there
57:00there are some utilities of models that
57:02are only trained on I don't know like
57:04the EKGs or um RNA sequencing or
57:08something like that and they they'll be
57:10very very good at that like the the the
57:12best chess player AI model or um the
57:15best go player AI models but they will
57:19miss that uh connection because again I
57:22view medicine as a kind of a holistic um
57:26art in a way. Uh if you are just trying
57:29to analyze one set of data the the
57:32specialized models could be could be
57:34very very useful. In fact, you know, I
57:36gave the example of EKGs. Um, most
57:38generalized models were not terribly
57:41great at um, for some reason, you know,
57:43the the EKG images were were not they
57:47were not very good at diagnosing what
57:49what what it was showing. Um, and and
57:52you know, specialized models were very
57:55good because they were trained with, you
57:57know, millions more EKG data sets than
58:00the generalized model was. Um but but I
58:04think you know if if we can train the
58:07generalized model or fine-tune it or
58:10overtrain it I I don't know how to say
58:12it. Um
58:14then they will be better than
58:16specialized models all the time. Um
58:19because not only they have they know all
58:22about EKGs but they know about
58:25radiology, they know about RNA, they
58:27know about pro proteins. So they can
58:30take that information and and excuse me
58:33analyze the EKG the the electroc
58:36cardiogram your your your heart beats in
58:38the context of all the other biology. So
58:42that will that's very enriching uh
58:44knowledge. Um but um I think you know
58:49specialize in the sense that you can
58:51take these big models and you can sort
58:54of I don't know harness them or
58:56fine-tune them because there's a lot of
58:58data sets that's not public. So these
59:00these models they don't have access to
59:02that. You might have some um data you
59:06know locked in certain because of
59:08regulatory reasons whatever. So you can
59:11take take a big model. In fact, you
59:13don't you may not even need the the
59:16closed models. You can even take some of
59:17the open- source models which are which
59:19are now getting very good. If you uh you
59:22can train them on that uh and they also
59:26have the generalized knowledge and
59:28combined with that they they'll probably
59:29do better.
59:31>> So I wanna I want to talk about
59:34there's treating disease, there's curing
59:37disease, and then there's reversing
59:39aging. So let's let's start with
59:42curing disease, treating diseases,
59:44curing diseases because you know
59:45obviously we do die of age related
59:47diseases, cardiovascular disease being
59:50the number one killer in in most
59:52developed countries. We have cancer.
59:55That's a really big one. And and with
59:57cancer, it's just such an awful disease
1:00:01to have. And anyone that's listening
1:00:03that has either had cancer or knows
1:00:06someone that has had it, you know, knows
1:00:08this is this is true. But also, I think,
1:00:11you know, cancer, a lot of people think
1:00:12about it as one disease. Non-scientists,
1:00:15non, you know, physicians, they kind of
1:00:17think about cancer as this this one
1:00:18disease, right? As you and I both know,
1:00:21it is definitely not one disease. It's
1:00:24hundreds of diseases. Um, I'm curious on
1:00:29first of all, you know, we still don't
1:00:32have a cure for cancer. I mean, we've
1:00:33we've made a lot of progress, right? And
1:00:35different cancers and can be treated
1:00:37better than others, but can you talk a
1:00:39little bit about why it's been so hard
1:00:42to find a treatment for cancer?
1:00:46>> Yeah, I think uh the the important thing
1:00:48to clarify is that cancer is not one
1:00:51disease. It's probably 100 disease h 100
1:00:54different diseases that have probably
1:00:57hundreds of sub sub diseases or sub
1:01:00subtypes if you like. Um in fact certain
1:01:04cancers are 100% curable or 95% curable.
1:01:08Uh you know like child some of the
1:01:10childhood leukemias which were
1:01:12completely fatal you know couple of
1:01:14decades ago are now you know 90% or or
1:01:17close to 100% curable. if you catch uh
1:01:21certain cancers early enough again 100%
1:01:24uh cure rates almost uh so so because
1:01:29it's it's a very different set of uh uh
1:01:32diseases um the the cancer of pancreas
1:01:37is very different than cancer of uh lung
1:01:40cancer or breast cancer or there are
1:01:42some cancers that are so slow like if
1:01:46you get um certain types of cancers
1:01:49If you're age 80, doctors don't even
1:01:52bother to treat it because by the time
1:01:55that will unless we cure aging first uh
1:01:58because by the time you die of aging,
1:02:00you know that that cancer is not going
1:02:02to kill you. Aging is going to kill you
1:02:03first or you know there's certain
1:02:04prostate cancers at certain age. So that
1:02:08that's why we have to really understand
1:02:10that this is a very complex biology. But
1:02:13more importantly, why cancer is such a
1:02:16challenge is that the the cancer cells
1:02:19are part of us, right? So, uh if you're
1:02:23infected with a bacteria or a virus, you
1:02:26know, it can kill you, right? They're
1:02:28extremely dangerous, but we are able to
1:02:31recognize them as an enemy, as a threat,
1:02:34your immune system, and we can fight
1:02:36back, you know, uh not always
1:02:39successfully, but most of the time very
1:02:40successfully.
1:02:42And we can also target them very
1:02:44specifically like we have an antibiotic
1:02:46that will only act on the bacteria. It's
1:02:48not going to touch your normal cells
1:02:50because it's only uh a foreign or
1:02:52organism. But cancer is not like that.
1:02:54So if I try to stop cancer with
1:02:57something I'm also trying I'm also
1:03:00stopping some other cells that are
1:03:01normal, right? That's why people lose
1:03:03their hair, their immune system is
1:03:05greatly weakened because the immune
1:03:07system has to divide. your um hair has
1:03:10to hair cells have to divide. So you you
1:03:13block them because they the cancer cell
1:03:15is also dividing and and you your side
1:03:18effects of chemotherapy sometimes worse
1:03:20than having the cancer like hundreds of
1:03:22thousands of people die because of that.
1:03:24So the revolution in cancer was uh
1:03:28recently because of what we call
1:03:30imunotherapy.
1:03:32The question was why uh can we make the
1:03:36immune system to recognize cancer as
1:03:40foreign threats like they're kind of
1:03:42like terrorists, right? So a terrorist
1:03:44you will not know if that's an enemy or
1:03:47not. They look like you, you know, they
1:03:48just come in and then they they create
1:03:50uh so the immune system is seeing it
1:03:52that way that it thinks that the breast
1:03:54cancer cell is not so different than a
1:03:57normal breast breast cell, you know,
1:03:59like epithelial cell, whatever. And so
1:04:02it doesn't know what to do. If we could
1:04:04teach the immune system or if we could
1:04:07remove some of the breaks that it has
1:04:09regulation and let it recognize and
1:04:12attack the cancer cells, then that could
1:04:14have a a tremendous effect. That was the
1:04:16hypothes and it it actually worked. So
1:04:19cancer imunotherapy I think uh is is
1:04:22more powerful now than than chemotherapy
1:04:25and radiotherapy put together. I mean
1:04:27they still have a have a role. Um and of
1:04:30course the other thing is that how can
1:04:32we make the treatments very specific
1:04:35right so if I give a chemotherapy that's
1:04:38not specific it's like trying to hit the
1:04:40patient on the head and hope that the
1:04:41cancer will die before the patient dies
1:04:44but if I know this single mutation
1:04:47that's happening on you know whatever
1:04:50EGF receptor uh in certain cancers I can
1:04:54develop a small molecule which will only
1:04:56act if there's that mutation on the EGF
1:04:59receptor or EG whatever and so it's not
1:05:02going to touch anywhere else it's only
1:05:04going to target the the and in fact
1:05:06people call them smart drugs and they're
1:05:08they're extremely effective right so uh
1:05:11if you have that particular mutation
1:05:13your 1% of the lung cancer patients you
1:05:16get treated with that drug you get
1:05:17almost 100% cure rate um but again you
1:05:21know uh we can make this even much
1:05:23better so for example immune system can
1:05:27be engineered something that we work on
1:05:29in the lab uh to recognize like
1:05:33literally engineer we take the cells out
1:05:35we train them we put genes into them say
1:05:38okay so if this gene binds to a cell
1:05:42assume that that's a threat and kill
1:05:44that that and so it's called carti
1:05:46therapy and they will go and seek out
1:05:49whatever uh the cancer cells that have
1:05:52that marker and kill them the advantage
1:05:54of that is that cancer doesn't have much
1:05:57weight to escape escape that. It can try
1:05:59to suppress the immune system, but other
1:06:01than that, even if it mutates, you know,
1:06:04the the immune system will still
1:06:06recognize it and find that few cells
1:06:08that are hiding somewhere and and
1:06:10destroy it and and that's showing
1:06:12incredible results. So, the mRNA
1:06:15vaccines, which I think is going to be
1:06:17revolutionary, is is on that basis,
1:06:21right? So, and that really personalized
1:06:24the cancer. So I have a breast cancer
1:06:27but my breast cancer has certain type of
1:06:30mutations that other patients don't
1:06:33have. So even if the immune system can
1:06:36recognize X patient, it won't recognize
1:06:38mine because the cancer has different
1:06:40mutations. If I take those mutations and
1:06:44synthesize what's called RNA and then
1:06:47give it back as a vaccine and train my
1:06:50immune system and tell the immune
1:06:52system, look, if you see these mutations
1:06:54in these genes, that's an enemy. Go
1:06:57destroy that. That's mRNA vaccine. And
1:07:01that becomes extraordinarily powerful
1:07:03because now you are directing your
1:07:04immune system to to an internal threat
1:07:07just in you. and let's say the the the
1:07:10cancer required different mutations, you
1:07:12can create another mRNA vaccine and then
1:07:15train the immune system to that as well.
1:07:18Um so uh you know I think that's those
1:07:22those are the the the the difficult
1:07:24parts but but we see the light at the
1:07:26end of the tunnel.
1:07:28Can cancer cancer is going to be 100%
1:07:31curable uh
1:07:34probably less than a decade.
1:07:35>> How is AI going to make that happen?
1:07:38Yeah. So, in fact, it's already making
1:07:41that happen. You probably heard of this
1:07:42story from Australia. This computer
1:07:44scientist um had chat GPT and and some
1:07:48other AI models to develop an mRNA
1:07:51vaccine for his dog. His dog had I think
1:07:55a melanoma
1:07:57and he um he got it sequenced. He took
1:08:00the sequence and gave it gave it to to
1:08:02an AI model and the AI model designed
1:08:05the precise mRNA molecule that needs
1:08:09that the dog needs dog's immune system
1:08:11needs to be trained got it synthesized
1:08:14and I think it was able to apply it in 3
1:08:16months. Um probably could have been
1:08:18shorter if there wasn't regulations and
1:08:20and the tumor started to to regress and
1:08:23the dog is was was alive when it was
1:08:26supposed to to die. So I mean that that
1:08:28that's a very uh obvious and simple
1:08:32version but because
1:08:35there are hundreds of difference of
1:08:37cancer types you can imagine that we'll
1:08:41have maybe hundred different treatments
1:08:45for just a type of a lung cancer someone
1:08:48will be mRNA someone will be small
1:08:51molecule targeting that so to be able to
1:08:54develop those ondemand and or very very
1:08:58rapidly, we're going to need AI. So the
1:09:00AI is going to model every possible
1:09:03mutation and we'll screen millions and
1:09:06millions of compounds. And so we'll
1:09:08we'll we'll get to a point where we'll
1:09:10have hundreds of new drugs coming out
1:09:12every month maybe, you know, uh uh and
1:09:16we'll you know this this thousand drugs
1:09:18is for breast cancer patients. But you
1:09:20know if you have this and this this
1:09:21mutations and if it's stage four then
1:09:24you take this combination. And if it's
1:09:26that yeah you you you take this protocol
1:09:30um and um that that's that's how AI is
1:09:33going to of course you know if you get
1:09:34to digital twin that that will
1:09:35accelerate
1:09:36>> right and that's that's the next
1:09:37question is you know so let's let's say
1:09:40we have the true personalized medicine
1:09:42and personalized cancer treatment but
1:09:45you also need to know about side effects
1:09:47like am I going to take this mRNA
1:09:49vaccine and my immune system is going to
1:09:50go crazy and start to inflame my heart
1:09:52and give me myocarditis or right
1:09:55How do you also see this the digital
1:09:58twin which now has you know genomic
1:10:01information all your proteins
1:10:02metabolites and everything in real time
1:10:04data then it can also simulate well
1:10:08what's going to happen if we give this
1:10:10specific mRNA vaccine cancer vaccine or
1:10:13this small molecule to this person
1:10:16>> absolutely I mean you know so so you
1:10:19mentioned myocarditis which by the way
1:10:22happened during covid pandmic IC and
1:10:24that's why there was a lot of uh
1:10:26antivaccine sentiment but people uh
1:10:29didn't appreciate that you know co virus
1:10:32itself caused myocarditis. Yes, the
1:10:35vaccinated people young people at one in
1:10:395,000 to one in 10,000 rate got
1:10:42myocarditis. It wasn't it was mostly
1:10:45fatal. But the question should be asked
1:10:47like why is it that one out of 10,000
1:10:49got myocarditis and the other ones
1:10:51didn't? Or in fact, we can reverse that
1:10:54question. You know, we g we vaccinated
1:10:56everybody, but if you were a young
1:10:59person, your your um chance of dying
1:11:02from COVID was let's say 1 in,000 or one
1:11:04in 10,000. So n 999 people got didn't
1:11:08have to be vaccinated. But to save that
1:11:11one person, we have to give that
1:11:12vaccine. Or I'll give another more uh
1:11:15general uh you know, we give statins to
1:11:18anyone who has high cholesterol. So I I
1:11:21think like one out of five or one out of
1:11:2210 people truly benefit from that. Uh
1:11:26high cholesterol doesn't automatically
1:11:28doesn't mean you're going to get
1:11:29atheroscllerosis. You need to have
1:11:30inflammation this and that. But because
1:11:33we don't have the data, we cannot
1:11:34predict that. It's not personalized.
1:11:38Millions of people take statins and to
1:11:41save few thousand people. Yes, that's
1:11:43that's a good thing because you don't
1:11:44know. Um so AI will be able to do that.
1:11:48So we'll we'll tell you okay not only um
1:11:53we'll create the drug just for you but
1:11:56also we'll say okay you don't have to
1:11:58take this this medicine like you you
1:12:00should take this or maybe you don't even
1:12:03need any any treatment at all like you
1:12:05have an infectious disease or whatever
1:12:07or maybe certain cancers this will be
1:12:09enough like we give extra chemotherapy
1:12:12plus imunotherapy plus radiotherapy
1:12:15why are we doing that because we're not
1:12:17sure if one is going to be enough or
1:12:19not? Um and and so uh that will
1:12:22dramatically reduce the the the side
1:12:24effect issue. You might still have some
1:12:27side effect of course, but it's manage
1:12:30it will be manageable side effect. It's
1:12:31not going to kill you, for example.
1:12:34What about using AI to predict
1:12:38cancer a decade or years before it forms
1:12:42based on your proteins and metabolites
1:12:45and your biomarkers and maybe perhaps
1:12:48your genetics too, right? Like how do
1:12:51you see that? We're talking about
1:12:53personalized cancer treatment, but what
1:12:55about being able to prevent cancer
1:12:58before it happens, you know, years
1:13:01before it happens? Yeah,
1:13:03>> again great question because I think
1:13:05this is um this is so important that
1:13:09people don't think about very much. um
1:13:12we say health care you know we don't
1:13:14have health care we have sick care right
1:13:16so we we never take care of healthy
1:13:19people like you don't go to a doctor to
1:13:21say oh how healthy I am or just just go
1:13:24to a doctor and say can you check my
1:13:26immune system you know is it is it
1:13:28healthy am I going to am I going to get
1:13:30sick am I going to have cancer they
1:13:32won't be able to answer that question
1:13:34only if you get sick they will treat
1:13:36what the problem is um and so the
1:13:40preventative medicine is going to be so
1:13:42absolutely critical. I think not all but
1:13:47most diseases can be prevented. Some are
1:13:51just bad luck. You know, it happens no
1:13:54matter what you do. If even if you live
1:13:55the perfect life, you might still get
1:13:57certain disease but but a lot of them
1:13:59were because of your genes and so on. A
1:14:01lot of them can be prevented and I think
1:14:03AI is going to be amazing in that
1:14:05because it's already able to do it. uh
1:14:07that there was a study from um a UK bio
1:14:10bank um UK has this amazing bio bank
1:14:14with 500,000 people uh lots of data sets
1:14:18incredible data sets and so and this was
1:14:21actually done I think more than a year
1:14:22ago with models that were a year or two
1:14:26years old they took a lot of that data
1:14:29and they were able to predict about
1:14:32thousand diseases
1:14:34before they happened of course this was
1:14:36kind of retroactive. So they knew what
1:14:38what people were going to get based on
1:14:40their data that was collected years
1:14:42before. But the AI was telling you,
1:14:44okay, this patient's going to have this
1:14:46disease that, but not patient, normal
1:14:48healthy people, they're going to get
1:14:50this and that. So that to me that was
1:14:52that was amazing and that's going to get
1:14:54better and better because there are
1:14:56there are there are always signs
1:14:59like cancer doesn't just develop in
1:15:01days. It takes years. uh if we probably
1:15:06most of us might have some cancer cells
1:15:08you know most of it controlled by immune
1:15:10system and so on and it you know slowly
1:15:13grows it has to have another mutation
1:15:15another mutation but but there's
1:15:17probably some signs of that somewhere
1:15:20you know whether it's in your metabolism
1:15:22or this you know and a AI even if it's
1:15:26100% will be able to say okay look um I
1:15:30think that you know if this is if this
1:15:32is the lifestyle that you continue your
1:15:34chances of getting this disease is now
1:15:36is 85% or whatever. Like I wear a
1:15:40glucose monitor. Um I'm not diabetic,
1:15:43you know, uh but I I want to see every
1:15:47minute or every five minutes what my
1:15:49sugar levels are in a continuum or if I
1:15:52eat something, you know, is it spiking?
1:15:54Is it coming down? Because I want to
1:15:56prevent insulin resistance. one of the
1:15:58the worst things that can happen to you.
1:16:00If I uh if I don't uh do that, I won't
1:16:04know until I get diabetes.
1:16:07My my insulin if if if if my sugar is
1:16:10constantly spiking and then you know
1:16:13insulin is just working too hard and
1:16:14hard that that could continue for years.
1:16:16By the way, um that at some point it's
1:16:20going to break, right? Um for some
1:16:22people it might continue 50 years,
1:16:24nothing happens. Some it might be five
1:16:26years, but that data set probably has
1:16:30that predictive value that plus my my
1:16:33age, my genes, but whatever. So, uh
1:16:36yeah, that I think um everyone's going
1:16:39to have their own um AI, I don't know
1:16:42how what to call it, uh health coach or
1:16:45something. Uh but but it will it will
1:16:47continuously analyze the data. Um and
1:16:50and hopefully we'll it will be much
1:16:52easier to collect data because that
1:16:54that's another issue you know we don't
1:16:57collect data like we know nothing about
1:17:00uh you know there there are more than
1:17:02thousand metabolites in our bloodstream
1:17:04so we look at maybe you know 10 of them
1:17:0620 of them only if we get sick not even
1:17:08for a checkup so we have to have a
1:17:10continuous um like a glucose monitor I
1:17:14want to see what my you know uh proteins
1:17:17are changing hormones are changing you
1:17:19know in
1:17:20reasonably continuous manner.
1:17:22>> Such a good point and I'm so glad you
1:17:24brought up the UK bio bank study. I
1:17:26remember I think the model was like
1:17:27called Milton or something and it was
1:17:29it's a Astroenica own like developed it
1:17:33or something and and I remember looking
1:17:35at this study because like you mentioned
1:17:37the bioank data is huge data set and it
1:17:40just spanning many decades and so I
1:17:43think they looked at you know like over
1:17:45200 plasma proteins you're talking about
1:17:4710 we're talking about 200 right? Oh
1:17:49yeah.
1:17:50>> And and all the other data, right? And
1:17:52they were able to predict and I think
1:17:54cancer and neurogenerative disease were
1:17:55at the top of like 10 years before and
1:17:58they were able to look at the people. So
1:18:00the AI AI predicted it based on based on
1:18:02all this biometric data. And then they
1:18:05looked and said, "Oh, yep. Those people
1:18:06actually did end up getting cancer and
1:18:08Alzheimer's disease." And it was very
1:18:10accurate. Yes.
1:18:11>> And and to me, the exciting thing here
1:18:13is that you can intervene before it
1:18:17happens. You can make
1:18:19lifestyle changes, you can make dietary
1:18:21changes. I mean, these things matter.
1:18:23They do matter. And and that is
1:18:27exciting. Uh because then you don't even
1:18:29have to get to the drug part, which you
1:18:31know, maybe you will, but if you can
1:18:33make these changes, if you know, hey,
1:18:35I'm on this trajectory to get cancer. I
1:18:38have all this inflammation. I have all
1:18:40these things happening. If I don't make
1:18:41a change now, then in 10 years, I might
1:18:44have a cancer.
1:18:45>> It's very motivating, you know, for for
1:18:47someone. So, it's very exciting um as
1:18:51well and and and then having AI in there
1:18:53is just going to make it even even
1:18:54better. Um and then I I want to get in I
1:18:57want to get into age reversal and and
1:18:59before we get to that, you know, you
1:19:02you've really been a pioneer in this the
1:19:06field of AI being involved in biology.
1:19:10You know, you were talking to me about
1:19:11your your blog I don't know was it 30
1:19:13years
1:19:14>> bios singularity. Yeah. 25 years ago. 20
1:19:1725 years ago.
1:19:17>> Yeah. So you have this blog bios
1:19:19singularity predicting. Can you can you
1:19:21talk a little bit about it?
1:19:23>> Yeah, sure. Uh so uh in fact I I gotten
1:19:26interested in AI early 90s uh after I
1:19:29graduated medical school. Um you I was
1:19:32very interested in computers uh when I
1:19:34was a teenager. The the first computers
1:19:35had come out at the time and you know I
1:19:37was trying to code and uh you know just
1:19:41I mean I loved it. It was it was just so
1:19:44wonderful. Um, but you know, I went to
1:19:46medicine because I figured biology is
1:19:48much more complex, so I should first try
1:19:50to figure that out. But then immediately
1:19:53I realized, and I'm sure you did too,
1:19:55you're a scientist as well. Um, that
1:19:58biology is so incredibly complex. I
1:20:00said, well, I mean, you know, we don't
1:20:02have any any chance of figuring this
1:20:05out, you know, because there's going to
1:20:06be so many so many data sets. So that's
1:20:09when I first got interested in uh in AI.
1:20:11Uh, of course at the time, you know, AI
1:20:14was was very primitive. Um, uh, but fast
1:20:17forward uh, you know, one of the one of
1:20:19the books that influenced me was, uh,
1:20:21from, uh, um, Ray Kurtzwell. I'm sure a
1:20:24lot of people follow technology know
1:20:26him. Um, he wrote this book techn
1:20:29singularity is near. So he he called a
1:20:32point of singularity where uh the the
1:20:36computation or technology is advances
1:20:38exponentially
1:20:40so much that you cannot even predict
1:20:43what will happen next day. I mean
1:20:45because it's sort of like a
1:20:47self-training AI models and he he had
1:20:50these figures where he would plot the
1:20:52advances of AI you know say you know by
1:20:552029 it will be at the human brain level
1:20:58and you know we'll reach AGI and you
1:21:00know it was just unbelievable and most
1:21:03people thought that he was just talking
1:21:07crap or you know science fiction you
1:21:09know they didn't believe it how could
1:21:11that happen and so on but you know I got
1:21:13I got very excited In fact, I have a
1:21:15signed copy from Ray uh for for the
1:21:17book. And so being inspired from that, I
1:21:19I started this um blog called bios
1:21:22singularity. So I said, okay, so so
1:21:25computation is going exponential, but uh
1:21:28biology is sort of a computation as
1:21:30well. I mean it's it's based on
1:21:31information and so uh but it's just much
1:21:35more complex. So it should also expand
1:21:38exponentially. And if you if you plot
1:21:40that curve that that means that by you
1:21:43know based on my calculations 25 years
1:21:46ago in fact I wrote it on the on the
1:21:48about page of the blog by by year 2035
1:21:51or so we should be able to treat all
1:21:53diseases uh and by 2045 or so that we
1:21:58should be able to completely reverse
1:21:59aging. In fact by 2050s we will get to a
1:22:03point what I call human 2.0 you know
1:22:05because at that point we have a complete
1:22:07understanding of biology. Uh then we can
1:22:10truly engineer it. We can create new
1:22:12biological organisms. We can you know
1:22:15change our biology, our genome,
1:22:17reprogram it. Um
1:22:19>> rewrite our immune system. Yeah.
1:22:21Exactly. um in in in many possible ways
1:22:24because it's kind of a messed up if you
1:22:27if you think about it like you know
1:22:28biology we think is a miracle but it's a
1:22:31it's a bad kind of a a legacy
1:22:33engineering right it's not it's not a
1:22:35bad engineering it's a legacy because
1:22:37biological system finds something it
1:22:39can't get rid of it can't start from
1:22:41clean slate so it builds on top of it so
1:22:43you get regulation over regulation over
1:22:45regulation and then of course you know
1:22:47with like immune system that I study you
1:22:49know you get lots of autoimmune disease
1:22:51is his immune system kills a lot of
1:22:53people you know even during like
1:22:55pandemics and things like that or it
1:22:57doesn't recognize the cancer cell and
1:22:59things like that. So why you know we
1:23:01should be able to design like immune
1:23:03system 2.0 like clean slate really
1:23:06greatly engineered the immune system.
1:23:08Well, and I said you know by 2045 50
1:23:11we'll get to that point. Um and and
1:23:14actually, you know, again, at the time
1:23:16it sounded really crazy to people. Uh
1:23:19but now I feel that I was I was too
1:23:21conservative. We'll probably get there.
1:23:23Uh but but the key point is that I wrote
1:23:26specifically in the about we will do
1:23:29this because of artificial intelligence.
1:23:31You know, I was just taking the plot
1:23:33that Rey uh plotted. You know, I said,
1:23:36okay, by 2029, AI is going to be at that
1:23:40point. it will be good enough to apply
1:23:42to the biology and that will allow us to
1:23:45solve diseases and then and then the
1:23:46aging the fact that you know the timing
1:23:50was was pretty good uh uh again even
1:23:53even a bit conservative uh uh I feel
1:23:56great about it that's why you know I I'm
1:23:59all in on AI like wow um that it's
1:24:02happening it's really happening
1:24:04>> so aging
1:24:07is very complex and you know as you know
1:24:09it's not one process. We've got these 12
1:24:13hallmarks of biology. We now have 12.
1:24:16Genomic instability, mitochondrial
1:24:18dysfunction, you know, cellular
1:24:20scinessence, on and on. We've got
1:24:21there's 12 of them.
1:24:22>> And we know organs are are aging at
1:24:25different rates. They're they reach
1:24:27their peak at different rates and they
1:24:28age at different rates and everything is
1:24:31interacting in a very complex way.
1:24:34What do you see as the bottleneck
1:24:39for understanding the aging process and
1:24:42also reversing it?
1:24:44>> Um I mean more so than the bottleneck uh
1:24:49this is the way we have to think of
1:24:51aging. Uh biology actually um uh is
1:24:57programmed to prevent aging. Right? So,
1:25:00it's not like um uh it's not like a car
1:25:04in a way because once you make a car um
1:25:08you have to constantly bring it to a
1:25:10repair shop or you have to repaint it.
1:25:13Biology does that internally. If it
1:25:16didn't, we would age immediately. Like
1:25:17there is a disease called progeria.
1:25:19These children get aged uh by the age of
1:25:2378 they become like a 89 year old
1:25:25because of single point mutation in one
1:25:27of their one of their genes because they
1:25:29lose their ability to repair um whether
1:25:31it's the DNA repair whether it's getting
1:25:34reg rid of the old cells or cleaning up
1:25:37the tissues and then regenerating like
1:25:39stem cells creating new cells. So this
1:25:42program continues for for sometimes
1:25:45decades otherwise we we wouldn't survive
1:25:48for some animals for some organisms is
1:25:51only a couple of years for for us is
1:25:53about you know maybe 50undred years uh
1:25:56for some veils is hundreds of years so
1:25:58so you know same biology it's just that
1:26:01one of them decided that you know I can
1:26:03keep a veil um or you know whatever some
1:26:07animals um you know older longer because
1:26:11they don't they're not getting hunted or
1:26:13they can reproduce later and so on. So
1:26:15what happens in in in the in the
1:26:17biological system is that somehow uh
1:26:20this program breaks down and you start
1:26:22to lose what's called the res
1:26:24resilience, right? So when you are uh
1:26:28age 30 or 40, you're a you're resilient.
1:26:32you can tolerate much more damage than
1:26:36someone who's 70 years old, 80 years old
1:26:39because your your your systems are uh
1:26:42you know even if you get wounded or if
1:26:44you uh get sick you can recover uh
1:26:47easier. Um uh but that that sort that
1:26:50resilience is lost and that the reason
1:26:53why it's low that there is a sort of an
1:26:55information loss because the biological
1:26:59system has a certain information that it
1:27:02knows when certain genes should be
1:27:04turned on when things should be
1:27:07regenerated when it needs to be like
1:27:09your skin. You know why you get
1:27:12wrinkles? because your cells stop making
1:27:14collagen and then all kinds of crap
1:27:16accumulates under your skin and then you
1:27:18know the guy the guys who like the
1:27:20macrofasages or whatever was supposed to
1:27:22clean there they don't do their job.
1:27:24There's some sort of a breakdown in
1:27:25information or communication or you know
1:27:28intracellular communication is one of
1:27:30the hallmarks of of aging. And then of
1:27:32course why that happens is is that 12
1:27:36hallmarks is is is the reason many
1:27:39reasons you know uh for example the
1:27:41bacteria in your gut is is is a reason.
1:27:44So so these bacteria produce all kinds
1:27:46of metabolites that help your immune
1:27:49system to constantly regenerate keep it
1:27:52in optimal shape. If that changes then
1:27:55you know your metabolism is changing
1:27:58your glucose levels your mitochondrial
1:28:00uh mutations and so on so forth. So all
1:28:03of these things accumulate you know
1:28:05epigenetic changes and DNA mut mutations
1:28:08and somehow the the biology forgets well
1:28:11what am I supposed to do like how how am
1:28:13I dealing with that also becau because
1:28:18when a damage happens it's harder to fix
1:28:21a damage than prevent it right so if if
1:28:24you're continuously taking care of your
1:28:26car or your house the likelihood of it
1:28:29you know breaking down is much less than
1:28:32If you wait until like okay nothing
1:28:35works yes you can reverse it but it's
1:28:38going to take a lot more effort and so I
1:28:42think uh what will happen is that for a
1:28:45younger uh individuals in the next
1:28:48decade or so uh there uh
1:28:53for them it's not just it's not going to
1:28:55be reversal it's going to be prevention
1:28:58of the aging process it's going to be
1:29:00maintaining
1:29:02that process the resilience decades
1:29:04more. So we will come to a point where
1:29:07if you are 20 30 whatever years old you
1:29:10won't age anymore because it's going to
1:29:12be constant reversal. But people who
1:29:15have already aged you let's say you're
1:29:1780 years old 90 years old then we're
1:29:20going to have to reverse that process.
1:29:22That's that's a more difficult we'll be
1:29:24able to do it. Definitely we'll be able
1:29:26to do it. Um uh uh but um uh it will
1:29:30require a lots of engineering approaches
1:29:33because you need to fix most of those
1:29:36hallmarks. If you're younger, you
1:29:39prevent those hallmarks from happening.
1:29:42You maintain the the information uh uh
1:29:45much much longer. Um both of those uh
1:29:48will will will happen. um uh uh we we we
1:29:51just need to figure out what that
1:29:53information is being lost and we put it
1:29:56back.
1:29:56>> Do you think so? Let's first talk about
1:29:58preventing the aging if you're a younger
1:30:00person because it's easier to to do
1:30:03always prevent if if you have a person
1:30:09you know who's 20 or 30 years old. Do
1:30:11you think that the approach would be
1:30:16finding first of all do we even know all
1:30:18the repair processes that are we we have
1:30:20discover we have what we know right
1:30:23>> but we still have a lot to discover
1:30:25>> we have a we probably have a lot to
1:30:26discover and so like do you think
1:30:28there's going to be a a discovery where
1:30:31we figure out like you know we know
1:30:33things like autophagy stem cell
1:30:35depletion you know all these stress
1:30:36response genes like antioxidant like all
1:30:39these things DNA repair mitochondrial
1:30:41the way mitochondrial repair itself,
1:30:42right? Um, are we going to be enhancing
1:30:45or like tuning these up so that they
1:30:47keep working at their prime continually
1:30:50or do you think we're going to have
1:30:51again this like information where we why
1:30:54why are those things going down? Are we
1:30:56going to just then go to the information
1:30:59of it, the epigenetics perhaps? Um, and
1:31:02is it going to be more targeted towards
1:31:05those genes or are we going to have more
1:31:07of this? you know, we'll get into this
1:31:09cellular reprogramming and partial
1:31:11reprogramming, but um I'm I'm curious
1:31:13like how you see AI coming into that
1:31:16process. Like I guess we don't know
1:31:18that's the part of the problem, but then
1:31:20we have to figure out how to give these
1:31:21del you know treatments to people,
1:31:24right? That's another part of the
1:31:26equation.
1:31:27Um, so I mean I think you know the the
1:31:29ones that you mentioned about sort of
1:31:30the lifestyle changes and they of course
1:31:34help a lot but they only slow down the
1:31:37aging process. There's I don't think
1:31:39there's anything that reverses that
1:31:41process. There might be some sort of
1:31:43local reversal for a temporary period of
1:31:46time. Uh maybe but it's still kind of
1:31:49trying to you know uh hope that things
1:31:53won't go bad a little bit longer. Like
1:31:56for example, some people can live to to
1:31:58to 100, others only to 60, right? So
1:32:01there's something good about those who
1:32:04live to and in fact there are super
1:32:05centinarians who couldn't make it to 110
1:32:07years old. Very very few people, but I
1:32:10think it's mostly genetics. I mean their
1:32:12lifestyle might have helped a little
1:32:13bit. Uh something about their biology is
1:32:16able to maintain that information much
1:32:19uh much longer that program. So we have
1:32:21to get to the core. what what are the
1:32:24things that are disrupting that
1:32:25information um uh loss? Um and um yeah,
1:32:30it's of course you you have to focus on
1:32:32the on the genome because that's that's
1:32:34sort of the blueprint. It's not just
1:32:36that. It's sort of what affects you
1:32:39afterwards, you know, that your your
1:32:41microbiome, your um metabolites, you
1:32:44know, how those things are changing,
1:32:46whether accelerating or uh reversing,
1:32:49you know, like and it has to be kind of
1:32:51an engineering approach as well, like
1:32:53you know, the skin aging is is a very
1:32:56different problem than immune aging,
1:32:58than the brain aging, right? So, uh your
1:33:01skin cells are constantly renewing. So
1:33:03all you have to do is to have sort of
1:33:06the uh programmed stem cells to go in
1:33:09there clean the environment sen cells
1:33:12and get it get it regenerated and
1:33:14produce collagen whatnot but the brain
1:33:16is not like that right so you don't you
1:33:17don't want to regenerate your your
1:33:19neurons uh you will lose your identity
1:33:21so they have to be dealt in a different
1:33:23different way some of it will be I think
1:33:26for the younger population u it seems
1:33:30like you know redesigning certain
1:33:32biology ology would be sounds radical
1:33:35but it would be uh more foolproof right
1:33:38so what if we could change the genome
1:33:42through genetic engineering like we add
1:33:45certain genes or we change certain genes
1:33:47such that the DNA damage um is checked
1:33:51you know much much longer it you know
1:33:54because there are in fact certain
1:33:55animals who have better DNA damage
1:33:58proteins they kind of evolve to do that
1:34:01like elephants rarely get cancer, right?
1:34:04Because they have this gene called P-53.
1:34:07They have multiple copies of that. P3 is
1:34:09kind of like the guardian of the genome.
1:34:11You know, it prevents the genome from
1:34:13getting too much mutations and prevents
1:34:15cancer. So, somehow elephants have pre I
1:34:19don't know how many copies, but they get
1:34:20very rarely cancer. Um, naked mole rats,
1:34:24you probably know that very well. Um,
1:34:27you know, they they're they're like
1:34:28rats. They live underground but normal
1:34:31rats live a couple of years and these
1:34:32guys live 30 40 years. So it turns out
1:34:35they have some mutation in some immune
1:34:38gene called seag gas that's also
1:34:40involved in immune optimization and DNA
1:34:44repair just like you know one or two
1:34:46genes make a huge difference. So can we
1:34:50uh engineer humans to uh block that
1:34:54degradation of of information uh uh for
1:34:58for those who have already had the
1:35:00damage then we're going to have to to
1:35:02think about repairing that reversing it
1:35:06and then maintaining it. Uh that's
1:35:08that's going to be a bit more
1:35:09challenging but uh uh we'll we'll get to
1:35:12that too.
1:35:13>> What do you think about so the gene
1:35:15going to gene therapy? There's obviously
1:35:17gene editing, gene therapy, and and um
1:35:21right now we only know what we know,
1:35:23right? Again, like with these longevity
1:35:25genes we know about, but do you think
1:35:28that that AI is going to be able to help
1:35:30us analyze the human genome?
1:35:34And I don't know what other data sets it
1:35:36will need but we'll give it everything
1:35:37and help us figure out well actually
1:35:40there's interaction of these genes
1:35:42together and when there you know like
1:35:43all these combinations is that something
1:35:45that you think is going to happen? We'll
1:35:47actually figure out there's a lot more
1:35:49to this equation than we originally
1:35:51knew.
1:35:52>> Yeah. That that's the critical problem
1:35:55because we know what all the genes are
1:35:57in the genome. like we have we have it
1:36:00decoded completely and then we pretty
1:36:02much know their functions most of them
1:36:05uh even if you don't know every single
1:36:07gene involved in aging we know a lot of
1:36:09them the problem is that different gene
1:36:13uh uh first of all can create different
1:36:15proteins you know there's all that
1:36:17splicing that happens and and so on but
1:36:20but even we doubt that in a different
1:36:22context so if you the same protein uh
1:36:26can kill a cell or causes survival like
1:36:29in immune system we have these receptors
1:36:32called TNF receptors or whatever they
1:36:34can they can have a survival signal or
1:36:36or a death signal suicide signal
1:36:38depending on the context of the of the
1:36:40cell. So that is very very uh critical
1:36:43that how as you pointed out how these
1:36:46genes and proteins uh in a network
1:36:49fashion in a sort of a topological
1:36:51network uh
1:36:54you know what do they do like if I
1:36:57interfere like these uh um probably
1:37:00we'll talk about that these things
1:37:02called Yamanaka factors where you can
1:37:04you can generate a stem cell from a
1:37:06normal cell right so like complete
1:37:08regeneration uh uh but But the the
1:37:11problem is that they can also cause
1:37:14cancer because they only need to be
1:37:17active in certain time. If they're
1:37:18active all the time, they can cause
1:37:20terteratomomas and things like that. So
1:37:22that that part is so complex that we
1:37:26absolutely going to need AI to simulate
1:37:29that for us. If I have this gene in the
1:37:33context of all the other things at
1:37:36certain age with these epigenetic
1:37:39programs plus all the metabolites and so
1:37:42on because those are constantly
1:37:44signaling the cell and you know doing
1:37:47letting the the proteins do something
1:37:49and so on. What would happen if I
1:37:51interfere with that particular gene or
1:37:54how can I improve that? Uh if if you
1:37:57have a because you have to consider the
1:37:59other genome too like your gene therapy
1:38:01might be very different than somebody
1:38:02else's because you might have some great
1:38:06genes that are synergistic with that
1:38:08other person might have not so great
1:38:11genes if even if you try to improve it
1:38:13that that would actually work or it
1:38:15wouldn't it wouldn't help. Um so uh it's
1:38:19just a matter of complexity. there's so
1:38:21much information that uh the AI has to
1:38:25not only put that together but have sort
1:38:28of almost a temporal simulation of the
1:38:31model like that's a very important point
1:38:33the because right now the models are
1:38:36kind of static they they have a good
1:38:39understanding but they don't know what
1:38:41would happen if a cell comes next to a
1:38:44tumor just 2 minutes earlier they the
1:38:48cell next to it what that context
1:38:51affects there's a behavioral issue. It's
1:38:55the same problem with the robotics,
1:38:56right? So, um kind of the physical
1:38:59intelligence or the biological
1:39:00intelligence once those models are
1:39:03evolved with with a lot of data. I think
1:39:06we'll we will be able to simulate this
1:39:08and and AI will will be able to decide
1:39:11this is the gene therapy you should get.
1:39:13So, you need a new copy of immune
1:39:15system, but let me design it for you.
1:39:17It's it's so exciting because not only
1:39:19are we talking about, you know,
1:39:21extending our lifespan and curing
1:39:23disease, but we're talking about like
1:39:26getting rid of side effects in a way. I
1:39:28mean, you know, people all respond to
1:39:30different foods and treatments and
1:39:33everything differently, right? That's
1:39:34why some people have a terrible response
1:39:36to perhaps maybe a vaccine
1:39:38>> um and others don't. And so, it's really
1:39:40exciting to think about that. Um,
1:39:42>> which which I, by the way, call human
1:39:442.0. And maybe we'll get to human 3 3.0
1:39:48uh which which will happen at this bios
1:39:50singularity moment. What that means is
1:39:52that you know we we kind of re-engineer
1:39:54ourselves. Um uh I always think about
1:40:00like most most scientists or most
1:40:02doctors think like what's wrong with
1:40:04this person or patient. Uh I always
1:40:06think the opposite. There are certain
1:40:08people I'm saying what's right about
1:40:11them? like this person is has smoked for
1:40:1450 years, never got a lung cancer or you
1:40:18know had a terrible diet or whatever.
1:40:20This one lived to be 110 for you know
1:40:23whatever reason. And so what is good
1:40:26about those people? Why can't we take
1:40:29what's good about all of those people
1:40:31and then re-engineer those that are not
1:40:34so lucky to be born with what's what's
1:40:36so good and then you know even make it
1:40:39better. So that's the human 2.0,
1:40:41>> right? I I I mean that's exciting to me
1:40:43as well, right? I mean we do know like
1:40:45you said we can live humans are capable
1:40:48right now of living to be is the whole I
1:40:50think the oldest was like 121 maybe
1:40:53>> 123
1:40:55French woman. I mean
1:40:57>> the fact that that right now in 2026 we
1:41:00know that humans can at least live to be
1:41:02123
1:41:03>> is exciting. 115 I mean at 115 116
1:41:07that's considered sort of the current
1:41:09limit but you know only 300 people in
1:41:12the world are 110 and older why is that
1:41:16why not the rest of the 8 billion
1:41:18>> right yeah it's it's fascinating and I'm
1:41:21I'm so excited for you know having this
1:41:24super computing power to help us figure
1:41:26that out what did you think when
1:41:30you know the Yamanaka factors were
1:41:32discovered by Shina Yamanaka and all of
1:41:34a sudden you could take this old cell
1:41:36and completely rever reverse it to
1:41:38revert it to an you know essentially
1:41:41induced you know pur potent stem cell.
1:41:43Do you remember like is that was that
1:41:44something did aging come into your mind
1:41:46at that point where you were thinking
1:41:47well that's the youngest almost you
1:41:49could get I mean
1:41:51>> yeah uh of course uh in fact at the time
1:41:54I was um part of some aging groups uh uh
1:41:58I think like an hour after the paper was
1:42:01published I was you know typing there
1:42:04you know like this is this is it this is
1:42:06amazing so I I should say that there
1:42:08were two um moments for me uh uh that
1:42:12that I thought that aging was was going
1:42:16to be uh reversible or curable, however
1:42:18you call it. Um kind of like the
1:42:20chachipit moment of biology. The first
1:42:23moment was uh the um the sheep uh that's
1:42:27called Dolly. Uh you probably know it
1:42:29was the first cloned ship sheep. Um it
1:42:34was 19967
1:42:36or something like that. I can't remember
1:42:37the exact date, but it was in '90s. And
1:42:40um so basically uh the um uh the
1:42:45scientists took a cell from uh you know
1:42:48uh from one sheep and then recreate
1:42:51exact copy of that sheep you know by by
1:42:53cloning it. Uh it was it was at the
1:42:56embryo level but it was sort of like
1:42:58exact copy of it. So that means that
1:43:01there was enough information that you
1:43:04could just like uh recreate the same
1:43:07person again and again and again. Right?
1:43:09And then the second of course uh uh the
1:43:11the Yamanaka factors uh in 2016 I think.
1:43:15Um and that was the moment that uh that
1:43:18we knew um that we could completely
1:43:23erase the um sort of the age of the cell
1:43:27on a cellular level and then bring it
1:43:30back to a purotinent stem cell level and
1:43:32then use that to recreate the whole
1:43:35biological organism. So, so it means
1:43:37that we have unlimited supply of
1:43:40regenerative capacity like it's there is
1:43:43there's there's no limit to it. In fact,
1:43:45we already know that like so our our DNA
1:43:47just keeps for billions of years. It
1:43:50keeps going on and the fact that you
1:43:52could do that in the lab and you could
1:43:54you could generate it was was was
1:43:57amazing. Um uh but of course the the
1:44:00problem was okay so then how do you
1:44:03apply that? In fact, I I think there was
1:44:06just a recent study that started in
1:44:08Japan using the Yamanakica factors uh uh
1:44:12in in clinical trials because you know
1:44:15it was not a very controlled system like
1:44:17you didn't know if those cells would
1:44:19develop tumors you know in mice they
1:44:22they did some of them tumor tumors you
1:44:24know whether um you can control them or
1:44:28importantly I think there's going to be
1:44:30a trial started by David Sinclair soon
1:44:33can we do like partial reprogramming
1:44:36because most of the time you don't want
1:44:38the plur potent cell all right you just
1:44:40want your skin cells to go early enough
1:44:44to their sort of more stem-like level
1:44:46like I work in immune system and for us
1:44:49um I can divide like immune cells into
1:44:52naive memory and aector and
1:44:54differentiated so the naive cells are
1:44:57kind of the young guys they have huge
1:44:59potential to expand and and make memory
1:45:03and and affect the population and the
1:45:05other ones constantly um die and get
1:45:08older. Can we actually revert the cells
1:45:11towards the naive? And I I actually
1:45:13spent a long time trying to do that. Um
1:45:15so maybe this partial programming will
1:45:18will will will enable that and and
1:45:20that's that will be uh revolutionary
1:45:23because uh then you can if you can also
1:45:26deliver those then you can make most of
1:45:29your old skin cells turn into a younger
1:45:32version. I think the trial is going to
1:45:34be for I uh with David Sinclair. Um
1:45:37yeah. So uh but again it's it's um the
1:45:42these these things showed us that uh we
1:45:46can reverse aging. But when people say
1:45:49oh that's impossible like this is this
1:45:51you can't you can't reverse aging like
1:45:53you know this entropy whatever. Um but
1:45:56we we we do it in the lab all the time.
1:45:59Uh why not do it in a in a total
1:46:01organism level? So with this partial
1:46:03cellular reprogramming as um as you
1:46:06mentioned you know you're you're
1:46:08basically taking an old cell and putting
1:46:10these four different proteins I think
1:46:12they can do it with fewer now but
1:46:14putting them on for a shorter period of
1:46:16time on the cell and that it's changing
1:46:18the the epigenetic program and in a way
1:46:21that it's still the cell keeps its
1:46:23identity. It doesn't become a stem cell
1:46:25but it seems to be more youthful. Um, I
1:46:28know there's been some work and I
1:46:29haven't followed all this literature
1:46:31since I the first, you know, some of the
1:46:34first studies that came out, but I think
1:46:36it was like Juan Carlos, um, Epizusa, he
1:46:40he's now, I think, at Altos Labs, but he
1:46:41at the time was at the Sulk Institute
1:46:43>> and, um, he had done this in in mice. I
1:46:46think they were even maybe perhaps
1:46:47progeria mice or some sort of
1:46:49accelerated aging model
1:46:51>> and there was some reversal of you know
1:46:53certain organs seemed to be rejuvenated
1:46:56in a sense and um the the life
1:46:58expectancy was extended in those animals
1:47:02but what's interesting is that not all
1:47:05of the 12 hallmarks of aging go away.
1:47:09>> Yeah.
1:47:09>> Right. And so you would hope that you
1:47:12would reverse aging totally but there's
1:47:14genomic you know somatic mutations are
1:47:17still there I think tieumir don't get
1:47:20mitochondria so
1:47:21>> do you think first of all I don't I I'd
1:47:24love to understand why that is so what
1:47:27is it if you're if you're essentially
1:47:29you know wiping out the epigenetic
1:47:32current epigenetic program and and
1:47:33reverting it back um why does not
1:47:36everything change I don't know if you
1:47:38have any ideas But do you think AI is
1:47:40going to help us understand that?
1:47:43>> Uh definitely. I mean we I should also
1:47:46point out that we um we do need to
1:47:49generate lots of data. So so I think um
1:47:53you know when whenever I talk about AI
1:47:55um people say okay well why can't AI do
1:47:58it now? Um for two reasons. One is that
1:48:01we don't have enough data. So we we we
1:48:04probably know maybe 10 20% of all the
1:48:06biology. we still have lots of data to
1:48:09to generate. The second is the
1:48:11>> you're talking about scientists.
1:48:12>> Yeah. Scientists or or automated lab
1:48:15whatever it is. Um so I mean right now
1:48:18we're able to generate millions of data
1:48:20points in one experiment you know and
1:48:22but but even that's not enough like we
1:48:24need to generate billions of data points
1:48:26and so on. So but of course to handle
1:48:29that we also need um super intelligence
1:48:32and supercompute. So, we have to have
1:48:35compute that's thousands of times than
1:48:37what's available. And people say, okay,
1:48:39well, you know, why are they building
1:48:40all these data centers? Isn't this
1:48:42enough? And so on. Well, we're going to
1:48:44need it. If you want if you want to cure
1:48:46all diseases and reverse aging, we're
1:48:48going to need probably we're going to
1:48:50need data centers in the space and and
1:48:52and and lot more because so much data
1:48:56has to be in real time sort of uh uh
1:48:59simulated. um uh and we might get much
1:49:02more efficient doing that as we learned
1:49:04algorithms. So so that's that's one
1:49:06issue. The other is that um as you
1:49:09pointed out something very important I
1:49:11mean this partial reprogramming or total
1:49:13reprogramming they're super exciting but
1:49:15they don't solve um they don't
1:49:17completely solve the the aging problem.
1:49:20They will um make your um eyes see
1:49:24better for a certain period if you're 80
1:49:28years old or your skin gets better. Um
1:49:32but will it work on your um your heart
1:49:35muscle uh or on your brain cells neurons
1:49:40which is the critical point because if
1:49:42you can have a perfect body but if your
1:49:44brain is aging then then that's it. Um
1:49:47so will it modify the sort of the
1:49:51microbiome that has now the environment
1:49:54of an old person because if if that
1:49:57happens if your metabolism is in old
1:50:00person's metabolism and and microbiome
1:50:02is old person's metabolism and your your
1:50:04DNA has accumulated a bunch of mutations
1:50:06and mitochondria has bor mutations you
1:50:09can reverse that a bit have some
1:50:14regenerative capacity but they will
1:50:16quickly
1:50:17become old again, right? You know,
1:50:19because the environment is not is not
1:50:21great, right? So, like if you live in a
1:50:23bad neighborhood and you created this
1:50:26beautiful house, you know, it's but it's
1:50:28very bad neighborhood, your house is not
1:50:30going to last very long there. So, your
1:50:32your neighbors has to be clean as well.
1:50:35So, I think it's it's a great thing and
1:50:37that's probably going to add certain uh
1:50:40years to lifespan and the quality of
1:50:42life uh for sure. uh but we we have to
1:50:45push that much much further um and then
1:50:49really understand whether it's 12
1:50:51hallmarks actually I asked JPT recently
1:50:54came up with another four or five
1:50:55hallmarks
1:50:56>> what were they
1:50:56>> I I can't remember exactly uh it was one
1:50:59of them was related to immune system I
1:51:01just this was recently um uh but yeah it
1:51:06was it was it was quite interesting um
1:51:09trying to remember one had to do with
1:51:10metabolism
1:51:12um uh uh you know because we we we kind
1:51:15of classify hallmarks based on what we
1:51:18can measure and see and I think AI can
1:51:20see a little bit more than we can. So
1:51:23anyway um this is going to be uh a
1:51:27serious engineering uh problem. I I
1:51:30would be very surprised if we have like
1:51:33one pill you take and then you suddenly
1:51:36become young again. That's that seems
1:51:38very unrealistic to me.
1:51:40>> Yeah. I mean you know and then the other
1:51:41question is like in the lab we're we're
1:51:43the way we're delivering these
1:51:45treatments is like an adino virus right
1:51:47and and then it's like well is that
1:51:50going to cause cancer because they virus
1:51:52go to right cell
1:51:53>> is it going to go to the right cell
1:51:54exactly I mean the there's definitely a
1:51:56lot of engineering
1:51:57>> we we have to develop uh so one of the
1:51:59things that I like doing with AI models
1:52:01is to develop some new methods new new
1:52:04technologies they have a bit too much
1:52:06guard rail so uh they don't allow me to
1:52:09to go too deep in it but you know
1:52:11because uh I don't think we have we have
1:52:13enough tools like of course we have
1:52:15crisper now but actually Dudana's lab
1:52:19just came out with something even better
1:52:20for bacteria for genome editing so
1:52:23imagine there's there's there's probably
1:52:24all kinds of other tools that we can
1:52:26build that will make this localization
1:52:29the editing much more perfect and and
1:52:33has to be programmable you have to
1:52:35literally create circuits we can program
1:52:38immune cells in in in culture like we
1:52:40can give a a drug it will shut down
1:52:42their response or we can create end or
1:52:45gates and not gates if they see two
1:52:47molecules then they respond if they see
1:52:48one they don't like you can literally
1:52:50program the biology so we have to
1:52:53develop these new tools that are better
1:52:55than viruses maybe uh generate lots of
1:52:58data sets um and they manipulate the the
1:53:02organs and so on could be that for some
1:53:05organs when they're too old it might
1:53:08might be just too difficult to repair
1:53:11them. So you might consider just putting
1:53:13a new one,
1:53:14>> you know, like it might be a point of no
1:53:16return, your your kidneys or whatever.
1:53:19Then you'll have these u u organ
1:53:22factories which 3D printed and actually
1:53:26>> uh we we did a lot of collaboration with
1:53:28a colleague of mine, you know, he can
1:53:30print, you know, small tissues, lungs
1:53:32and and pieces like that. So some of
1:53:35them will be kind of transplanting new
1:53:37organs. Some of them will be
1:53:39pre-engineering and
1:53:40>> and then the digital twin the analysis
1:53:42and simulation will be able to figure
1:53:44out is are you going to reject this or
1:53:46would you need to not reject it?
1:53:48>> That's right.
1:53:48>> Right. Um what do you think of the new
1:53:51data that came out using this this model
1:53:54called GPT micro 4B
1:53:57GPT micro 4B? um where I guess there's
1:54:01this model that was used to figure out
1:54:05how to make certain mutations in the
1:54:08four different Yamanaka factors to make
1:54:10them more effective. So they were able
1:54:12to basically 50fold more um be more
1:54:16effective or efficient at increasing
1:54:18this induced pur potency.
1:54:20>> Yeah.
1:54:20>> How do you interpret that data?
1:54:22>> So I I don't think that model uh is is
1:54:26any better than what we have right now.
1:54:28uh probably um the current models are
1:54:30are much better. Um the I think probably
1:54:34there might have been two two
1:54:36differences and I don't know all the
1:54:37details but one is that they probably
1:54:40remove the the guard rails because
1:54:42there's a lot of biocurity guard rails
1:54:44in in the current models. Um if you ask
1:54:48the same question to GPT5.5
1:54:50it will refuse to do it. It will say oh
1:54:52this is a biohazard like what if you
1:54:54mutate and create a new virus or new
1:54:56cancer whatever. So that might be one
1:54:58reason and then the other is like if you
1:55:00let these models think longer. So like
1:55:03GPT5.5
1:55:05pro and and the thinking and in model is
1:55:08the same pre-training but pro model can
1:55:12think two hours thinking can take two
1:55:14minutes. Uh so the longer they can think
1:55:17the the more they can iterate. They can
1:55:20run these scenarios again and again and
1:55:22again. So my speculation is that that
1:55:25model probably ran for for a long period
1:55:28of time. Of course you need a lot of
1:55:29compute and a lot of tokens not a
1:55:31problem for open AAI. um uh then you you
1:55:35will probably come up with the solution
1:55:37that even a a more intelligent model
1:55:40couldn't come up in a in a shorter
1:55:42period of time because that that
1:55:44particular case is really running
1:55:47experimental scenarios like okay if I do
1:55:50this mutation what would be the
1:55:51potential outcome like it's running all
1:55:53the simulation oh yeah okay so so what
1:55:56if I change that mutation to here and
1:55:57then what if I add another mutation and
1:55:59running the experiment again and again
1:56:01again so you you're constantly making
1:56:03the the solution better and better and
1:56:06better as as you think longer. Um so uh
1:56:10and and this will get better. So if if
1:56:12you have much more compute, much more
1:56:14intelligence and you say okay um GPT7 or
1:56:19six whatever is go and think for a
1:56:21month, you know, find the perfect
1:56:25molecule that will bind to this receptor
1:56:27and this will cause that. It'll it'll
1:56:29probably figure that out. What is it? It
1:56:32sounds like we're going to need to do a
1:56:33lot of this type of simulation and by
1:56:35were I mean researchers and scientists.
1:56:38What is it going to take to remove some
1:56:40of those guardrails in that environment
1:56:42for researchers to be able to to make
1:56:45these new discoveries and and what sort
1:56:48of I guess I mean how how do we protect
1:56:50from a a new crazy
1:56:52>> biohazard or you know biosafety issue?
1:56:55Well, I mean I think like OpenAI is
1:56:57partnering uh with with um you know
1:57:00trusted people. Uh so you have to be
1:57:03approved by them. So I think then uh
1:57:05whether it's a company or something like
1:57:07that. It's the same problem with uh with
1:57:09cyber security, right? So Anthropic has
1:57:12this new model called mitos and they
1:57:14decided not to release it because they
1:57:16said it's too dangerous for cyber
1:57:18security because this this model can
1:57:21just crack into any can find all these
1:57:24things that that other uh others cannot
1:57:26see. So, in fact, even the governments
1:57:29thought that that was important that
1:57:31they should uh I don't know if they're
1:57:33exaggerating if it's if it's if it's
1:57:35true or not, but so you have to put that
1:57:37guard rail if you release it to the
1:57:39world because somebody can use that
1:57:41model and then hack into your bank
1:57:43account or somebody can use it to create
1:57:46a a a new virus gene or something like
1:57:49that. So I think there you know that
1:57:51will be made individual purses or
1:57:56institution
1:57:57basis that these these um uh hopefully
1:58:02these AI companies will share that
1:58:05because they might decide not to share
1:58:07it. Uh might say well okay why don't we
1:58:10just develop all the drugs internally
1:58:13and not release any of these models. um
1:58:16uh some some might be doing that for
1:58:19example I don't think that would be a
1:58:21good thing because uh what you really
1:58:23need is again as I said you need a lot
1:58:26of data you need a lot of scientists
1:58:29uh putting all that data into the models
1:58:32but not only the data but their
1:58:34experience in a way you in in the let's
1:58:37call it the wild or or the world you're
1:58:40you're actually training those models
1:58:42even even if it's super intelligence
1:58:44it's going to
1:58:45so hungry for data that you're going to
1:58:48have to um collaborate or release it to
1:58:51to to others. Um also I think this will
1:58:55be important to democratize uh health
1:58:58care because one question everybody
1:59:00asked okay well you know if you find the
1:59:03the treatment for aging this is only
1:59:05going to be available for the super rich
1:59:07I'm never going to be able to afford it
1:59:09or or treatment for cancer I say the
1:59:12opposite actually thanks to AI it will
1:59:15be super affordable because if you can
1:59:17create a drug like in a startup let's
1:59:20say cannot compete with a big pharmace
1:59:22to a company they can find a drug uh
1:59:25using AI 100 times cheaper and if you
1:59:28can do the clinical trial using digital
1:59:30twin that's where all the money goes
1:59:32like you could develop a drug for a
1:59:33couple of million dollars rather than a
1:59:35couple of billion dollars so the cost of
1:59:37drug development or treatment uh
1:59:39development will be magnitudes lower and
1:59:42that will give uh a huge number of
1:59:44people u access to that uh but of course
1:59:47you know AI AI has to be um shared. It's
1:59:53it's uh I think it's it's a product of
1:59:56all humanity and it should be the
1:59:58possession of all humanity. That's how I
2:00:00view it
2:00:01>> except for going back to the the thing
2:00:03that you mentioned at the beginning of
2:00:04this podcast which is that you know
2:00:06humans in the wrong hands that is the
2:00:09problem and that's and that's that is
2:00:11something that needs to be very taken
2:00:13very seriously.
2:00:14>> But but the the solution to that is also
2:00:16AI. So right now, I mean, I hear that
2:00:19like MTOS um basically finds all these
2:00:23loopholes in in this cyber security uh
2:00:26issues that people couldn't figure out
2:00:29for decades. They didn't even know they
2:00:31existed. So it's just patching all those
2:00:34uh uh all these security bugs. So it
2:00:38will create almost a perfect secure
2:00:41systems like it will be unhackable
2:00:44because uh MTOS is actually preventing
2:00:47that. So to to prevent that from
2:00:49happening you still need AI. You might
2:00:51still have some bad actor trying to
2:00:54develop a virus that will cause a
2:00:57pandemic. To prevent that you also need
2:00:59AI. So the AI should be able to predict
2:01:02it and already create the vaccine ready.
2:01:05will say, well, somebody might make this
2:01:07virus, so let's let's get ready for it.
2:01:09Um, so, uh, AI is the solution to all
2:01:13that.
2:01:13>> Interesting perspective. Always seems to
2:01:15you always seem to have a positive
2:01:16outlook. Um, I wanted to ask you another
2:01:18question about, you know, we're talking
2:01:20about these simulations and how we're
2:01:22going to, you know, using AI to to
2:01:24essentially run these clinical trials
2:01:26cheaper because we're going to do this,
2:01:28you know, these simulations and have,
2:01:30you know, biomarker data and it'll just
2:01:32be, you know, shorter and and cheaper
2:01:34and easier. The question is always what
2:01:38do you measure, right? What is the
2:01:39biioarker? What are what's the end
2:01:41point, right? And in aging, you can now
2:01:44see I mean every a study almost a new
2:01:47study every day coming out looking at
2:01:49these epigenetic aging clocks. And
2:01:52that's you know the the so as most
2:01:57people listening to this podcast know
2:01:58I've had Steve Horbath on um a couple of
2:02:00times and he's sort of the pioneer in
2:02:02these epigenetic aging clocks and
2:02:04they've now developed over you know the
2:02:06last decade or so and become much more
2:02:08of a biological marker of age like your
2:02:10biological age not just to be able to
2:02:12predict your actual chronological age.
2:02:15And so, um, you'll find now studies that
2:02:17are looking at treatments and whether or
2:02:19not it can reverse quote unquote reverse
2:02:21biological aging or epigenetic aging,
2:02:25but it's not clear that that's
2:02:29necessarily,
2:02:30you know, if that's
2:02:33really reversing aging, right? So, how
2:02:35what do what do you think um from your
2:02:38perspective, what should we be looking
2:02:40at in terms of some of these functional
2:02:43>> outputs? Um yeah, I mean the those
2:02:46epigenetic u markers are very useful um
2:02:51but I don't believe that um they are
2:02:55terribly useful as um sort of as as
2:02:58predicting true aging. I mean there's
2:03:01there's a very um uh significant problem
2:03:04with with those markers. uh usually
2:03:07they're they're done through through
2:03:08blood analysis
2:03:10but in the blood you have u like you
2:03:13know I work with te- cells so you have
2:03:15these cells that we call aector cells
2:03:19that have um lots of epigenetic change
2:03:22because they differentiate it and they
2:03:24continue to accumulate in in old age and
2:03:27then you have these naive cells that
2:03:28have you know more pristine uh kind so
2:03:31it's a combination so depending on um
2:03:34what that combination is is going to
2:03:37affect the output of of the um so you
2:03:40you you can actually just look at the
2:03:42proportion of your uh T- cell
2:03:44differentiate T cells you'll probably
2:03:46get the same same kind of information um
2:03:49and it doesn't tell you like what's
2:03:51happening in the skin or the brain or
2:03:53the heart you know that it doesn't mean
2:03:55that uh if if the immune cells are
2:03:58getting younger or the young ones are
2:04:00expanding and the old ones are dying
2:04:02that doesn't mean that your skin is
2:04:04getting younger or your liver is getting
2:04:06younger. So that it has a very limited
2:04:08use in my opinion. But we really don't
2:04:12need that because like aging is probably
2:04:15the easiest way to measure. We know
2:04:19exactly what goes wrong in in in old
2:04:22age, right? So like you can't breathe
2:04:25that well. Your heart doesn't work that
2:04:28well. Your muscles don't work. you can
2:04:32only, you know, raise so much because it
2:04:35your weakened muscles or your VO max is
2:04:38is is lower. Um, these are all
2:04:43phenotypic like you don't even have to
2:04:45probably withdraw a blood just measuring
2:04:48the ability of uh of an elderly person.
2:04:52Uh, can they walk uh better, you know,
2:04:56100 meters than they used to? like
2:04:59because that's looking at the total
2:05:01biology like you know your cells your
2:05:03metabolism or whatever uh muscle to me
2:05:07that's or or your cognitive abilities
2:05:09>> but those can't be simulated I mean
2:05:12>> the they eventually they can be right
2:05:17now they can't they can't be simulated
2:05:19uh um because as I mentioned the AI is
2:05:24missing that behavioral physical
2:05:26intelligence in the real world because
2:05:29that's a that's most things are
2:05:31happening in real life. uh but um I
2:05:35think I think they can be simulated but
2:05:37more importantly uh I think eventually
2:05:40you have to tr whatever the AI comes out
2:05:42with you need to try it on on the humans
2:05:44right so uh my point is that you don't
2:05:48have to uh do anything too fancy or wait
2:05:53decades to see the effect if I give this
2:05:56treatment to um I don't know 80 year old
2:06:00and they're suddenly able to breathe
2:06:03Well, you know, their VMX went up. Um,
2:06:06they're sharper, they can think better,
2:06:08uh, they can remember better. Um, you
2:06:12you can look at their immune system and
2:06:13we can see that the cells are we know
2:06:16which cells are younger or worse. Or you
2:06:18can look at their skin like, oh wow, the
2:06:20skin is getting young. Like you see it,
2:06:22you don't even have to do anything. Um
2:06:25so so there are so many features
2:06:26phenotypic features of aging that could
2:06:29be um objectively measured actually and
2:06:33not just subjectively you will see the
2:06:35effect very very quickly like this
2:06:38partial reprogramming trial they're
2:06:40doing it's it's done for glaucoma
2:06:42patients I I guess uh because that
2:06:45happens in old age right so your your
2:06:47cells are aging so I mean if these
2:06:50people start to see it works right their
2:06:53their cells cells got regenerated. Um
2:06:55you don't need to look at the
2:06:56epigenetic. Um so I think uh it will be
2:07:00a combination of those um measurements
2:07:03probably we will come up with and AI
2:07:06will probably come up with this set of
2:07:08biomarkers. I don't think we know
2:07:10because it's going to be a set of
2:07:12biomarkers like um you know your glucose
2:07:15your cholesterol might be high when
2:07:18you're 30 and it will be high or low
2:07:20when you're 80. I mean there's not a
2:07:22very specific marker that will tell you
2:07:24your your age for just looking at that.
2:07:26But the combinatorial effect will will
2:07:30AI probably will be able to predict your
2:07:32age looking at all kinds of data sets
2:07:35and say oh this guy must be uh you know
2:07:38um 52 years old based on this. You know
2:07:40>> I know we have uh that model clock base
2:07:42that's looking now at a variety of small
2:07:46molecules that might reverse epigenetic
2:07:48aging.
2:07:49Now there are some data sets showing
2:07:51that you if you reverse epigenetic aging
2:07:54there is some functional correlation
2:07:56with some functional improvements like
2:07:58pre-frailty things like that you know
2:08:00like improve but um it at the end of the
2:08:03day you know I think it it'll be
2:08:05interesting to see if there's going to
2:08:06be companies that come out trying to
2:08:08sell some sort of drug to claiming it
2:08:11reverses aging when they're really just
2:08:13looking at one
2:08:14>> biioarker which is reversing
2:08:16>> it's most as I say it's mostly the
2:08:18immune aging that they're looking at or
2:08:20or sort of maybe getting rid of the
2:08:23terminally differentiated immune cells
2:08:26like for example in old age you you
2:08:29accumulate these CME specific tea cells
2:08:32um CMV is a virus that you can't really
2:08:35get rid of so the immune system
2:08:37constantly have to keep it under check
2:08:40um and those immune cells they kind of
2:08:43become like missionaries they should
2:08:44retire but they keep on expanding and
2:08:47some indiv individuals might have like
2:08:4920 30% of all their tea cells just
2:08:51dedicated to like one peptide of this
2:08:54this CMV and they're they're not helpful
2:08:57but they become uh harmful because those
2:09:00guys are are old they should retire they
2:09:03don't and they cause inflammation
2:09:05because they're they're active and um
2:09:07and they don't give place for the young
2:09:09guys to come in uh and they're they are
2:09:11epigenetically you know closed because
2:09:14um they're differentiated their
2:09:15telomeres are shorter So, uh, you know,
2:09:19you might be getting rid of some of
2:09:20those cells with certain treatments,
2:09:22which is great. Um, but then you have
2:09:25the indirect effects, right? So, if you
2:09:26can if you can control the immune system
2:09:28and inflammation, that's going to have
2:09:31huge effect all over your that doesn't
2:09:33mean your skin got just regenerated, but
2:09:36it it will it will help clean up
2:09:39>> aging. Yeah. Yeah. Exactly. Um, also the
2:09:43other thing I was thinking about is
2:09:46like, you know, we you're mentioning V2
2:09:47max and, you know, muscle strength,
2:09:49muscle mass. We have all these markers
2:09:51that sort of like decrease with age and
2:09:54yet we don't know necessarily that they
2:09:57cause aging in a way. So the question is
2:10:00like will AI be able to take all this
2:10:03correlational data like we have all this
2:10:05you know all these different functional
2:10:07out you know endpoints that we look at
2:10:09and and be able to differentiate it from
2:10:11like personalized you know this
2:10:14personalized um data set versus like
2:10:17actually like how do you cure aging like
2:10:20what do you change that's going to drive
2:10:22you know reverse the aging I mean
2:10:25there's there's a lot of questions Um,
2:10:28you mentioned something interesting that
2:10:30had to do with the brain and that is
2:10:32something that I've been thinking about
2:10:34as well because you know we we have a
2:10:38lot of repair processes in our body
2:10:39right we can repair a lot of DNA damage
2:10:41and you know mitochondrial function and
2:10:44you know all these things but in the
2:10:46brain we can grow new cells replace the
2:10:48old cells in the brain it's not as
2:10:50robust right there's some parts of the
2:10:53brain that can um you can grow new
2:10:55neurons neurogenesis there's neurop
2:10:57plasticity. That's a big part of of the
2:11:00repair process in a way, but it's not
2:11:02like a big you you're not you're not
2:11:06totally replacing the brain and you
2:11:07don't want to, you know, as you
2:11:09mentioned because then memories go away
2:11:11and your identity and you know, it gets
2:11:13very complicated.
2:11:15>> Um, how do you see AI
2:11:18>> intervening in that? Like everything's
2:11:20great if we can reverse our heart aging
2:11:21and all this, but our brains that's so
2:11:23important
2:11:24>> now. Is it just going to be a, you know,
2:11:26delay age related disease,
2:11:29neuroinflammation, all that stuff? We
2:11:30can we can fix that, but like are we
2:11:32going to be able to really
2:11:35reverse brain aging?
2:11:37>> Um, you know, I I would have to ask AI
2:11:40to to to figure that out. But, you know,
2:11:42I can I can think of several scenarios
2:11:45how that might happen. uh first of all
2:11:47you know neurons um or the brain overall
2:11:51must have some very good maintenance
2:11:54policy right so so there are neurons
2:11:57that live for decades maybe 70 80 years
2:12:00and not just neurons but there are other
2:12:02cell types that can live for very long
2:12:04they don't divide very much there is
2:12:06some regeneration uh it's not like zero
2:12:09and that's very important because that
2:12:12means that if you let's just do a total
2:12:15experiment let's just say that you
2:12:17replace 0.01% of your neurons uh every
2:12:21month or every year something like that.
2:12:23I don't think that's going to make a
2:12:25huge difference in your brain structure
2:12:28because what they're doing is that
2:12:29they're probably you know there's some
2:12:31neurons somewhere interacting with bunch
2:12:33of other neurons synapses and then it
2:12:37gets replaced and the new neurons might
2:12:39have a few other synapses other than
2:12:42that but that's going to replace that
2:12:44network anyway because they have that
2:12:46capability. So if you do this slowly u I
2:12:49think um
2:12:51you you won't lose a lot. In fact, we we
2:12:54still lose memories, right? So, uh we uh
2:12:57we can't remember everything or we
2:12:59hallucinate all the time. Uh talk about
2:13:01hallucination, right? Uh imagine that
2:13:03this happened to me. No, no, no, it
2:13:05didn't happen. No, no, I I I remember
2:13:07that. So, that's like brain uh brain uh
2:13:11maybe part of it is new neurons that
2:13:13they just didn't know. So, they just
2:13:14made it up, right? So, um so that's one
2:13:17thing. The other thing is that uh these
2:13:19neurons probably have some internal
2:13:21abilities to regenerate. What I mean by
2:13:23that is that you know the cell can
2:13:25maintain itself if it has you know sort
2:13:28of um a great way to clean up internally
2:13:32like autofagy is is a very important
2:13:34mechanism as you know um or it has some
2:13:38really special DNA damage correction
2:13:41ability uh like stem cells have that
2:13:43right so pristine stem cells they don't
2:13:46get old you know even in 100 years old
2:13:48they they're still like like a a young
2:13:51person so uh And then you have all these
2:13:54other cells like GA cells and and and so
2:13:57on that are there to prevent all the
2:14:01other stuff that happens the
2:14:03inflammation you know GA cells of course
2:14:05are are are part of the immune system in
2:14:07a way but they they are like the immune
2:14:10system is not allowed into the brain in
2:14:12very rare cases uh uh it's like a
2:14:15protected area um because the immune
2:14:18system causes too much damage and if you
2:14:20can't replace it quickly that's that's a
2:14:22huge problem, but they have their own
2:14:24network of cleaning up and they probably
2:14:26have some sort of like a lymphatic
2:14:28system and and so on. Um um so if we can
2:14:33figure that out or if I can figure that
2:14:35out, we might be able to really maybe
2:14:38not completely regenerate but extend it
2:14:42um quite significantly. Maybe another 10
2:14:4610 years, 20 years, 30 years for
2:14:48whatever. And then we might come to a
2:14:50point and this goes into a little bit of
2:14:52a science fiction now you know uh let's
2:14:54say in 50 years time AI might be able to
2:14:58figure out all of the synaptic
2:15:00connections in your brain like every
2:15:02single neural network and the
2:15:05neurotransmitters and everything else.
2:15:07So eventually you might be able to like
2:15:09literally simulate your your brain um
2:15:12you go into the matrix level. So that
2:15:15might allow AI to like say okay I'm
2:15:18going to replace all these neurons but
2:15:20I'm going to make sure that they
2:15:22reconnect all these signapses
2:15:26>> so so that you don't lose your identity.
2:15:28Um or alternately I can keep a copy here
2:15:32and then we can create a new brain and
2:15:34then transfer to that new brain that
2:15:37exact u uh state uh that I that I found.
2:15:41Uh I I'm not saying that this is
2:15:44possible right now. That's really
2:15:45science fiction error, but you can
2:15:47imagine that at some point we might get
2:15:50to that level. So I'm not I'm not too
2:15:52worried. I I think if it can pass this
2:15:55couple of decades and then keep the
2:15:57brain um healthy and and self-preserving
2:16:00for for maybe uh you know age 120, 130.
2:16:05And in fact, you know, the people people
2:16:07actually who live to to age 100, they
2:16:10they have very sharp minds, right?
2:16:11>> Because if you don't have sharp mind,
2:16:12you don't live very old. So that's like
2:16:15super correlated. So if we can keep it
2:16:18for a couple of more decades and uh
2:16:20we'll probably find some other
2:16:21solutions. So if we can if we can keep
2:16:23the neuroinflammation low, if we can
2:16:26increase brain drive, neurotrophic
2:16:28factors, some of these things that we
2:16:30know does play a role in improving
2:16:31neuroplasticity and
2:16:33>> you know and growing new neurons and to
2:16:35do all the things that we can at least
2:16:37in some predictable way
2:16:39>> and we can have like chips for the for
2:16:41the memory part, you know, we could
2:16:43always supplement that. So
2:16:45>> increase the capacity
2:16:46>> and and and hopefully um AI will help us
2:16:49figure out how to deliver these
2:16:50therapies to the brain.
2:16:52Yeah, delivery is always the biggest
2:16:54problem,
2:16:54>> right? Well, this has been such a
2:16:56fascinating and exciting conversation.
2:16:58Uh, Duria, I have a couple of more
2:17:00questions, closing questions for you.
2:17:03And I really kind of was just wanting to
2:17:04know
2:17:06if you had access, let's say there was
2:17:08no guard rail rails and you had access
2:17:11to
2:17:13all this data in aging biology, you
2:17:15know, the T- cell, you know, all the T-
2:17:18cell repertoire, long, you know,
2:17:20longitudinal uh longitudinal
2:17:22longitudinal cohorts, um, centinarian
2:17:26data, like everything, just anything you
2:17:28can imagine. You had it all and you had
2:17:31this model that was amazing that you
2:17:33could
2:17:33>> You're describing heaven for me.
2:17:35>> Yes. Yes. What what what would be the
2:17:37the the prompt? What would be the
2:17:39question you would you would ask it? I
2:17:41mean, there'd be more than one, but what
2:17:42would be the first?
2:17:43>> Yeah. Hoping that the AI won't answer uh
2:17:4642 as an answer. Um the so so so uh the
2:17:50the first thing I would probably ask is
2:17:53um not not saying that just go figure
2:17:57out aging or whatever because I think
2:17:59there has to be there there has to be
2:18:01certain sequence. So imagine that you
2:18:04have all this data. Uh what would be the
2:18:09the most practical
2:18:11um uh quickest way you can develop uh an
2:18:15intervention to an elderly person say
2:18:18age 70 80 years old that will uh
2:18:22immediately add five years to their
2:18:24lifespan. So to me that would be uh the
2:18:28most critical immediate question to ask
2:18:31because do that population doesn't have
2:18:34a lot of time and so we have to develop
2:18:37these these technologies extremely
2:18:40quickly and will should have you know
2:18:42even two years three years extend so
2:18:45that I can come up with the next prompt
2:18:47uh after that. Um so I guess that that
2:18:51that would be the the the first prompt I
2:18:54would ask. That's great. What um Okay,
2:18:57there's another question. So, this this
2:18:58one is
2:19:00there's no there's no money. Money is no
2:19:02object. Okay. There's no like you have
2:19:05complete like access.
2:19:07>> You're describing so many heavens now.
2:19:09>> I know. I'm just I'm curious what what
2:19:11your your answer is. You're going to
2:19:14personally build your own digital twin,
2:19:17>> which which I plan to
2:19:18>> right now. Um
2:19:21what test would you prioritize? like
2:19:23what data sets would you prioritize, how
2:19:26can a person get them, um how often
2:19:29would you take these tests, how would
2:19:31you organize this information into the
2:19:32AI to really get the biggest bang, you
2:19:36know, benefit from the information it's
2:19:38going to give you,
2:19:39>> right? But you said money is not
2:19:40>> money is not an issue, right? Money's
2:19:42not an issue.
2:19:43>> Um so, so I would divide it into two
2:19:45parts. Uh one part is that we have to um
2:19:50um so what I would do is set up a a huge
2:19:53lab um you know partially automated lab
2:19:58where I would generate enormous amount
2:20:01of data on the um on the cells on the
2:20:04tissues in the in the lab because uh we
2:20:08have to go by the first principles to
2:20:10understand what's going on on let's say
2:20:13in an individual T- cell uh all these
2:20:16thousands of proteins, metabolites, what
2:20:18are they doing? Then then that will
2:20:20enable me to create what's called the
2:20:23virtual cells um and then eventually
2:20:25virtual tissues and you know how cells
2:20:27are in a special temporal manner are are
2:20:30are behaving and so on. So that would be
2:20:33that probably be the most expensive part
2:20:35of it and uh I'll need a lot of money.
2:20:37You said no limit, right? So okay. Um uh
2:20:41but the second part would be sort of
2:20:43what we talked earlier uh kind of the
2:20:46behavioral data from from the human
2:20:49humans and that data is not just of
2:20:52course you know all kinds of you know
2:20:55plasma levels of proteins metabolize
2:20:58your full microbiome your full genome
2:21:01sequencing and all of these things are
2:21:03are possible by the way I mean it's you
2:21:05know if the cost is not an issue you can
2:21:07easily like UK bio Bio bank has done it
2:21:09for 500,000 people. You can do it for a
2:21:12million people. Uh and I think if you
2:21:14did it in a million people that would
2:21:16pretty much cover all the possible human
2:21:18humanity. I mean I it's not like
2:21:20everybody's perfectly uh different. You
2:21:23know we share a lot of things and and
2:21:25and and so you know from the humans
2:21:29collect lots of biological data but very
2:21:32importantly behavioral data. I think
2:21:35this this is something that's totally
2:21:37missing in a digital twin. Um like you
2:21:40know we were talking earlier uh ability
2:21:42of someone to walk certain distance,
2:21:46ability to to you know raise some some
2:21:49weights. These don't show up in any
2:21:52biomarker sets but they could be
2:21:54extremely important. um uh or ability to
2:21:58think, you know, their their cognitive
2:22:01level that that could be directly uh
2:22:04brain uh brain aging related and and I
2:22:07mean you lots of things and and you know
2:22:10what happens when humans are in certain
2:22:12environments, you know, in certain
2:22:14environments you even if you if you are
2:22:18having a very sort of healthy lifestyle
2:22:21that may not help you much. for example,
2:22:24you know, I lived in New York City for a
2:22:26decade. Uh, you know, my stress level
2:22:28was so high uh uh uh and that stress
2:22:32level is so harmful for you because the
2:22:35immune system is constantly thinking
2:22:37there's a threat out there and it's
2:22:38causing a lot of inflammation. In fact,
2:22:40I think people who live in New York has
2:22:41twice as much heart attack risk or
2:22:43something like that. you know that that
2:22:46your environment, your um uh your
2:22:49emotional states and how you interact
2:22:51with other people. All of these things
2:22:54will impact your aging process, your
2:22:57your resilience to the life, your
2:22:59optimistic level. By the way, being
2:23:01optimistic is one of the best things you
2:23:03can do for for aging and study after
2:23:06study show that. So being able to absorb
2:23:10um bad things that happen to you and
2:23:13then keep keep going. So resilience. So
2:23:16but these are behavioral data that's not
2:23:18available in in the biological set. So
2:23:21uh yeah uh I would do that for for a
2:23:24million people all over the world
2:23:25different parts. Um and then on the lab
2:23:28uh every single cell type that I can
2:23:30find uh decode those uh put them all
2:23:33together to the super intelligence and
2:23:35then voila we have digital twin.
2:23:38>> Okay Doria. So let's say someone wants
2:23:40to build their little mini digital twin
2:23:43right now using the models we have
2:23:45access to today. The type of data that
2:23:47we can aggregate you know at the
2:23:49consumer level today biometric data that
2:23:52we can that we can put in. um how would
2:23:54you build that mini digital twin today?
2:23:59>> Yeah, great question. I mean uh in fact
2:24:01it is possible to build a sort of a mini
2:24:04digital twin uh that doesn't have to be
2:24:07as sophisticated as I described because
2:24:09that that one is more uh sort of
2:24:11clinical trials and developing
2:24:13treatments. Uh but you know going back
2:24:16to the u example of the UK bio bank you
2:24:19know they didn't have trillions of data
2:24:21sets. they only used a few hundred data
2:24:24points from from each person and they
2:24:26were able to predict a lot of diseases.
2:24:28So that means that you know we we can
2:24:29have a lot of predictive power with the
2:24:32data that we're collecting uh today. Um
2:24:35you know another example is this uh
2:24:37glucose u meter that I have um you know
2:24:41every five minutes it shows my glucose
2:24:43level and then I take that data and of
2:24:45course I put it to chat GPT um and once
2:24:48you uh additional
2:24:51data set that becomes very uh very
2:24:54valuable because uh let's say that you
2:24:56have your lab values your cholesterol
2:24:59your glucose um your uh every day the
2:25:03the steps that you took and your sleep
2:25:06uh and so on. So these are actually very
2:25:08rich data on their own because they're
2:25:11uh their accumulation of lots of under
2:25:14uh uh underbiology that that results in
2:25:17that but also that puts uh AI into a
2:25:21context your mini digital uh twin. So my
2:25:24my suggestion would be uh to uh u you
2:25:27know provide the AI as much data as they
2:25:31can and on a daily basis so that so and
2:25:35keep it in the same context so same
2:25:38window so they so the model can remember
2:25:41that um actually there are there are
2:25:43some tricks uh uh to do that as well.
2:25:45You can keep it as like a database and
2:25:48tell li model go check my database and
2:25:51see what my new uh you know based on my
2:25:53new data how things have changed what
2:25:56suggestion you could give. Um I for
2:25:58example uh provide all the supplements
2:26:01that I take you know um you know the
2:26:04type of food that I eat um all of these
2:26:06things will make uh will make a big
2:26:09difference. Um so the model start to
2:26:12really personalize um you know sort of
2:26:15the uh style. It will know your style
2:26:18and will make uh suggestions for you. Uh
2:26:22rather than giving blanket statement
2:26:24like you should walk 10,000 steps. Well
2:26:26you know knows that like Daria cannot
2:26:29walk 10,000 steps every day but I think
2:26:323,000 would be enough for him.
2:26:34>> And what kind of model are we talking
2:26:35about? Would you be using the GPT5.5
2:26:38Pro? And then what about you know these
2:26:40agents and codecs and how does that come
2:26:42into helping analyze that that database
2:26:44that you're creating?
2:26:46>> Yeah, I I think you know these models
2:26:48are becoming more agentic all the time.
2:26:50I I know OpenAI for example they
2:26:53integrated agents into um their codeex
2:26:56model the coding model and soon I'm sure
2:26:58it will be part of all of chat GPT. Um
2:27:01you don't you don't I don't think you
2:27:03need very sophisticated models for that.
2:27:06What is important is that uh really
2:27:09maintaining that context. Uh so
2:27:11hopefully the models will have a larger
2:27:14memory and they can remember. So you
2:27:17chat can keep certain memories about you
2:27:20but it's still kind of limited. Uh um
2:27:23it's not just CHP like you can use JNI
2:27:25for example which has a longer um uh
2:27:28context windows um or or cloud for that
2:27:30matter. I think most of the models can
2:27:32handle that uh information. And they
2:27:34don't have problem dealing with large
2:27:37data sets. As I mentioned, I can put
2:27:39millions of data sets and they're able
2:27:41to analyze that. What they need is that
2:27:44they need to remember how things were a
2:27:47month ago
2:27:49because that's before and after. Before
2:27:51and after is extremely valuable. So the
2:27:55model will know
2:27:57he started taking vitamin D3. Oh, these
2:28:01things changed after that that you may
2:28:03not notice or is glucose looks better
2:28:06because of you know when that's that
2:28:09change happened. So it's starts to make
2:28:12those lengths and and that's I think the
2:28:14critical point because you need all of
2:28:17that context in the in the AI model to
2:28:19to give you sort of a better uh
2:28:22prediction on what to use and what not
2:28:24to use. Okay, you were using that well
2:28:27maybe that was not a great idea so
2:28:28change it. um or change the dolls or or
2:28:31whatnot.
2:28:32>> Yeah, that's interesting. It kind of
2:28:34reminded me of a question that I did
2:28:36want to ask you about, you know, these
2:28:38AI models and future AI advances
2:28:42when you think about these qualities.
2:28:43So, like persistent memory, expanded
2:28:45context handling, it seems like those
2:28:49seem to be more important.
2:28:51>> Absolutely. that I I think um for me uh
2:28:55memory which which brings the context uh
2:28:58so the models are now able to think for
2:29:02quite long time and they don't because
2:29:04previously the models would just um even
2:29:07in the same context window if you had a
2:29:09million context windows uh after a while
2:29:12they would just fall off because they
2:29:14would forget even what they were
2:29:16thinking about. Now they have this
2:29:18ability to constantly um go and check on
2:29:21it. So uh I think in the next few months
2:29:26this is going to happen. Uh so that that
2:29:28will that will have a tremendous impact.
2:29:30Well that's memory is everything.
2:29:32>> So so you think so how long are we
2:29:34talking like let's say you know you
2:29:37started a vitamin D supplement 6 months
2:29:40ago. Put that you have the same window
2:29:42and you start in that window you have
2:29:44that you know entry point that the date
2:29:46and then you keep adding about you know
2:29:48you add your your your data in. has got
2:29:50all the data um right now. Can it go
2:29:54back that far or how far can it go back?
2:29:57And
2:29:58>> um if you have that data somewhere in
2:30:00your database uh for example um I
2:30:03adapted a a technique that uh Karpathi
2:30:07who's a famous AI researcher described.
2:30:10Uh so you can um turn you can create
2:30:13your own Viki sort of Wikipedia kind of
2:30:16a thing like personal um you take uh you
2:30:19know uh if you have all your data
2:30:21somewhere uh you can ask AI just pull
2:30:24all that and put it into a Wikipedia
2:30:26like you know you can do it daily or
2:30:28weekly depending on the environment
2:30:30whatever um and so now you're building
2:30:33your own database health database uh
2:30:36which AI can help you update it if you
2:30:39have that data
2:30:40it can go years doesn't matter like you
2:30:43can have 10 years of data it will
2:30:45analyze all of that um uh but
2:30:48>> it has that memory it can like
2:30:50>> yeah so in in in in the same context if
2:30:53you provide all of that I mean it's
2:30:56still limited with you know maybe a
2:30:58million tokens or something but no one's
2:31:00going to have million token data set
2:31:02even if you if you calculate 10 years so
2:31:05so that's that's not that's not a
2:31:07problem the problem is like if if you
2:31:09have
2:31:10If you want this to be continuous like
2:31:12you just give AI okay here's the data
2:31:16today that it should be able to remember
2:31:18what was yesterday what was 2 months ago
2:31:21so you don't have to give you know all
2:31:23of the um uh you don't have to keep your
2:31:25own database and and give all that again
2:31:28and again because you have to do that
2:31:30every time right so the your whole and
2:31:32that will spend a lot of tokens and
2:31:34stuff like that so uh but but I think
2:31:36this is this is going to be uh this is
2:31:38going to be sold
2:31:39How do you not bias? How do you lower
2:31:43the the ability of yourself to bias what
2:31:47you know GPT 5.5 Pro is is going to feed
2:31:51you back, right? Like based on what
2:31:53you're asking it and I mean I I find
2:31:55sometimes I I might be able to bias it a
2:31:58little bit. Do you do you do you know
2:32:00what I'm talking about?
2:32:01>> Yeah, sure. I mean that's why I think uh
2:32:04we are in sort of the experimental phase
2:32:07um in a way um everyone has to do their
2:32:10own kind of validation
2:32:13um as the models are getting better.
2:32:14What I mean by that is that again you
2:32:17know of course don't try harmful things
2:32:18and then you know uh don't go into risk
2:32:21but you know for daily daily use um you
2:32:25might be taking vitamin D
2:32:28and then you you stop taking vitamin D
2:32:31so that you're just doing an experiment
2:32:32like before and after and then you
2:32:34collect that data before and after u and
2:32:37then AI gives you one solution says well
2:32:41you know taking this dose of vitamin D I
2:32:43think is important So then you can start
2:32:45that dose again and then see see what
2:32:48happens. If if you reach the same level
2:32:51as before means that AI made a good
2:32:53prediction like you need to see after
2:32:56you have to have that record before and
2:32:59after so that you you are the judge.
2:33:02Well what this was a good idea so I'm
2:33:04I'm glad that I listened to Jupy. Well
2:33:07if it wasn't a good idea it didn't kill
2:33:09you. It didn't make you sick. So that's
2:33:11that's also fine.
2:33:13>> Yeah. I guess for someone that's already
2:33:14taking a lot of supplements for example,
2:33:17they're not going to have that before
2:33:18and after. Then also you have to know
2:33:19like how long do you wait you know for
2:33:22example to for the wash out period and
2:33:25>> yeah and whatnot. the the hope is that
2:33:28if you provide that very frequently um
2:33:32in fact um I can mention one thing uh
2:33:35for example the the lab values like you
2:33:37go and measure your cholesterol glucose
2:33:39sodium whatever they always give you a
2:33:42range right so if it's within this range
2:33:45it's normal well how do you know that uh
2:33:49because you can be at the top of the
2:33:51range uh that might be your abnormal
2:33:54somebody else's normal somebody might be
2:33:57a little bit over the normal and might
2:33:59still be okay um or vice versa because
2:34:03we don't know the level on a
2:34:04personalized level. So we we calculate
2:34:07population base. So okay so this range
2:34:09is good for this population. So in a way
2:34:12um if you have three or four
2:34:15measurements let's say every few months
2:34:18you can develop your own set point
2:34:21normal. you know the the AI will know
2:34:24your normal for glucose is 90 not 70 not
2:34:30100 or not 105 somebody else might be
2:34:34102. So it knows that based on that that
2:34:37measurements. So then then it starts to
2:34:40give you advice based on your data set
2:34:43your set points because if yours is 100
2:34:46and suddenly dropped to 70 maybe that's
2:34:49not a good thing you know I'm just I'm
2:34:51just uh giving an example uh uh so uh
2:34:54that's why that continuous data
2:34:57collection is so important uh uh with
2:34:59with glucose meter I collected every 5
2:35:02minutes uh the more data the better
2:35:05>> well Duria thank you so much for sitting
2:35:07down with me today and talking about
2:35:09this exciting I mean frontier that we're
2:35:13exploring
2:35:14you know curing disease extending human
2:35:17life expectancy obviously health span
2:35:20reversing aging perhaps getting to human
2:35:232.0 you know where we're enhancing
2:35:25you know genetic you know features as
2:35:29well um very exciting time to be in and
2:35:33if we cannot die in the next 10 to 15
2:35:35years
2:35:36>> it may be even more exciting. Yes,
2:35:38absolutely. Because uh you know the last
2:35:40thing I will say uh
2:35:43this is so unique in human history. Uh
2:35:46because a decade ago uh if you set
2:35:50someone well you should be very healthy
2:35:53you know uh do this do that and they can
2:35:56say well it's only going to extend my
2:35:58life maybe two years or three years. I
2:36:00just want to live my life and I don't
2:36:02care about living few more years as an
2:36:04old age. and that that was perfectly,
2:36:07you know, relevant. That's not the case
2:36:10now. Living an extra one year could make
2:36:15you reach that threshold where there's
2:36:17going to be the ability to to treat many
2:36:21diseases and reverse your aging and give
2:36:23you another decade, give you another 20
2:36:25years and then once you reach that, you
2:36:29get another 10 years, another so like
2:36:32even every day counts now in my opinion.
2:36:35Uh so uh that's why don't die for
2:36:39>> um where people can find out more about
2:36:42your research and they can follow you. I
2:36:44follow you on X. Um maybe you can tell
2:36:46people how to follow you, what your user
2:36:49your Twitter follower or sorry your ex
2:36:52user handle is and where else they can
2:36:54find you.
2:36:55>> Uh yeah, my my main account is an X. Uh,
2:36:58it's at Daria D E R Y A T R under dash.
2:37:05Um, if they write Dario Nutmas, I think
2:37:07I'll I'll show up. Um, that's that's
2:37:09where I, you know, do most of my
2:37:11communication. Um, I have a LinkedIn
2:37:13account, but I don't post that often
2:37:15there. Um, I I've been planning to start
2:37:19up a a sort of a YouTube channel, but
2:37:21um, I don't think I'll ever do that
2:37:23because I'll never have the time. you
2:37:25know, it's it's really amazing what what
2:37:27you're doing because
2:37:30video takes a lot of a lot of effort. Uh
2:37:32so for me is the the fastest way. Uh in
2:37:35fact, I I even had a Substack uh
2:37:37account, but just couldn't find the time
2:37:39to write long uh uh long messages. So So
2:37:43uh X is the best way.
2:37:44>> Well, I really encourage people to
2:37:46follow you on X. post. I mean, just
2:37:48every day there's something interesting
2:37:50that you're posting on X and so I highly
2:37:53recommend that people do follow you
2:37:55>> as um many already do. So, thanks again
2:37:58for the research you're doing and for
2:38:00I'm I'm excited to see what um what's
2:38:03going to happen in the next couple of
2:38:04months.
2:38:06>> Looking forward to it. Very optimistic.
2:38:08Thank you. Thank you very much. It was
2:38:10great.

About this transcript

This page contains the full transcript of Why the Next 10 Years May Add 50 to Your Lifespan | Dr. Derya Unutmaz by FoundMyFitness, generated from the public captions YouTube serves with the video. The transcript has 26,832 words across 3,708 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.