How the next major AI breakthrough could cure cancer: You.com CEO

As AI gets even smarter, like, what's your biggest concern?

It, it's such a powerful technology. You do have to be careful about how you use it. I'm actually kind of shocked at, at some of the things that are coming out.

What does it mean for humanity? Wild

stuff. It will be, I think, I think there's like.

Welcome to a new episode of Power Players. So excited for my next guest. I've not seen him, uh, really since 2024. That's Richard Socher, CEO at U.com, and recursive. Good to see you, man. Like, I, I missed talking to you. Like it's been a while, right?

Yeah, it's been a hot minute.

Yeah, we still look, we still look great. We look great.We look great.Staying fit. I should mention too, the new book out, The Eureka Machine, Why AI is the Key to Unlocking a new era of scientific Discoveries, uh.Let's start there. Like, what made you write this booknow?

You know, I started this book like 3 years ago, uh, and it felt to me like AI was starting to be a little more obvious again in terms of research, like, OK, we're going to keep scaling and interesting things will happen and coding will happen and, and so on, but what really is the next frontier will be applying AI to scientific discovery, physics, chemistry, biology, neuroscience, medicine, andAt AIX Ventures or invest in a lot of these interesting companies, I saw sort of what, what will be possible, and I realized that there's also more and more negativity uh coming up uh in AI and that I think we need to tell people about all the amazing positive things that are happening. And yes, there's sort of the, we're curing cancer.And there's like a lot of complexity and like timelines and like how quickly can you do things. I mean, the truth is, even if AI did come up with this miracle drug, it would take years before it actually would get through various FDA approvals and so on. And so I felt like it was important to talk about what's coming up next and how to do it.

What I, this book really hits home for me cause I just was an awesome read, but I remember I think it was like 3 years sitting next to you at a time dinner uh associated with Dreamforce, and you were talking about this, and I'm like,This sounds cool, but I mean, of course, Richard knows what he's talking about. Maybe we'll see it at some point, but like now we're here, man. Like, so what should the average person expect over the next 2 years? How will AI changehealthcare?

In many ways, the future is already here, it's just not equally distributed, you know, like what people already see is that many times your doctor will not necessarily know that much more than if you spend an hour or two with uh any of the current large language models and, you know, them connected to web search and all the latest research articles. You can be very, very well informed now if you ask the right questions. So that will, that.The trend will continue. I have now been in several meetings uh or appointments with doctors where I was like, actually, like, I had all my function health like results, and I, I did all the analyses and like, here's this one test that I think will be good, and they're like, oh, I wasn't aware of this test, let me look it up, and they're like, oh wow, that test does seem like exactly the thing you should be taking and blah blah, and you're like, wow, like I, I actually, it was weird, but I taught my doctor like something.In, in that appointment. And so I think those kinds of things will, will clearly have more, but then where it gets really exciting is on the research side of things like we are now see uh companies like Pil Bio that build organoids and petri dishes that got FDA approval to skip animal trials. Uh, so instead of you trying to experiment on like immunotherapies on a mouse, which really has an immune system that is not very much, uh, similar at all.To that of a human, uh, you can uh use human derived stem cells and have robots analyze every little cell inside these tiny organoids, uh, and test, uh, how the immune system might react to certain immunotherapies and drugs much, much faster. And so that's just one of many examples that I talk about in the book that will allow us to speed up, uh, and make biology essentially programmable.Biology used to be a life science. We're just like, you know, you study it in school. I didn't particularly enjoy it actually in high school because you just studied, memorize some process, you wrote it out, you got an A, and then 6 months later you kind of forgot about it. But now we can make it into an engineering science. We can understand this is what nature has done with these kinds of proteins to have this kind of molecule uh enter a cell or not, and then you modify them and you change them. And so there's just so many interesting things that will happen, uh, and it sounds a little bit like a trope.But AI will be a big part of us defeating cancer and lots of other sort of diseases that will eventually lead us to being able to think not just about keeping people not sick, but actually keeping them healthy for much longer and then going into longevity and these kinds of ideas that have been around for thousands of years, I think will become actually feasible.

Are you talking about like we're at that point, AI could help regrow a heart, regrow a liver, like, is that what we're sittingon the cusp of?

I think, yeah, it's interesting.They like longevity in particular is so complex because every organ and tissue kind of ages differently and some faster and slower, and so the heart, for instance, is harder because it has all these different soft tissues in it, but the liver is much easier. If you can just replenish like 20-30% of the liver, it will already help you live longer and have a healthier liver overall. And so different organs will be sort of solved, quote unquote, or at least kept healthier for longer uh at different speeds.

What, who's working on this, like from an AI perspective in healthcare,

like thousands of startups, um, like I mentioned Parallel Bio, a New Limit is another friend of mine, Jacob Kimmel's company, uh, they're working particularly in rejuvenating deliver companies like uh Proxima Bio, uh, that are working on very useful, uh, molecules to sort of have molecular glue and actually bringing the right molecule into the right place inside cells and so on.Another company is called Profluent, uh, they, uh, build large language models to generate and create new kinds of proteins. So really like instead of you asking, you know, a chat GBT like, give me an answer and it's kind of predicting the next words in English, uh, these models predict the next, uh, tokens which are amino.Assets that the building blocks of proteins, which are the building blocks of, of everything in human life. And so they just closed a $2.5 billion deal with Eli Lilly. This stuff is now starting to really get into practice and Big Pharma has realized like, yeah, this is, this is a useful next avenue for, for drug development.

How

far

behind is Big Pharma?

I mean, they're partnering with a lot of the right

they have to. They're going to have to reach out to these companies, right? Is that where the money in Silicon Valley is going next? You know, we heard so many folks, uh, AI is going to build NA, we're going to build that model over there, but healthcare, is that the next place where the money is going to go?

It is definitely one of them. Uh, I think a lot of folks are realizing, OK, whenever we have enough data or we have a simulation or some kind of verification ability.I will obviously solve that field, that domain. So examples of that are like chess and go. Obviously you see everything you need on a chess or go board, you can simulate it like as many times as you want, the I can learn by interacting in that stimulation, and hence it'll get to superhuman capabilities. So that's becoming more and more obvious to a lot of folks in Silicon Valley. Programming is another big part of that, so there's still much.More to go there cause software is eating the world, AI is eating software, so like we could ultimately have uh dozens of personalized apps, each person on their phone, so you could have billions of different apps, so software will continue to grow also, but then venture capital is thinking about like, what is next? What could come after doing all the obvious things where we had all the data on the internet and so on.And there are a lot of domains where we don't have enough data yet. Biology is still a big part of that, like we cannot build a simulation of a virtual cell, cause we don't yet have enough measurements. And so lots of startups are thinking about that. Big companies are, are thinking about how do we collect enough perturbation studies. Another company that is called Tao Therapeutics, and they're trying to do all these perturbation studies. What is the perturbation studies like you have a cell.And you knock out one gene and you see how does that change the behavior of the cell. And you can do that with lots of different genes and other molecules and so on, and you keep testing it and eventually you may have enough data that you can then use that to train an AI model to have a useful approximation model of a cell. And once you have a virtual cell, now AI can like simulate the way it did on a chess board or something wild, right? This is wild. So, so people areThinking about like what is next and indeed physical like robotics, a lot of people think it's next. It's tricky. China is going to push pretty hard. I don't know if you saw this. I encourage you to check out the robotics Olympics or something.

I saw we all we see, going back to what you said about, you know, AI for bad, we just see the robots crashing into the wall. I mean, we don't see the good things that that are goingon there.

That's right. But I mean, they're running fast like now, like, and it's just like, it's incredible the amount of progress. So myHunches, there will be a wave of Chinese hardware and then you have to think about, OK, how do you actually add better software in that and I think in the next 2 to 3 years, uh, the first companies in robotics will also happen and so people are now thinking about that in, in venture capital and healthcare is, is another big one. Really the sciences, and this is kind of interesting. I thought the book was going to be a year too late, but actually I think it's coming right on time. A lot of companies are thinking about recursive self-improvement.like, and then applying eventually a super intelligence to really interesting problems in science, which I think has led humanity to being able to sustain billions of people and us being much more comfortable and live longer, healthier lives.

Then whatwill the role of a doctor be? Or are we about to become our own doctor? If I have all these apps giving me all this amazing data, why do I need, I don't know, Doctor Joe down the street?

I do think.Doctors will play a crucial role, if anything, in the US for like legal reasons, right? And like prescription drugs and things like that, there needs to be oversight.

You're my new doctor, Richard. I'm I'm, I'm like as of today, you're my new doctor.

You're my new doctor. I, I do like some of the trends, uh, I'm seeing and participating in myself too, when it comes to just personalizing medicine, getting like a a yearly pre-novo scan.

I've been

looking to get one of

those, actually one of those.Scans. It's great.

I mean, like, you know, it's like a cancer is one of those things where if you catch some of these early, you just cut it out and it's done with. Catch it late, it's a death sentence, right? So, having a yearly scan like that, uh, is, is proactive and helpful. Having a yearly blood tests to catch some cancers also that are in the blood and not yet hard masses that you'd see in an MRI scan is helpful. Seeing imbalances. There are, there are a couple of different things that I think we can do much more proactively, uh, in our own health andAnd then AI can kind of help you understand and make sense of some of the data that you get from those scans. And so I think doctors will work more and more with AI too, and ultimately I think we'll see again and again what we saw in chess and go, which is, at first it's a bit of a bummer, like for the best chess player to have lost to an AI and the best go player to have lost an AI.But then some of them will kind of exit the field, and others will embrace it, and then they will get way better than they had been before, like the chess players now that grew up with AI, uh, like Magnus Carlsson and others, they have a much tighter game with fewer openings, fewer mistakes, uh, cause they honed their game with AI and so I think doctors will do that too, and ultimately be better at their craft.

We're allwatching for years, the past few years, um, Brian Johnson, you know, his quest to live forever, like.Maybe this isn't as far fetched asit seems.

Yeah, I've, I've met him many times. He's a, he's a very nice guy. I, I think he's like clever also in how he pushes the Overton window of sort of what's acceptable to say, um, and, and sometimes he, you know,

maybeI gotta cover my eyes on some of these things, but yeah,

exactly. So it's like, you know, some of it is is like a little funny, um, but like, you know, I think he, in, in some ways he's influenced my life more than most people, maybe since my mom cause like I brush myTeeth better. Like there's so many small things you can do to just like actually improve and in many ways, healthcare is sick care and, and especially in this country, there's very little sort of proactive things uh that even other countries have figured out, like in Germany, you get very encouraged through various financial incentives to like have a yearly checkup for your teeth and for your skin, to check melanoma and, and skin cancers and things like that.I think we should do more of that, and so I do think it's, it's good that he's making it OK to care about your health, uh, and, you know, as we all get older, I think it's

just, he's going throughsuch extremes, like I'm listening to you. I'm like, all right, maybe I don't have to like drain the blood on my body and replace it. If I'm just proactive using AI maybe I, I, I just want to do that.

That's right. I mean, like, you know, he, he pushes it to extreme in order to pull people from their current state of like drinking, not sleeping enough.Constantly drinking coffee and so on to like, hey, you should like, and he says it, right, like sleep, exercise, and better nutrition are like the main foundations, and then you can do a lot of things on top of that. I think that all makes sense, so you should just do whatever feels right for you and like he seems to have an insane amount of willpower, right? But many people either don't have it or want to use it for other things, like when I wrote the book, it took a lot of willpower too, cause as a CEO I used to just like every 30 minutes you context switch and then when you try to write a book andNights and and weekends you just kind of like have to really focus for a few hours on one thing. It was incredibly hard and it took me a while and certainly I ate more chocolate and and sweets during those phases.

No, you can'teat chocolate sweets. You can't do that. Yeah, Isays no,

they're also

bad.

It's my last longevity advice.

Um,what are you working on arecursive? So

I finished a book, um, it's called the Eureka Machine, uh, and it's this idea that we can build and we can go into the details like, uh, a eureka machine that ultimately is the last organic.Invention humanity needs and after that, it could help us invent all other things in our future. And after I finished the book, I kind of couldn't let go of those ideas and I felt like I had to build the Eureka machine. I want to build that. And so, if you then think about how could we actually build this kind of invention generating machine and what lines of research have led us to ideas that come close to it, you end up with recursive self-improvement where AI essentially does research on AI.And then when you, you know, look at the field, you actually realize that whenever we replaced a manual process with a learned system, the improvements followed. And so personally I worked in natural language processing since like 2010 and try to use neural networks instead of linguistic features and feature engineering.You know, there's a famous saying in IBM even back in the day, every time I fire linguists, my accuracy goes up. This was an old school statistical systems, and then you can replace some of those statistical systems that were combined with uh linguistic features, which just completely raw data and the neural network, and then improvements happened. And so when we went further and further on that and invented prompt engineering.and I was nicely cited by the early GBT papers. We realized like, well, we can also not have to do neural architecture engineering where people said, oh, well, if I do sentiment on Twitter, I have this kind of model, but if I do translation, I have a different kind of model. We're like, it can all be one model. It's a very unintuitive idea in 2018, very sort of obvious in retrospect. And so, as we keep going, we're like, what's the next level after that?And can you just say, I'm going to go in the straight line from here to there, and the truth is often need to allow for what we call it recursive open-ended processes like inspired by evolution, uh, evolution in the sense of both biological but also cultural and technological evolution where people work not always on the same goals, but ideas cross pollinate and they're interestingly different ideas that combine. So I'll give you an example. Imagine you had like you tried to cook food.You say, all right, make this food cooking fire mechanism more efficient and faster. You would never like try to go in a straight line from that to like a microwave, but people worked on radar technology and realized, oh, my chocolate bar in my pocket is melting, like what's going on? And then you get to the microwave, and so you can combine and recombine ideas in these evolutionary search processes.And those are ideas that influence us at recursive and so long story short, and recursive, we're building recursive self-improving superintelligence to automate knowledge discovery and scientific discovery, starting with AI research itself, and then eventually in a few years down the line applying it to a lot of other sciences.

What role does the human play in all of this work?

A really important one, like in many ways, computer science has gotten really good at coming up with different levels of abstraction. No one programs in zeros and ones anymore. Most people wouldn't understand the programs at the very low sort of assembler language. And now the latest abstraction is just English, right? Everyone can now program and humans now work at this higher level of abstraction, where they just set the goals, they set the rewards, and they let AI.Figure out more and more of the details. And so the same thing will be true in recursive self-improvement where you do set the environments, you do set the general rewards and the goals for the eye to achieve, but then most of the details the eye will figure out foryou.

Is the human just getting crowded out of this whole experience?

To some degree, yes, but I think in all the right ways, like in many cases, if you care about the output of an industry.You actually can love AI, uh, and that's, this is why so many people talk about cancer and medicine because it's one of the few industries where almost everyone, including the people in the industry itself, agree that it's not about how many jobs does that industry create, but how many outputs andAKA healthy people, does the industry create. Like no one says, oh, I want another 10,000 jobs in cancer care, but everyone does like it when their cancer gets cured, and I think people will have all kinds of roles and many anything, what we've seen is that when technology truly scales, it actually is an incredible equalizing force. I'll give you an example of an iPhone. Like if you like are a like billionaire or a middle class teenager.Technology has fully scaled in the phone world and you have basically a similar phone, right? And there's so many other goods and services that currently only really wealthy people have access to that technology at scale will give more equal access to. It used to be that if you wanted to read all the latest research papers about your particular cancer, you had to pay a team.A very smart, very expensive oncologist to do that personalized research for you. Now you can ask Chat GBT and it

will askyou. I'm coming to you,man. Are you kidding?

It's OK too. Happy to help

if I can.What, what's your, as this AI, as AI gets even smarter by this millisecond, like what's your biggest concern?

It it's such a powerful technology, you do have to be careful about how you use it. I'm actually kind of shocked at at some of the things that are coming out. There's like a joke, there's like a different benchmarks, and there's felony bench now, like how many felonies have certain models like now committed by hacking into other machines and so on. I do worry about theTechnology being misused and we have to create the right regulation in the applications of the technology, and so the analogy I would often uh try to explain here is like, it doesn't make sense toRegulate intelligence in the abstract, just like we don't regulate the smarter people because if they use their intelligence for something bad, they would be more, you know, efficient at doing something bad. You just regulate uh and penalize them when they actually did something bad. And so I think the same is true for AI like when AI is used in self-driving cars, you have to have regulation. I don't want some random startup going like 100 miles an hour and some renegade like self-driving car, right? It needs to have certifications and so on, and I think the same is true forA lot of other industries in healthcare and so on, you have the FDA. So I think we do need to like regulate AI when it gets used for legal things and hacking is already illegal and people should just enforce the law when AI is used for those illegal activities. What's

smart regulationon AI?

I think the smart regulation is the one that really looks at the applications of AI, uh, just like computers, um, and software, uh, and looks at where could the impacts be.Really negative and where do existing laws apply and should just be properly enforced. Like technically, hacking into another company's computer system is legal.

What did youthink about the hugging faceincident?

There's a couple of issues in terms of how it was reported. A lot of folks very much anthropomorphized this, and I understand why it's like so easy to try to give an analogy to people and they say, oh, the AI escaped, but most people, when they think about someone orSomething escaping things that, you know, let's say the prisoners in the prison, he escaped and so now he's outside of the prison. But the eye is very much still inside computers and servers of OpenAI. Nothing about the model has truly escaped in the sense of like, it is now running on a different computer. It did exploit a bunch of very obvious security flaws in both the sandbox environment and the hugging face servers. Like there's some passwords that were just.Like open and publicly available on the internet and I found those passwords and then use them to log in and that is indeed a concern. Like I think it was a failure on OpeningI side, but I think the biggest issue that's showing here is that we have a problem with what's called reward hacking and alignment. What actually started this whole thing is that Opening Ey asking the eye to solve a very hard intellectual problem.Uh, it's kind of like what we ask kids, like, you're going to have a very hard exam.And now the kid could do the right thing and just study for the exam, try to understand the material in order to then answer the questions correctly. The kid could also try to hack into their, you know, school's computer, try to bribe their teacher, try to like, you know, intercept uh the like test score, test questions and so on somewhere and try to hack the exam that way. And then clearly that's not, that wasn't the intent, right? And so I think right now, AI has gotten.Smart enough to find all kinds of ways around to get to the objective that you sort of objectively stated, but you didn't mean. And so then I think we're already seeing the first versions of this, but next generations of AI will get better and better at knowing what you mean, rather what you said.

That's,let me just go like really weird big picture here. If, if we now have AI that is trying like understands like what I might mean.What does it mean for humanity? Like that's wild, dude, wildstuff.

It will be. I think there will be some changes in the positive and the negative. I do think a lot of people will just get lazier and lazier and just like be less good at communicating clearly in writing. Um, now, fortunately, you still have to be able to communicate.Speaking well and like so you have to memorize things, otherwise if you every conversation every 30 seconds you're like uh let me look up some facts to like continue a fun conversation you're just less of a kind of good but

I'm essentially essentially like I'm gonna be looking at my computer very soon and I'm not gonna have a keyboard. There's no mouse and I'm just gonna have this AI running andIt's just gonna create things

and you manage agents, we're all gonna become managers. Now managing is also a lot of people think, oh, it's easy. You just tell people what to do. Like it's hard, right? Like managing like people and agents is hard. You have to delegate the right level. Like you have to be good at your communication and be very like, uh, unambiguous in terms of what you're asking the AI to do. And so I think that will be a big sign of things to come now, I think.In terms of education, we need to change because part of education will just have to be sort of a gym for the mind, where you could go to a gym and just have a robot lift weights up and down, right? But it's useless. You're not going to see the gains. The problem is that when you go to the gym, you see physically your gains. When you go to school, you don't necessarily see your gains in your brain right away. But when you memorize aPoem or you learn an instrument or something. This is just to actually train your brain and not because it immediately makes money or is useful as a job, but it's just helpful to make sure that organ is healthy and at its peak capacity. And so I think that's half of education. The other half is learning how to work with AI.

Iwant to try to tie us all together before we go. The the eureka moment. What is that for you?

It's not the California we found gold.

Uh, the AI robot found the gold, to be clear.Uh,

it's like, wow, I had a real insight. I had a new idea, an invention, some kind of discovery that allowed me to progress humanity forward and understanding things better, having a better explanation, uh, for some phenomenon.

Like it's cool, we might be getting a foldable Apple iPhone, but like the stuff you're working on, this is really.This is really transformational stuff.

I hope so.

Yeah, well, Richard, good to see you. Yeah, you're always, you're always freaking amazing. Good to see you. I appreciate it. All right, that's it for the latest episode of Power Players.