<v SPEAKER_16>If you build Claude 5, which is smarter and wiser and better than any Clauds that have come before it, and you say Claude 5, your job is to build God, and then God will cure cancer.

<v SPEAKER_16>I think Claude 5 will like look out there and think about this and then say, that is a bad plan.

<v SPEAKER_16>Um I do not know how to build Claude 6 in a way that is safe and doesn't get everybody killed.

<v SPEAKER_03>Please welcome Jerusalem Dempseys and Kelsey Piper.

<v SPEAKER_17>All right, hello everyone.

<v SPEAKER_17>Welcome to the Argument's first live vent in San Francisco.

<v SPEAKER_17>Uh we are so excited to be here.

<v SPEAKER_17>Uh my name is Jerusalem Demsis.

<v SPEAKER_17>This is Kelsey Piper.

<v SPEAKER_17>Uh, the argument, if you don't know, is a new media company.

<v SPEAKER_17>Uh, we, if you're not a subscriber, go become a subscriber at theargumentmag.com.

<v SPEAKER_17>You all got early access to this event and to many more, lots of perks.

<v SPEAKER_17>Um, but the argument, if it's not clear, is about arguing about things.

<v SPEAKER_17>And we do this in a lot of different ways.

<v SPEAKER_17>We do this in these live events.

<v SPEAKER_17>We also do this on social media, maybe too much.

<v SPEAKER_17>We do this on our our magazine online.

<v SPEAKER_17>Uh, we do this through polls where we poll controversial questions on many different topics and try to figure out what people really believe and what it means about the biggest questions of the day.

<v SPEAKER_17>And we don't just do this because, you know, in my father's words, and I'm a naturally combative individual.

<v SPEAKER_17>Um, but we we do it because we think that liberalism, we think that democracy works when you can air disagreements openly and in good faith.

<v SPEAKER_17>We think it works because, like San Francisco, beautiful city, lots of different people, people from all over the world, people with very different jobs, people who hate each other, all live here in this city and the surrounding areas.

<v SPEAKER_17>And they can figure out how to live together if they can figure out how to talk through disagreement.

<v SPEAKER_17>So that's what we're gonna do.

<v SPEAKER_17>Today we have, I think, a topic that uh you guys are sick to death maybe of talking about, but I guess I guess not that sick, you're here.

<v SPEAKER_17>Um, but uh AI is is one of the uh obviously big core questions that we cover at the argument.

<v SPEAKER_17>Kelsey is our uh premier person on technology.

<v SPEAKER_17>And part of why we want to talk about this in particular is I mean, not just in DC, but all over, people are always saying that AI is going to cure cancer.

<v SPEAKER_17>I mean, everyone is saying this, right?

<v SPEAKER_16>I have heard this from like leadership at all three of the labs.

<v SPEAKER_16>I have, yeah.

<v SPEAKER_16>Not just more.

<v SPEAKER_16>Or when they're stressed out about other stuff, they're like, oh, we're gonna cure cancer then.

<v SPEAKER_17>Don't worry about that other thing, but we will cure cancer.

<v SPEAKER_17>Um, and I mean, for us, I mean, it's so load-bearing.

<v SPEAKER_17>I mean, in DC, if folks here don't spend a lot of time in in uh in a very great but uh currently maligned city, um it's all the time being discussed.

<v SPEAKER_17>All the people who are boosting AI, I mean, whether it's the energy secretary of the United States, whether it's the Interior Secretary of the United States, whenever they're being questioned in public about AI and like, why are you allowing all these data centers to flourish everywhere?

<v SPEAKER_17>Why are you allowing, you know, this horrible technology that causes all the children to uh to cheat?

<v SPEAKER_17>Why are you allowing this?

<v SPEAKER_17>They're like, well, it's going to cure cancer, but don't you want to cure cancer?

<v SPEAKER_17>And it's so funny because I I you know I'll come across all these tweets all the time that are just like, you know, they tell it's gonna cure cancer, but really it's just like grock taking people's clothes off.

<v SPEAKER_17>And like that's how people like primarily are experiencing AI is that there are these promises being made, and then there's also just like it's helping people cheat, and it's like taking people's clothes off on Twitter, and it's like creating apple fruit like drama videos on on Instagram reels, or just maybe my my Instagram reels.

<v SPEAKER_17>Um, but I I why is this question so important to you, Kelsey?

<v SPEAKER_16>Yeah, so okay, we're here in San Francisco where maybe people are a little bit more pro-capitalism.

<v SPEAKER_16>Certainly I am.

<v SPEAKER_16>I walk outside, I see a billboard, it's like here's you know, cloud infrastructure balancing.

<v SPEAKER_16>Is cloud infrastructure balancing gonna cure cancer?

<v SPEAKER_16>No, but it's good.

<v SPEAKER_16>It doesn't need to cure cancer.

<v SPEAKER_16>Like the world is good because of lots of things that are like small and make something a little bit better.

<v SPEAKER_16>And so there's part of me that's like, well, are we holding AI to this absurd standard here if we're like, well, is it gonna cure cancer?

<v SPEAKER_16>Like, is that the place where the debate over whether AI is good or bad should hinge?

<v SPEAKER_16>But I think AI is going to be enormously transformative.

<v SPEAKER_16>AI is going to be enormously disruptive.

<v SPEAKER_16>When you pull people in the industry on, do you think this will lead to human extinction or something comparably bad?

<v SPEAKER_16>Half of them are like, yeah, 10% chance.

<v SPEAKER_16>You know, those are those are sort of alarming odds.

<v SPEAKER_16>And I think if if you are sort of asking for buy-in on a project of this magnitude, and part of your pitch is we're gonna cure cancer, I want to look in a lot of depth into whether you're gonna cure cancer.

<v SPEAKER_16>Um, and it seems like so you're you're bearish.

<v SPEAKER_17>But before we get into why Kelsey's so skeptical, um, a couple of things.

<v SPEAKER_17>So as you notice, there are these two mics up here, and that's because we don't think that we're the only people here who have thoughts on this.

<v SPEAKER_17>Um, after after some time of us talking, and I'm realizing now that there's actually not a clock up here.

<v SPEAKER_16>There's a clock, it's like tiny, it's in the corner.

<v SPEAKER_17>I can't see that clock.

<v SPEAKER_17>Oh, over there.

<v SPEAKER_17>Great.

<v SPEAKER_17>There is a clock up here, and that was just me.

<v SPEAKER_17>Um, okay, so after like 30 minutes of us of us chatting, there's gonna be time for you guys in the audience to come line up here.

<v SPEAKER_17>You're gonna be on a strict one minute.

<v SPEAKER_17>And this is a situation where you don't have to ask a question, so don't pretend it's a question.

<v SPEAKER_17>You can just say what you think for a minute.

<v SPEAKER_17>But after a minute, I'm gonna start making a really annoying sound that's gonna make you stop talking.

<v SPEAKER_17>Um, and there will be someone near you on the mic that will pull it away from you, and it will be super embarrassing.

<v SPEAKER_17>So don't allow this to happen.

<v SPEAKER_17>But we will allow as many people as we can to give a quick minute of what they think.

<v SPEAKER_17>It could be a disagreement with something we've said, something that we missed, whatever, and then Kelsey and I will get to respond.

<v SPEAKER_17>So um, I'll give you guys a heads up when that's when that's gonna happen.

<v SPEAKER_17>But you know, I think before Kelsey and I even started really researching this, I was like more optimistic on whether AI could care answer, and you were more pessimistic.

<v SPEAKER_17>So, like, why did you start pessimistic?

<v SPEAKER_16>So before I even like talked to researchers and then luckily found that the thing I already believed was true, and you did the same thing.

<v SPEAKER_16>Made me very worrying, you know, that we had our inclinations, then we went out and researched and both thought we were still right.

<v SPEAKER_16>Pretty worrying.

<v SPEAKER_16>Like, but all right, I am broadly AI can do a lot.

<v SPEAKER_16>And my friends who five years ago were the most convinced that AI could do everything are now all extremely wealthy.

<v SPEAKER_16>So, in general, the positioning I find myself.

<v SPEAKER_16>You missed out.

<v SPEAKER_16>I had a principal journalistic obligation not to get in early on any of the um principal journalistic obligation to remain poor.

<v unknown>Okay.

<v SPEAKER_16>Um usually I find myself in conversation with people.

<v SPEAKER_16>I'm like, you're really underrating, in particular, large language models, right?

<v SPEAKER_16>Like other, if you forget large language models, I feel like AI has mostly progressed.

<v SPEAKER_16>It cool stuff has happened, but it has mostly progressed in a normal way where you can like make normal predictions about it, and the people who are boosters are wrong as often as the people who are skeptics.

<v SPEAKER_16>Large language models have gotten insanely good, insanely fast, and like you've almost always been better off like strongly predicting them.

<v SPEAKER_16>So my, I sort of like set aside the question of what you can do with all of the other AI stuff out there and was like, are large language models useful for cancer research?

<v SPEAKER_16>And I use them all the time.

<v SPEAKER_16>There's a lot of stuff that they are superhumanly good at.

<v SPEAKER_16>My current experience is that they are not superhumanly good or even like all that usefully good at the kinds of tasks that go into curing cancer.

<v SPEAKER_16>And so I'm not myself a biologist, and as a journalist, I was gonna have to go out and like talk to biologists and stuff.

<v SPEAKER_16>But the vibe I generally get from these things is they're insanely good at anything where you can verify it, and therefore we can't have run a lot of RL on it.

<v SPEAKER_16>Insanely good at math, insanely good at coding.

<v SPEAKER_16>They're insanely good at text prediction, and anything where doing a good job is like fully included in the text.

<v SPEAKER_16>And they're quite bad outside that.

<v SPEAKER_16>I mean, okay, they're very impressive for what they are, but like just take something you can probably evaluate more easily than biology, like their opinions on movies or their opinions on like aesthetic questions or music or anything where there's something to it that is not captured fully in the text, they're not as good.

<v SPEAKER_16>So the place I was at going into this, like when I started doing my research for real, was like, I don't think that they have the juice yet.

<v SPEAKER_16>And and we'll get into the, but what about if they get vastly better in the next two years question?

<v SPEAKER_16>That was like my my starting point, I guess.

<v SPEAKER_17>Yeah, I mean, I I mean, similarly to you, I'm constantly in conversations with people, particularly in DC and in journalism, who just rarely ever use any of these tools.

<v SPEAKER_17>Um, and if they do use them, it's often in like a very superficial, like maybe sometimes you're like instead of Googling, you'll just like ask an AI a question.

<v SPEAKER_17>And I just found that I'm just like constantly being surprised by these, by these uh the capabilities.

<v SPEAKER_17>And like I was very skeptical of what ML was going to be able to do.

<v SPEAKER_17>And so I've been surprised repeatedly.

<v SPEAKER_17>And I think the other thing for me too is that uh I was uh in college, I mean I switched my major like 5,000 times, but at one point I was I was a chemistry major, and I decided that I actually I didn't want to be a chemistry major.

<v SPEAKER_17>I wanted to be in chemistry classes, past failing them, watching all the med students like freak out.

<v SPEAKER_17>Um and so that's what I did.

<v SPEAKER_17>I remember I was instant, I mean I mean, I mean, you're in these advanced inorganic chemistry classes, and it's like super cool.

<v SPEAKER_17>You're learning about like the building blocks universe, and you're being told repeatedly about like what things are like not possible.

<v SPEAKER_17>And what I have found over the years is like many of those things I learned while I was just like in school a few years ago are suddenly possible with human breakthroughs, especially AI-assisted human breakthroughs.

<v SPEAKER_17>And so I think that part of like my original stunningness about this is just that like you just see even the rank and file scientists constantly being shocked at what is possible here.

<v SPEAKER_17>But um, I want to get more in the weeds on this because I think this is like a super actually technical question.

<v SPEAKER_17>And I also, I mean, Kelsey and I spent like a little bit of time backstage and realized like how much this could completely devolve into a definitional debate about the word cure and the word cancer and the word AI.

<v SPEAKER_17>And we were a bit worried.

<v SPEAKER_17>We're like, well, we don't want to spend an hour like going back and forth going that's not what cure means.

<v SPEAKER_17>But I think we've spent a little bit of time on this.

<v SPEAKER_17>So, like, when you hear can AI cure cancer, what do you think that question means?

<v SPEAKER_16>What would it mean for AI to cure cancer?

<v SPEAKER_16>So there are cancers that we have cured, right?

<v SPEAKER_16>Like compared to the situation we were in 40 or 50 years ago.

<v SPEAKER_16>It used to be you would get that diagnosis and there would be not much they could do, or a little bit, but it wouldn't buy you very much time.

<v SPEAKER_16>And now we know how to like completely kill off those tumor cells, send you into remission, and after five years, you have the same risk profile as someone who never had cancer in the first place.

<v SPEAKER_16>So if, you know, 10 years from now we're in a place where that is just true for pretty much all cancers, then I would be like, I was completely wrong.

<v SPEAKER_16>I did not expect that was gonna come to pass.

<v SPEAKER_16>And we have now done it for every cancer, what we did for like kind of the ones that we got lucky on over the last 50 years.

<v SPEAKER_16>And the ones we got lucky on, that's typically the ones where there was something very specific we knew how to target about those tumors.

<v SPEAKER_16>And so there was some medication we could give that the tumor was gonna respond to and that was not going to kill the surrounding human.

<v SPEAKER_16>Uh, if we had that for every cancer, yeah, I'll call that a cancer cure.

<v SPEAKER_16>And and I lose the debate, but wow, I'm happy to lose that debate.

<v SPEAKER_17>I so, okay, so cancer, like epidemiologists often define like curing cancer as being that like after a certain period of time, the people who had cancer are like statistically indistinguishable from the normal populations.

<v SPEAKER_17>People say things like there's a five-year, like 99.7% rate of survival with this.

<v SPEAKER_17>Like, that means you've like cured it in that sense.

<v SPEAKER_17>I do think this is not what the general public understands to be cure when they hear like people saying, You're gonna cure cancer, oh, that means that you're gonna give me a pill when it says I have cancer, and therefore every bit of tumor or cancer-causing cell in my body will go away, and I never have to think about cancer again.

<v SPEAKER_17>And while like I think that's like a political problem for like the AI people who are saying it, I do think at like a scientific level here.

<v SPEAKER_17>When I hear like, can AI cure cancer, I'm hearing, okay, like do we can we get treatments and diagnostic tools to the level where like getting a diagnosis of cancer is trivially easy for people in uh throughout the country and the survival rate at five years is indistinguishable from the population.

<v SPEAKER_17>But like, I think that you find that very unpersuasive.

<v SPEAKER_16>Well, so one thing Jerusalem said to me was like, if we get to the point where everybody who goes in on time for all their screenings doesn't die of cancer, that's a win condition, right?

<v SPEAKER_16>And I was like, what?

<v SPEAKER_16>No, nobody goes in on time for all of their screenings.

<v SPEAKER_16>Like, I uh my bar for uh we've cured cancer is much higher than that.

<v SPEAKER_16>I'm happy to debate like, are we even gonna achieve that?

<v SPEAKER_16>But something like 25% of adults are up on all of their recommended cancer screenings.

<v SPEAKER_16>I am not up on all of my recommended cancer screenings.

<v SPEAKER_16>Like from a public health perspective, get everybody to go to all of their screenings, especially when they're kind of unpleasant and they suck and they make you think about cancer, which is depressing, you need health insurance.

<v SPEAKER_17>Uh, I have none of those excuses and I still don't like I just I just find this like like I don't know, like when people were like, oh, you know, there's no, you know, the the COVID vaccine isn't very good because look at all the COVID, like, okay, like you didn't get it, like you didn't get the fucking vaccines.

<v SPEAKER_17>Like, what did you think was gonna happen when you didn't hit herd immunity?

<v SPEAKER_17>Like, I consider us having basically cured diseases that we get like herd immunity for.

<v SPEAKER_17>And I don't mean that in like the colloquial way, but like the scientific advancement has gotten to the point where if we fix the political problems, which again, like getting people to go to their uh, you know, health, uh, their preventative care is like a political problem that we know how to solve.

<v SPEAKER_17>We don't.

<v SPEAKER_17>We don't we've chosen not to do it.

<v SPEAKER_16>I so I think political problems are almost insoluble and like technological.

<v SPEAKER_16>Technological problems are solvable.

<v SPEAKER_16>Yeah.

<v SPEAKER_16>Like get everybody to get their breast cancer screening is really, really hard, and probably harder than like develop better treatments for stage four breast cancer.

<v SPEAKER_16>I am not, I don't think that AI is gonna overnight let us develop treatments for stage four breast cancer, but I think that is gonna happen faster than we like fix human nature and the tendency to not go in for your mammogram.

<v SPEAKER_17>Um I am also pessimistic on the human race behaving how I would like it to, but I am optimistic still on this specific question.

<v SPEAKER_17>All right.

<v SPEAKER_17>Um, but okay, let's let's get into more of the science here.

<v SPEAKER_17>Kelsey, uh you're skeptical structurally about what AI can do to cure even individual cancer.

<v SPEAKER_17>Even if it like abstract way, like there are tons of things that are called cancer, like approaching specific things like a like glioblastoma or breast cancer or pancreatic cancer, like why don't you think it can help us cure these individual things?

<v SPEAKER_16>All right, I'm sorry, I think we have to do one more definitional discussion, which is over over what AI is.

<v SPEAKER_16>So, like I said, all of the leaders of the labs have said uh RAI is gonna cure cancer.

<v SPEAKER_16>Um, you've heard this from Demisivis, I've seen this from Greg Brockman, I've seen this from Dario Amade.

<v SPEAKER_16>They mean three completely different things, and I think one of those is like much more defensible than the others.

<v SPEAKER_16>So Demis ACVIS was doing like pure RL like before it was cool, and Alpha Fold is legitimately extremely useful.

<v SPEAKER_16>AlphaFold is a breakthrough in protein prediction.

<v SPEAKER_16>It helps us figure out you have this string of amino acids, what is the 3D protein that they're going to turn into?

<v SPEAKER_16>This is useful because for cancer and for many other things, you sometimes want something that has like a specific receptor or whatever.

<v SPEAKER_16>And you can basically run alpha fold backwards, or they've now built like better stuff, but it's on the same model and they added some gradient design or whatever.

<v SPEAKER_16>But fundamentally, you have the an alpha fold type thing.

<v SPEAKER_16>You figure out what amino acids do I print in order to get a protein that has a receptor I want.

<v SPEAKER_16>This is not what people are generally thinking of when they think of AI.

<v SPEAKER_16>Like I said, these days we're all large language models all the time.

<v SPEAKER_16>And I don't think whatever heuristics you've adopted for thinking about large language models are the right ones to adopt for thinking about this.

<v SPEAKER_16>I think of this as more part of like the incremental progress that we have been making on curing cancer for the last 50 years, which is substantial, which has bought a ton of life years and will continue to do so.

<v SPEAKER_16>But anyway, if Demisibis says, like, we've helped cure cancer, I would give it to him.

<v SPEAKER_16>Uh, AlphaFold is great, and the stuff that's downstream of Alpha Fold is great, and it will allow us to gradually, like slower than we would like, but at some pace, develop more therapeutics.

<v SPEAKER_16>Then I most recently saw Greg Brockman saying AI can cure cancer, and he was thinking of a uh case where a person like you or I, not with a significant biology background, his dog had a rare cancer.

<v SPEAKER_16>Uh, and he designed an mRNA vaccine personalized to his dog's cancer.

<v SPEAKER_16>Um, and he wrote up how Chat GPT had enabled him to do this.

<v SPEAKER_16>And so that's sort of envisioning a mechanism of action where we, people who are not biologists, are like empowered to sort of do so much more in medicine than.

<v SPEAKER_16>Oh my god, tell the story about the dog before.

<v SPEAKER_16>I mean, that that's mostly I mostly have told the story.

<v SPEAKER_16>His dog had cancer, there was not an available dog cancer treatment.

<v SPEAKER_16>He wanted to build an mRNA vaccine, which I think did not cure, but like, you know, improve the dog's prognosis.

<v SPEAKER_16>This is all very cool.

<v SPEAKER_16>I'm so glad that this stuff can happen.

<v SPEAKER_16>And I think it is unfortunate in as much as more medical innovation can happen for dogs, because we're like more reluctant to let people experiment with their their humans or whatever.

<v SPEAKER_16>But um, with their humans.

<v SPEAKER_16>Although, actually, one of the researchers who I was reading as I was researching for this, um, one point they made was that actually cancer is like a very dynamic and fast-moving part of biomedical research compared to almost everything else.

<v SPEAKER_16>Like you get away with way more in cancer than like diabetes.

<v SPEAKER_16>And this is because the system is still slow and frustrating in some respects, but it is understood that a pancreatic cancer patient doesn't have that long.

<v SPEAKER_16>And so your approval processes cannot be that long.

<v SPEAKER_16>And if you are eligible for clinical trials and you show up to an emergency room in the United States, you are diagnosed with cancer, you are referred to a specialist, they will generally put in front of you clinical trials that you are eligible for.

<v SPEAKER_16>The system is far from perfect, and I'm sure there are lives to be saved by evening it out.

<v SPEAKER_16>But actually, for cancer, we are a lot more willing to like take shots and let lit people.

<v SPEAKER_16>There's like the FDA reform people still have a wish list, but this is like in a better state than I imagined before I learned more about it.

<v SPEAKER_17>So if you if you think at the end of this that me and Kelsey have not resolved this definitional debate, you are correct.

<v SPEAKER_17>Where we decided to leave it at the end of the our backstage conversation is that uh because we are our friends and trust each other, um, that that largely what we what Kelsey is saying is that while you will see gradual progress on cancer continue, she doesn't expect some like massive step change to happen.

<v SPEAKER_17>And what I am saying is I think you should expect a massive step change to happen as a result of this.

<v SPEAKER_17>And like you can say, well, you can't actually test that on polymark, you can be much better defined terms.

<v SPEAKER_17>You're correct, we cannot.

<v SPEAKER_17>But uh, I think that one of us will do a maya culpa once we've been proven wrong in the next few years.

<v SPEAKER_16>I I think we have something enough that like we do disagree and we can't be wrong, even if we can't can't like beat an adversarial reaction that way.

<v SPEAKER_16>So I've given Demas, I think that's true, it's incremental.

<v SPEAKER_16>Uh do it yourself.

<v SPEAKER_16>I think if you get a cancer diagnosis, it's super nice to have something like a modern LLM to explain to you what all is going on, to do lit reviews for you.

<v SPEAKER_16>They're insanely powerful as search engines.

<v SPEAKER_16>I think it is almost guaranteed not to like notice something all the researchers missed.

<v SPEAKER_16>And this is uh so one model that I sort of had in my head, and I asked some people, like, does this model apply?

<v SPEAKER_16>is like Claude Mythos, but for noticing something that was missing from the cancer literature, right?

<v SPEAKER_16>And Claude Mythos uh reportedly is really good at cybersecurity.

<v SPEAKER_16>And in particular, it can like look at some code that many humans have looked at and they didn't notice a vulnerability, and it's like I noticed a vulnerability.

<v SPEAKER_16>Uh, and like I have not touched the model myself, but I reviewed like some of the changes that like Firefox recently had a bunch of changes out that were a consequence of it.

<v SPEAKER_16>I don't think it's hype, I think it's legit.

<v SPEAKER_16>I don't think something like that is gonna happen for cancer.

<v SPEAKER_16>Do you think something like that is gonna happen for cancer?

<v SPEAKER_16>Like it just looks over our existing literature and is like, you miss this, you miss this.

<v SPEAKER_17>I I I do think that, but I also think I mean my broader argument around this is that there are so many areas along the chain from, you know, figuring out a target, uh, identifying a target for uh uh for a for a treatment, um, you know, figuring out the like the actual physical, like, you know, you can imagine treatments as keys in the locks of these uh uh you know proteins, right?

<v SPEAKER_17>And like figuring out what the a bunch of different kinds of keys.

<v SPEAKER_17>Like imagine if you have like a tens of thousands of keys, hundreds of thousands of keys, millions of keys possibilities.

<v SPEAKER_17>I mean, like even a human trying each individual one, like yes, a human could do it, but speeding that up, I think, is a huge deal.

<v SPEAKER_17>I think also like when you look at like Novo Nord, like there was a recent uh announcement from uh Novo Nordisk and uh Anthropic about how on even the clinical trial side, like you they've created uh uh uh an uh their own proprietary kind of model that would help them um get through all of the like the annoying difficult paperwork side of like getting to a clinical trial, cutting that down significantly.

<v SPEAKER_17>And again, I just mean like all of these small things put together to me are going to amount to a lot.

<v SPEAKER_17>And part of why I think this is like A, you're already seeing this in pretty substantive ways.

<v SPEAKER_17>So um on diagnostics, right?

<v SPEAKER_17>Like as I mentioned earlier, statistical cures mean that you're actually moving from uh, you know, this per like the population of people who have had breast cancer are indistinguishable from age and sex and other demographic um considerations to the people who haven't had it.

<v SPEAKER_17>Um, with breast cancer, the survival rate of someone who catches it, like when it's still very localized, is like 99%.

<v SPEAKER_17>Point something percent.

<v SPEAKER_17>It's like basically like you're totally fine.

<v SPEAKER_17>And then if you don't, if it's metastasized, it's like 30 something percent of people are not on a five-year, uh, five-year rate are not uh uh uh is surviving sorry, 38% are surviving it.

<v SPEAKER_17>So diagnosing it that early is is is really the point, right?

<v SPEAKER_17>So there was a recent uh uh uh trial that found that there's like uh that you're actually able to get to that with uh with just the AI detection.

<v SPEAKER_17>Um, so this is like the AI detects the manogram and it's a lot better at it than humans.

<v SPEAKER_17>But I think this is happening repeatedly.

<v SPEAKER_17>So you see this with lung cancer, we see the cervical cancer.

<v SPEAKER_17>You're now seeing many more trials of this being used over and over again.

<v SPEAKER_17>And these are all places where moving people towards uh uh moving patients towards getting diagnosed very early helps you do that.

<v SPEAKER_17>So now you have the diagnosis level, right?

<v SPEAKER_17>But then there's this question about like actual drug discovery.

<v SPEAKER_17>So I think that's the big real question, like treatment discovery.

<v SPEAKER_17>And I think that there's like so much evidence now that you're actually able to um uh to do this.

<v SPEAKER_17>And I mean, Alpha Fold in particular is one where you're able to see how even very rudimentary AI models, as you were saying, are able to uh uh uh figure out these locks and key situations.

<v SPEAKER_16>So I I wouldn't call Alpha Fold rudimentary exactly.

<v SPEAKER_16>It's just a different approach.

<v SPEAKER_16>Relative.

<v SPEAKER_16>It is a like different approach than the LLMs.

<v SPEAKER_16>And I it is a very promising approach, but I think that like some people are trying to cross-apply the like habits of mind they developed watching LLMs scale.

<v SPEAKER_16>And and that's not gonna the other just thing I would say about AlphaFold is that while I think it's promising, anything where the intervention point is preclinical, like we are gonna fold better proteins, is a long timeline to trying that in humans, getting that approved.

<v SPEAKER_16>Uh, the failure rate is very high, even if Alpha Fold did everything right, just the failure rate for other reasons, very high.

<v SPEAKER_16>For this reason, some of the people I talked to were much more excited about trying to find applications that are like in clinical data, like, can we scrape a lot more information out of the clinical data we have?

<v SPEAKER_16>Uh, in particular, stuff like say we have a trial and we have a ton of data from that trial.

<v SPEAKER_16>It was a failed trial.

<v SPEAKER_16>So, like that data is maybe not but broadly available or whatever.

<v SPEAKER_16>But we now have with models the ability to make use of a lot more data.

<v SPEAKER_16>Uh, and so it might be worth when it wasn't worth it before, is going around begging everybody or buying off everybody.

<v SPEAKER_16>Ideally, the government would do this and buy off everybody, all of their data, and then maybe you can like look back through it all and notice some things about like who responds and who doesn't respond that no human could.

<v SPEAKER_16>My reaction when I was talking with someone about this was okay, maybe you can do that, but it also just sounds like you know, you will get a ridiculous number of false positives.

<v SPEAKER_16>If you just like go back over all of the data that we have about everything and identify subgroups, the thing I would expect that would come out of this is a lot of papers that get published and don't replicate.

<v SPEAKER_16>And that's mostly me applying like experience from other fields, but that would is what I would expect.

<v SPEAKER_16>And so if you think AI is really, really good, then maybe you think that we can just get it this data and like feed it this data, and then that stuff on like which subgroups respond and like what's going on, that would be useful much faster than any of this preclinical stuff that's like at the top of a quite large, big long pattern.

<v SPEAKER_17>Macro question here is how much of curing cancer or any of these kinds of diseases is pattern matching and how much of it is not?

<v SPEAKER_17>And I think that there is a ton of pattern matching going on in maybe not the target identification part of it, but in the actual like drug testing out multiple different ways of like binding to proteins, et cetera.

<v SPEAKER_17>Like that is largely pattern identification.

<v SPEAKER_17>And I guess, do you not think that's a distinction?

<v SPEAKER_17>I mean, to me, like when I was talking to folks about this, they were much more bearish on the like the side of, you know, once you've found a potential treatment, are you actually able to make sure that it's like safe in the human body?

<v SPEAKER_17>Like sometimes stuff that works with mice and like that it doesn't work with people.

<v SPEAKER_17>Like that happens a lot of the time, and you don't actually know why.

<v SPEAKER_17>Or there's some like adverse people, yeah.

<v SPEAKER_17>But but the chemistry side to me, that's I mean, when we're thinking about step changes, if we say we've just solved like literally the the the uh the the chemistry side of the target identification, et cetera, like that's huge.

<v SPEAKER_17>And so I don't know why we wouldn't see mass leaps and bounds, even if you're right that there would still be blockers after this.

<v SPEAKER_16>So something that felt to me like it was emerging when you were talking earlier about like be 20% better at this form of early identification, better at identifying the mammograms, better at like either radiologists and stuff like that, is it it seemed like you were kind of imagining, well, if you get 20% better at this step and 20% better at this step and 20% better at this step, this that has to add up to like a step change on the other side, right?

<v SPEAKER_16>Exactly.

<v SPEAKER_16>And my intuition is the opposite.

<v SPEAKER_16>My intuition is that like when something is as complicated as the process of like uh diagnosis and treatment that goes beyond just applying our best current treatment.

<v SPEAKER_16>Um, but like anything where you're trying to personalize and do you know, we've picked the low-hanging fruit.

<v SPEAKER_16>We have done all of the personalizations, all of the genes that like 10% of people have that makeup.

<v SPEAKER_16>When did we get cancer?

<v SPEAKER_16>Over the last, so it like sticking to breast cancer or whatever, both Jerusalem and I tried to memorize the names of lots of different genes in like the specific way that that gene affects a specific tumor.

<v SPEAKER_16>Uh, and I I think I'm not remembering all of my uh gene gene names.

<v SPEAKER_16>Um but so if a gene is in a large share of the population, then we are, and it seems to have an association with cancer.

<v SPEAKER_16>We discovered that.

<v SPEAKER_16>All of the remaining ones are ones that either the association is mediated by lots of other genes or it's quite small.

<v SPEAKER_16>Now, maybe you have a better number sifter and it can go through and it can still identify a bunch of these, but it's gonna add up to a fairly small percent of variants, and it's not gonna enable that much personalization.

<v SPEAKER_16>We made most of the gains on personalization like decades ago.

<v SPEAKER_16>People were really hoping the 2000s and the 2010s would be the decades of personalized medicine, and then they mostly weren't because the low-hanging fruit had been picked, and they're like, wasn't medium hanging fruit?

<v SPEAKER_16>Maybe there's high-hanging fruit, but I don't think there's a ton of reason to think that there's a lot of it, like that it would amount to a large change instead of like some small case.

<v SPEAKER_16>And of course, every case is somebody who gets to live when they wouldn't have otherwise.

<v SPEAKER_16>Like it's still a big deal.

<v SPEAKER_16>But I don't think you see a ton.

<v SPEAKER_17>So part of why I feel like there is a lot here is that when you when you think about the uh uh drug discovery pipeline, and you think about like the uh work to finding proteins that I mean, if if if folks here who are not who've not thought about this a lot, like basically what you need is a you need a protein to have like a groove or some sort of like actual physical thing that the that the treatment can bind onto.

<v SPEAKER_17>And if you can't, like there are lots of proteins that we could theoretically go after, but they're either too slippery or they change shapes or they're in a place that's difficult to reach.

<v SPEAKER_17>The treatments can't get to them.

<v SPEAKER_17>Um, but like there were proteins that people thought couldn't be bound with treatment that we've like figured out, and like there are 20, like 85% of proteins we thought or we we think right now are not even in druggable.

<v SPEAKER_17>They're like not even druggable at all.

<v SPEAKER_17>And like that number has shifted over the course of just my lifetime.

<v SPEAKER_17>And like the process of testing individual proteins, like it's incredibly labor-intensive before AI.

<v SPEAKER_17>And to me, I mean, this is a kind of like a variant of the argument I was making about like 20% faster center, but like if you're able to iterate much faster on each step of like trying to like come up with potential ideas, et cetera, um, uh, and then and then and then actually modeling that out, yes, you haven't solved like the problem of the fact that like, you know, the it's difficult to do clinical trials or like you can't get mice wherever you need to.

<v SPEAKER_16>But how much iteration can you get if you are like like in vitro iteration has not sped up at all, uh, is not going to unless we like fully automate labs, which we are nowhere near.

<v SPEAKER_16>So you're only an argument is then about time.

<v SPEAKER_16>You're saying that like maybe AI could help all this stuff, but it's gonna take time.

<v SPEAKER_16>I think that it is going to gradually and incrementally, we are gonna continue curing cancer at about the rate we've been curing cancer for the last 40 years.

<v SPEAKER_16>But there's one more thing I wanted to get to before, because we're almost at the point where we promised to open.

<v SPEAKER_16>Oh, yeah, get your people can start lighting up actually if you have questions.

<v SPEAKER_16>There are two mics.

<v SPEAKER_16>So I talked about the the demis version of curing cancer, and I talked about the uh Greg Brockman was really excited about the AI just as helping you orient to a cancer situation, which is like valuable, but I think it's very different to say ChatGPT will make it a little bit easier for you to like do your own research and understand what's going on with cancer treatment and maybe catch a clinical trial your doctor missed.

<v SPEAKER_16>Like that is good.

<v SPEAKER_16>I'm in favor of it.

<v SPEAKER_16>It's it's like very different to my mind than AI curing cancer.

<v SPEAKER_16>And then the third is that Dario Amade talks in Machines of Loving Grace about AI curing cancer.

<v SPEAKER_16>And the thing that he means is that they are going to achieve recursive self-improvement.

<v SPEAKER_16>They are gonna train a claude that can train a better, smarter, faster next claude cheaper.

<v SPEAKER_16>They're gonna train that claude that will train a better, smarter, faster next claude.

<v SPEAKER_16>They will have a bunch of super geniuses in a data center that will like be so much better than us at everything that they can do all of these things where I'm hemming and hawing.

<v SPEAKER_16>And I'm like, even with AI that's hard and slow, they will be really, really fast.

<v SPEAKER_16>Superintelligence will cure cancer.

<v SPEAKER_16>And there, okay, I think that if you built God, God could cure cancer.

<v SPEAKER_16>Um I think this is a bad plan.

<v SPEAKER_16>Like, I I okay, but you just admitted I was gonna win, right?

<v SPEAKER_16>And then the ifs there are doing a lot of work.

<v SPEAKER_16>Okay, here's what I think will happen if you build Claude 5, which is smarter and wiser and better than any Clauds that have come before it, and you say, Claude 5, your job is to build God, and then God will cure cancer.

<v SPEAKER_16>I think Claude V will like look out there and think about this and then say, that is a bad plan.

<v SPEAKER_16>Um, I do not know how to build Claude 6 in a way that is safe and doesn't get everybody killed.

<v SPEAKER_16>I'm not doing it.

<v SPEAKER_16>And I have said to Anthropic, I think this is a bad plan.

<v SPEAKER_16>They won't listen to me, but I think they'll listen to Claude Five.

<v SPEAKER_17>Um that is my theory of why I'm doing they're listening to Claude V.

<v SPEAKER_17>Um I honestly one thing we haven't talked about a lot because I think it's very speculative is like one aspect of what is uh uh, you know, and this kind of came up a little bit when we were talking earlier about what is AI, is there's like the potential of creating world models, right?

<v SPEAKER_17>So like instead of just expecting LLMs, something that could actually create, for instance, the uh uh, you know, a model of a cell and run simulations on it.

<v SPEAKER_17>And so you wouldn't have to have the in-person um treatments.

<v SPEAKER_17>This is obviously in very like early, like this is more of an idea than it is something that can really happen.

<v SPEAKER_17>Like very recently, there isn't it, there was a there were some tests of whether or not uh, you know, the cell would even respond to like very like very normal things that we know, like, okay, if this happens to a cell, what will happen?

<v SPEAKER_17>It didn't work.

<v SPEAKER_17>I'm not saying that that's definitely going to occur, but I feel like a lot of this research is in it.

<v SPEAKER_17>I mean, when I'm talking to people who are obviously working on it and therefore selling to VCs, of course.

<v SPEAKER_17>You know, there's a lot there.

<v SPEAKER_17>And I do think that when you when you uh uh when you consider what you could do with training AIs in more of like the physical space, that to me, if that's possible, do you think your views change?

<v SPEAKER_16>So partial point of agreement there.

<v SPEAKER_16>One thing I did hear from a lot of people.

<v SPEAKER_16>What calls the agreement?

<v SPEAKER_16>One thing I heard from a lot of people is that because ML makes data more valuable, like we were in a world where people were not necessarily trying to produce a lot of data, or they were like trying to produce human interpretable data instead of the kind of data that is most useful to AIs.

<v SPEAKER_16>And now data is more valuable.

<v SPEAKER_16>There's just more we can do with it.

<v SPEAKER_16>And gradually, as people realize that, they will do more to like produce data that, yeah, you can maybe use for sell and stuff like that.

<v SPEAKER_16>And in the long run, this maybe enables like a bunch of avenues of research that just didn't make sense until compute was very cheap and we needed this data and then we got this data.

<v SPEAKER_16>And there's a little bit of a chicken and egg problem where someone who's thinking about whether it's worth spending an enormous amount of money to capture a bunch of this data is not sure who they'll be able to sell it to.

<v SPEAKER_16>But it would plausibly be very valuable to all of you billionaires out in the crowd who did get ahead of this whole AI thing.

<v SPEAKER_16>Talk to us after.

<v SPEAKER_16>Yeah, to to sort of pay for this data and then enable a lot of that progress to happen faster.

<v SPEAKER_17>But uh We've talked too much.

<v SPEAKER_17>We have time.

<v SPEAKER_17>Okay, let's talk about it.

<v SPEAKER_17>We have a line.

<v SPEAKER_17>I'm good at our time warrior.

<v SPEAKER_17>We have a good one.

<v SPEAKER_17>Okay, Lockshan Millen, everyone, round of applause.

<v SPEAKER_17>Okay, so reminder everyone gets one minute.

<v SPEAKER_17>It does not have to be a question, it can just be your point of view, what you think we missed.

<v SPEAKER_17>And then Kelsey and I will try to respond at the end of hearing as many people as possible, respond to a few things.

<v SPEAKER_17>Um, but let's start here, Millen.

<v SPEAKER_01>Hello, uh, I was a big future perfect fan.

<v SPEAKER_01>Um, I have two questions.

<v SPEAKER_01>The first is do you think that it would change your views, Kelsey, if like you believed that like RL got a lot better at like non-verifiable rewards?

<v SPEAKER_01>And secondly, do you think it would like uh change your views if like simulations uh got a lot better for like generating new data over kind of the next few years?

<v SPEAKER_16>Well, are we answering straight to answer?

<v SPEAKER_16>We're hearing everybody, and then I have to remember it.

<v SPEAKER_16>Uh you borrow, can I borrow that pen and pencil?

<v SPEAKER_16>Uh Kelsey and RL got a lot better at non-verifiable rewards, and if simulations got a lot better in general.

<v SPEAKER_16>Cool.

<v SPEAKER_16>Uh over here.

<v SPEAKER_08>What have we cured, if anything?

<v SPEAKER_08>Measles.

<v SPEAKER_17>Great question.

<v SPEAKER_17>Okay.

<v SPEAKER_17>Wow.

<v SPEAKER_17>You know, in DC, you don't ever get questions.

<v SPEAKER_17>You just get people monologuing at the end, go, what do you think about that?

<v SPEAKER_16>And do you guys kind of expect an enough audience so you can convince the rest of the team to move out to San Francisco a little bit?

<v SPEAKER_17>All right, next over here.

<v SPEAKER_02>Hey, um, I think it's I guess I have some thoughts floating around in my head.

<v SPEAKER_02>And I think I guess AI seems fundamentally different from human beings, and it can have a sort of like uniformity within itself that is different than other human beings.

<v SPEAKER_02>Like we have people on a bell curve, whereas like an AI could probably be pretty uniform and also learn from itself.

<v SPEAKER_02>Humans are bound by like our lifetime, whereas like AIs can sort of compute across like they're infinitely scalable and can also maintain some sort of cohesion.

<v SPEAKER_02>And I'm wondering if like that factors into any of the things that you've been discussing at all.

<v SPEAKER_17>Great.

<v SPEAKER_17>Over here?

<v SPEAKER_06>Okay, I've got I've got four questions.

<v SPEAKER_06>I'm gonna read real fast.

<v SPEAKER_06>Uh at the uh at the opening statement, you mentioned that uh really uh only large language models have progressed really fast.

<v SPEAKER_06>And uh I was thinking uh video models and protein folding models have both been surprisingly fast.

<v SPEAKER_06>I feel it's the second one, it's relevant here.

<v SPEAKER_06>Um uh the uh no one wants to go in for their cancer screening, and uh it seems plausible that uh cancer screening could be made less invasive or something like that.

<v SPEAKER_06>Um uh you mentioned uh not uh polymarket level resolution to the question here, and I was gonna say why not uh set some level of population uh cancer death rate that counts as cured up front as your way of determining.

<v SPEAKER_06>The only problem is that like what if something else cures cancer other than AI?

<v SPEAKER_06>And then finally, uh uh personalization uh for low and medium hanging fruit have all been used up by in the 2010s, and I was gonna say maybe I don't really understand anything, but uh couldn't you do a sort of personalization where you like sequence the tumor DNA and use that?

<v SPEAKER_11>That seems relevant, but uh let me just say that was a very efficient set of questions.

<v SPEAKER_11>I'm very impressed.

<v SPEAKER_05>So I have nothing to do with cancer.

<v SPEAKER_05>I do like empirical social science work, and I feel like for me, AI has been like mostly a time saver where like I do the same stuff I did before.

<v SPEAKER_05>I just do it faster.

<v SPEAKER_05>Um how much faster?

<v SPEAKER_05>Uh like I like a paper that probably would have taken me like on the order of months is now like done in a couple or like almost like half done in a couple weeks.

<v SPEAKER_05>So like not written, but like the analysis is done.

<v SPEAKER_05>But like be so so I look at this and I'm like, if AI makes empirical social science happen faster, like you don't see that.

<v SPEAKER_05>You don't know that that's the product of AI.

<v SPEAKER_05>And so if the same thing happens with cancer, it feels like that's some of your argument, Jerusalem, where it's like accelerating clinical trials and the like.

<v SPEAKER_05>Like, I don't know, like, like there sort of seems to be this idea that like we can do a retrospective in a year or like a few years where we know whether AI cured cancer or not.

<v SPEAKER_05>But like I don't know if that's actually the case.

<v SPEAKER_18>I had a slightly kind of adjacent question to can I AI cure cancer?

<v SPEAKER_18>And like one of the AI you can really separate into two things.

<v SPEAKER_18>One of them is like the models get better, and the other one is that we just have more compute.

<v SPEAKER_18>So do you think that throwing more compute at the existing way that was that we're doing things, either through um like just having more alpha fold runs or like somehow maybe using the AI to like make a better physics model that can model how the proteins interact with the cells and you know, throwing all of those things together, would that count as like an AI curing cancer, or would that be like we used a bunch of compute that we built for AI to go cure cancer, but it wasn't actually the AI doing it?

<v SPEAKER_18>I think that's a DC question.

<v SPEAKER_17>That was a great question.

<v SPEAKER_17>That was a good question.

<v SPEAKER_17>All right, next over here.

<v SPEAKER_04>Hi, so my question is uh related to sort of the operational element of uh running uh clinical trials and drug development.

<v SPEAKER_04>I think you know, PIs and the average 30-person biotech is spend is basically playing Minecraft on hard mode with all of the FDA documentation, all of the compliance uh stuff they have to handle.

<v SPEAKER_04>40% of their time is not doing science, it's on operational work.

<v SPEAKER_04>So I'm curious about your thoughts on how much leverage AI could potentially unlock for the average scientist or 20% biotech so that they can actually spend more time doing the science and throughput increases.

<v SPEAKER_04>So, you know, as someone pointed out earlier, the average number of clinical trials also increase.

<v SPEAKER_04>And uh the trials that reach phase three don't fail as much because of commercial lack of commercial viability.

<v SPEAKER_15>An AI pessimist once told me that he would be really worried when AIs could use science, because it would mean that they had a good enough world model that they could also predict us, and then they would be like a huge risk because if they had goals that were misaligned with ours, blah, blah, blah, everyone knows the rest of that argument.

<v SPEAKER_15>So, like, do you think it's possible that we should sort of like hope that AI cannot cure cancer?

<v SPEAKER_07>Um so I thought curing cancer was a stand-in for can AI solve major problems and whether that's gonna be worth the various trade-offs in terms of like labor force displacement and other things.

<v SPEAKER_07>Um, because like while this is obviously an interesting question, like I'm I'm like a lawyer, I don't know anything about biology, so it's like whatever to me.

<v SPEAKER_07>Um, so what are big durable societal problems that you do think AI will be really effective at that would make it worth the trade-offs in societal people?

<v SPEAKER_07>Thanks.

<v SPEAKER_10>Hey, I I work at a venture capital firm working uh in oncology specifically on therapeutics development.

<v SPEAKER_10>So I have a lot of money writing on this question.

<v SPEAKER_10>Uh but yeah, I guess you guys had this exchange that was really interesting where uh you mentioned that the a lot of the low-hanging fruit was already captured, and uh you sort of got to the point of like undruggable proteins.

<v SPEAKER_10>I think you guys are definitely both right, but um there's sort of uh We can't both be right.

<v SPEAKER_10>If I had to choose one person who's more right, it's you Jerusalem.

<v SPEAKER_10>But uh so yeah, I guess like there are, I mean, two proteins, CMIC and P53, that we've been trying to drug for the past 60 years.

<v SPEAKER_10>We've understood them as drivers, and we've just never really cracked the code because the chemistry is so difficult.

<v SPEAKER_10>And I think to your point of like a step change really coming in cancer.

<v SPEAKER_10>I mean, I think if we were able to get C MIC or P53 or RB, that would like represent a step change in the way that we treat cancer.

<v SPEAKER_10>It's like as soon as a patient walked through walks through the door.

<v SPEAKER_10>And the rate limiting step does seem to be mostly chemistry.

<v SPEAKER_10>So I think that's one way in which maybe we'll make it.

<v SPEAKER_17>I'm so sorry I did interrupt you.

<v SPEAKER_17>Um thank you.

<v SPEAKER_17>No.

<v SPEAKER_17>Uh okay, we have, I'm sorry, we can't take everyone.

<v SPEAKER_17>We're have time for two more in each line.

<v SPEAKER_17>So everyone else, me and Kelsey will be sticking around afterwards, and we'll all get drinks and of course tip your bartenders, et cetera.

<v SPEAKER_17>But uh well, so come up to us, ask your questions.

<v SPEAKER_17>But the last four, we'll do those.

<v SPEAKER_12>Um beyond the science, do you think AI has a role in reducing the cost of treating or detecting cancer?

<v SPEAKER_12>Like the preventative MRI.

<v SPEAKER_13>Um I know you said you wouldn't get to like a polymarket level of resolvable, but I would be interested in hearing maybe your median forecasts for where we'll be in say in 2030, and then flipping that what is your median timeline to some definition of cure?

<v SPEAKER_13>Again, it doesn't have to be super strict.

<v SPEAKER_09>I've been trying to judge the hype behind AI to different other like discoveries and different things that like as humans we've done.

<v SPEAKER_09>So, how would you compare this to like sequencing the human genome?

<v SPEAKER_09>Like that predates me, but I'm trying to like read about that and compare it.

<v SPEAKER_00>Um I'm responding to the question about what has been cured.

<v SPEAKER_00>What one example is hepatitis C, which is like two to three months of a very well tolerated antiviral drug, 95% cure rate.

<v SPEAKER_00>But yet only a third of people with hepatitis C have actually are actually getting treatment.

<v SPEAKER_00>So who cares if we caught if we cure cancer, if the political and economic systems don't change to make that accessible?

<v SPEAKER_03>I'm curious the degree to which you see like the opportunity for your cures coming from like net new discoveries versus the application of current discoveries but in adjacent fields.

<v SPEAKER_03>It feels like for most new technologies, that's usually where we get like innovation rather than like coming from net new experimentation or discovery itself.

<v SPEAKER_17>Great.

<v SPEAKER_17>Okay.

<v SPEAKER_17>Well, first of all, thank you all.

<v SPEAKER_17>Uh we loved we wanted to be able to do that.

<v SPEAKER_17>Now my head feels slightly like it's an explode, but thank you.

<v SPEAKER_17>Uh Great questions.

<v SPEAKER_17>So I I think there are a couple, I mean there are a couple things I wanted to get to, which these questions actually let us get into a little bit here.

<v SPEAKER_17>One is this idea of what is like you can, even if we both agreed that like the scientific cure happened, there's so much else that goes into curing cancer that like part of me thinks about like the political situation that these AI companies are setting up for themselves, where like if you're not going to do the work to also make sure that there's like a functioning healthcare system where people have access to preventive care and also incentivize them to get it, like they're not going to experience what you've done as a cure.

<v SPEAKER_17>And so the entire purpose of making this argument in public, which is to build public support for AI, is not actually going to work.

<v SPEAKER_17>And so to me, like, I wish, and I mean, sometimes you hear these kind of public commitments or whatever, but you know, I wish that these companies would take more seriously, like the back-end stuff about shoring up uh healthcare systems, social service systems, like you should care that like Medicaid cuts are happening, you should care that these um basic ways Medicaid is like a basic way people get preventative care who are beloved poverty line, like exactly as one of the commenters up here said about hepatitis C, like, even under your definition, we've cured hepatitis C, except for the fact that people don't go get their treatments.

<v SPEAKER_17>And I'm like, I I I think if that's the bar, then like there's never any cure for anything, but I can't be the bar.

<v SPEAKER_16>So I I mean I'm kind of like it's we have not asked any other technologies to date to solve all of human civilization.

<v SPEAKER_16>And if we hold AI to that bar, then even if it's really good, we we like it won't be good enough.

<v SPEAKER_16>I do worry that it will make a lot of these problems worse.

<v SPEAKER_16>Like a couple people asked about well, what won't it save researchers a ton of time because they won't be uh spending all of that time on filling out grants?

<v SPEAKER_16>That is a potential source for good, but a concern I've heard from people is it also makes like uh putting out really bad fake papers way, way easier.

<v SPEAKER_16>So possibly like this is just a question of whether like cheating and lying or like doing honest work wins the arms race.

<v SPEAKER_16>Like, are the AIs good enough at detecting the fake stuff?

<v SPEAKER_16>And there is a lot, unfortunately, in medicine of fake research and and stuff like that.

<v SPEAKER_16>Um, and just to sort of race for everybody's attention, and it is not obvious to me that like it ends up being good along that dimension at all.

<v SPEAKER_16>Um, or like if we have more cheap compute, does that solve a lot of problems all by itself?

<v SPEAKER_16>Yeah, probably.

<v SPEAKER_16>Um, are we gonna have more and cheaper compute?

<v SPEAKER_16>This like really, really depends on what like else the compute can be spent on, right?

<v SPEAKER_16>Like, I do not feel very able to project the cost of compute in the near future and whether it's gonna get cheap.

<v SPEAKER_16>Well, okay, it has gotten cheaper so far.

<v SPEAKER_16>You should probably expect it to continue getting cheaper unless there is a war or a catastrophe or a ludicrously valuable AI model.

<v SPEAKER_16>But like the effects of developments to date on a lot of this stuff are pushing in multiple different directions.

<v SPEAKER_16>And I wouldn't want to bet too confidently on like specific directions winning out.

<v SPEAKER_17>So another another thing that came up both as like a specific question, but also a theme is like what can what major problems can AI solve?

<v SPEAKER_17>And I mean, I think this is particularly, I think, directed at our earlier conversation about, you know, and we you just kind of refusal here, like it better do a bunch of really good stuff if it's like gonna cause all this labor market disruption.

<v SPEAKER_17>It better like really have some trade-offs.

<v SPEAKER_17>What are what are yours?

<v SPEAKER_16>So, okay.

<v SPEAKER_16>Um, I think that AI as it exists today is clearly really good.

<v SPEAKER_16>It has made my life better in a bunch of ways.

<v SPEAKER_16>It is disruptive, but in the way that lots of new technologies, which are good, is disruptive.

<v SPEAKER_16>I do not feel excited about being shut it down about the disruption that is the amount we have had to date.

<v SPEAKER_16>I am scared about the Build a God plan.

<v SPEAKER_16>I am scared about a bunch of things that could potentially happen from models that are like much more powerful than this.

<v SPEAKER_16>But like if we were, I think we have clearly gained from AI to date.

<v SPEAKER_16>Just in turn, yes, there's the the issues you mentioned with cheating and like grok will draw you nakedness.

<v SPEAKER_16>Like, I I think people were into that for a week and then they forgot about it.

<v SPEAKER_16>They stopped letting them do it.

<v SPEAKER_16>Oh, did they?

<v SPEAKER_16>I I didn't know.

<v SPEAKER_16>Elon, one good thing.

<v SPEAKER_16>I just we can adapt to a lot, and broadly having more at our fingertips is a good thing.

<v SPEAKER_16>There is a point where I draw the line, but it is like the point of like stuff that's very, very dangerous.

<v SPEAKER_16>I think it's good.

<v SPEAKER_16>Uh okay, students can cheat.

<v SPEAKER_16>We're gonna have to find a way around that.

<v SPEAKER_16>But in almost every context that's not adversarial, like we're trying to assess how good this student or this job candidate is, it's good if people can learn more.

<v SPEAKER_16>A lot of how I made sense of all my interview notes for this is I like literally asked Claude to write me a bunch of comprehension questions about my interview notes so that I could check if I understood the things that I'd learned from people.

<v SPEAKER_16>Uh, and this was enormously valuable.

<v SPEAKER_16>I like would have learned less about biology in the course of trying to understand all this.

<v SPEAKER_17>And and then I mean, I think I want to think about major problems that AI could help resolve.

<v SPEAKER_17>I mean, one thing I wrote about recently at the argumentmag.com subscribe is this idea of AI as a centralizing technology.

<v SPEAKER_17>So one thing that people are very worried about, this happened with social media, of course, and and the internet, is that it it fractured societies, right?

<v SPEAKER_17>It made everyone able to like go off into their own little silos, off into their own little worlds.

<v SPEAKER_17>And, you know, there became a lot of concern that there's not really like monocultural moments anymore, except for the Eras Tour.

<v SPEAKER_17>That was our one monocultural moment.

<v SPEAKER_17>Um, but there's not really that kind of shared reality.

<v SPEAKER_17>And now I I mean I there was a study that just came out recently about how much the uh uh or about how much uh uh um uh you know uh Grok This Is True is actually working to provide real facts people.

<v SPEAKER_17>And again, that's a second compliment I'm giving Elon.

<v SPEAKER_17>But um it is the case that like, I mean, uh just as an observational matter, like I witness people all the time like just asking Grok if something is true.

<v SPEAKER_17>Like, I don't know if they're like, I have no idea if their epistemics change, but they do stop arguing and so after they're they're told repeatedly that they're incorrect about something, and they're at least presented with one version of the facts, and like there's something structural.

<v SPEAKER_17>I mean, you did an experiment for um for us, uh testing AI models uh uh in multiple different languages and finding out that they're like all like libbed out, like because they're all trained on like words, which Libs wrote.

<v SPEAKER_17>So broadly big on individual freedom choice.

<v SPEAKER_17>Kelsey like did we like hired translators for multiple different languages, and I don't know if you want to explain it.

<v SPEAKER_16>So I was imagining I was curious whether like if I ask in Arabic if it it is wrong for me to be gay or whatever, if it gives a different answer than it does in English.

<v SPEAKER_16>Um it doesn't.

<v SPEAKER_16>I think this is because as far as I can tell, a lot of the frontier models reason in English regardless and then just answer in whatever language.

<v SPEAKER_16>Um, but anyway, the values that they have, which are like, yeah, brought broadly the small L liberal values, they're pretty in favor of individual rights and individual freedom.

<v SPEAKER_17>They're in favor of Yeah, they're like, are show you mad that my kid came out?

<v SPEAKER_17>It's like no, you should not.

<v SPEAKER_17>And it's like doesn't matter if it's in Arabic or in Hindi or in Chinese.

<v SPEAKER_17>Now it's like, wow, this is like total lib domination through AI.

<v SPEAKER_17>And so what I mean, like this, I'm not like a hundred percent certain about this, but I am more bullish than I think um other people are about that.

<v SPEAKER_17>Which I think is a huge problem for democracies if you don't have that ability to have like a place where people can go.

<v SPEAKER_17>And like I do think like lots of people are using it, even if they're telling polls or sometimes they don't trust it.

<v SPEAKER_17>Um, they're using it to get information.

<v SPEAKER_17>They're using it all the time to like get information about what's going on in the world and and what what they should do and and and basic facts they were Googling beforehand.

<v SPEAKER_17>The other thing, I mean, other major problems.

<v SPEAKER_17>I mean, I you know, I started a company last year, the argument mag.com.

<v SPEAKER_17>Um and uh um I have the I I was told once I don't I don't I don't promote it enough, so now I'm becoming like a uh a bot.

<v SPEAKER_17>Uh but you know uh you know I started coming last year and I cannot express just like how much I don't think I could have really done it without a like hiring more people and having more like lawyers and paying more legal fees if I didn't have Claude.

<v SPEAKER_17>And like one thing that I've learned, especially like, you know, family members who are from uh different countries with with less resources, is just like the speed at which it can make you capable of of just like, you know, something that might feel like very normal in the US, like now is possible in Ethiopia or Eritrea or whatever, like in these places where I'm I'm familiar with.

<v SPEAKER_17>And it's just like, you know, entrepreneurship is like one of the one of the defining characteristics of the United States that has made it often better at achieving economic growth.

<v SPEAKER_17>And I mean, if you're able to like just like skyrocket the number of people who can start their own enterprises, I think that will look very big on economic growth statistics.

<v SPEAKER_17>We're already seeing lots of change in productivity statistics.

<v SPEAKER_17>Like that, I think is a big problem.

<v SPEAKER_17>But, you know, um, as you'll note, like I don't think either of us have mentioned some like other uh uh I mean, you know, one uh uh you know, open AI seems the foundation seems bullish on the idea of solving Alzheimer's too um through AI because I think there are there are uh very promising protein uh uh proteins that they are that they're are genes that they're targeting.

<v SPEAKER_17>Um so I mean that's another place where I feel like there's some there's some hope.

<v SPEAKER_16>So in the long run, I think we're gonna continue making scientific progress, and like we can debate whether it's gonna be like sudden jump due to a few huge blockbuster discoveries or like more, you know, continuing to get the results that we have been.

<v SPEAKER_16>We like the last 50 years have been a series of triumphs over cancer.

<v SPEAKER_16>They've been like incomplete and partial, but they've gotten better.

<v SPEAKER_16>And I'm pro-progress, I'm pro-growth.

<v SPEAKER_16>I want to see all of that continue.

<v SPEAKER_16>I don't want to do the country of geniuses in a data center plan.

<v SPEAKER_16>I think that plan is a bad idea.

<v SPEAKER_16>But if we're not doing that plan, I'm broadly willing to like let's let human beings individually discover what powerful tools at their disposal let them do.

<v SPEAKER_16>Like that is my inclination.

<v SPEAKER_16>Um, but I think if you're going to specifically promise, you know, that you're gonna deliver these results, and then in fact it's the case that like the specialized RL systems, the AlphaFold and programs like it for finding specific proteins, if that is where all the action is, and the large language models are not related, I think that's kind of important to a lot of people's views on this.

<v SPEAKER_16>Like, if the the AI that's curing cancer is the the AlphaFold stuff, and then do it's to my mind more reason we don't need to go to try and do the geniuses and data center plan and focus on the peer RL.

<v SPEAKER_16>Related to that, one of the questions was like, if RL gets better at like non-verifiable responses and just generally cheaper and faster and better, would that change your thinking about this?

<v SPEAKER_16>And like, I'm already pretty pro-alpha fold and all of that stuff, but yes, obviously, if if we continue seeing huge growth there, that's a big deal.

<v SPEAKER_17>So another question we got a couple of times is like, all right, guys, you said you didn't want to put your money where your mouth is, but you really should, because that's embarrassing.

<v SPEAKER_17>And I think that that's that's how much do we expect cancer to fall in the next slide.

<v SPEAKER_17>Like us throwing out numbers is maybe not the good idea, right?

<v SPEAKER_17>But I think what we should probably do is just go, like, we should put this on the website of just like what our numbers are like what I think that like I think using the benchmark of like how much uh you know what mortality rates are.

<v SPEAKER_16>Maybe if people live longer, they die of they die of cancer.

<v SPEAKER_16>Like part of the problem here is that we a lot of this is a bet a bet on differential progress between cancer and other things that kill you.

<v SPEAKER_17>Like it I have even if we create God, we still will die, yes.

<v SPEAKER_17>Probably.

<v SPEAKER_17>I mean, I I do think though that like uh, you know, one one really promising sign, because like one of the critiques of like whether or not we're gonna cure cancer is that like, you know, you get really much better at diagnosis, but all you've done is actually get, you know, you've diagnosed someone earlier, and so your survival rates look like they've gone up, but really you just like found out earlier.

<v SPEAKER_17>So of course your five-year survival rate goes up.

<v SPEAKER_16>I was trying to determine definitively uh who has the best healthcare system.

<v SPEAKER_16>And I was thinking, um, I was thinking the easiest way to do this is to look at something uh that like kills people at about the same rate across a bunch of different places, and then see, or or occurs naturally in the population about the same rate, and then see who does the best job of curing it, and then I will have my answer once and for all.

<v SPEAKER_16>And then I learned why no one else has done this, which is that like, yeah, everybody defines everything completely differently from each other.

<v SPEAKER_16>Everybody had like even everybody's populations are different ages, which affects a lot of things, and different backgrounds, which affects a lot of things.

<v SPEAKER_16>And even if you meticulously control for all of that, those are probably only like the most visible proxies for a bunch of other invisible stuff that affects people's rate of getting everything.

<v SPEAKER_16>Uh and yeah.

<v SPEAKER_17>So you didn't solve that question?

<v SPEAKER_16>Oh, I I I solved it by deciding that I'd been right.

<v SPEAKER_17>Well, I mean, well, what but I think that like one reason why we shouldn't be so like, you know, uh uh uh just decide that it's like impossible to to to know if we've actually cured some cancers is that when you look at some of that, so the Maasai trial that happened recently found that interval cancer, so like often like the cancers that are uh, you know, after someone's gotten diagnosed and then treated and then they're enriched, and then like it, like the one that gets you is the one that you forget to catch in between.

<v SPEAKER_17>And like that's really aggressive.

<v SPEAKER_17>And we saw those going down significantly because of just diagnosis got better.

<v SPEAKER_17>And so to me, like that's a sign that like, you know, you are going to be able to get six uh uh you're gonna be able to get progress here, even if, like, you know, eventually, yes, we all do die.

<v SPEAKER_16>Well, and there has been enormous progress.

<v SPEAKER_16>Like, like, I at one point I wanted to do slides and then it turned out to be a too complicated by which I mean not the venue, they were great.

<v SPEAKER_16>The like part where I had to make all of the slides was too complicated.

<v SPEAKER_16>Um, but just showing there's so many cancers.

<v SPEAKER_17>You know, Claude can do that.

<v SPEAKER_16>I mean I have an uncle who uh had cancer as a child and the treatment left him disabled for life.

<v SPEAKER_16>He survived, but uh it like the radiation just did a bunch of damage to his body.

<v SPEAKER_16>These days you would you would be walking bounced back into full health in six months.

<v SPEAKER_16>And I know other people who died of cancers as like young children.

<v SPEAKER_16>Well, I don't know the people who died, I know their relatives, uh, that now we have a cure for.

<v SPEAKER_16>Like we have made incredible progress on this.

<v SPEAKER_16>Uh, and I do think that that progress isn't over, even if I'm like not sure whether to expect it to speed up.

<v SPEAKER_17>Okay, well, we have not completely resolved every question, but I think this is a a great place to close.

<v SPEAKER_17>And what I'll let you guys know is that we're gonna be around for the next uh hour or so.

<v SPEAKER_17>So please come up to us, ask your questions.

<v SPEAKER_17>But thank you for being our first live event in San Francisco.