Discussion on how we can all start using our AI skills to save lives

Introduction

So I think my intention with this presentation is just to insert a general idea into your heads. If we're really stretch-goaling it, then it would be great to discuss after the presentation or even in the Q&A how we think we can start to implement this idea. But yeah, so I'll try to give you context without spending too much time on it.

About Humanity and Its Mission

So quickly, what I do, run a company called Humanity.

Our focus is completely to allow people to monitor the rate of aging and then guide them to slow it down, whether it be directly or through our APIs. Our mission is to give a billion years of health back to humans by the end of the decade. So it's not about extending life specifically, it's more about extending healthy years where we can all enjoy ourselves and our families.

Background of the Founders

My background, wanted to be an astronaut. My co-founder wanted to be an entrepreneur. He reached his dreams. I'll get there one day.

I randomly created a dating site that became very large. And so, you know, depending on your experience with dating apps, you're either welcome or sorry.

And then went out to the valley and joined up and basically started, it didn't start, but joined up with a consumer VPN that got to about 900 million users. So different mission, but again, taking technology and expanding it.

Personal Motivations Behind Humanity

1Unfortunately, me and my co-founder experienced something that gets a lot of people into the health tech space. On my side, it was a relative found out stage four about breast cancer. and passed away within about a year. On my co-founder side, Pete's side, ended up, his dad had a stroke at 61, just out of the blue, was in a wheelchair for five years and then passed away.

And so you learn that the cliche is unfortunately very true. If you don't have your health, you don't have anything. And so we didn't immediately the next day start a health tech startup, but that became a bit of a obsession for both of us.

The Journey to Humanity

And then, you know, Unfortunately, we all experience these things.

My dad passed away five years ago, and it was, again, kind of this feeling of helplessness. What was lucky for me, maybe because of beginning with arrogance and then finding out that people are very kind, is I was looking at early detection of cancer for obvious reasons.

Exploring Early Cancer Detection

And I kind of, you know, normal founder stuff, like typed in, like, I heard that, you know, genetics was going to play a part in allowing us to screen for cancer better. And I was like, you know, top geneticists, I'm going to talk to these people, typed it into Google. Met up with George Church. He's, him and Craig Venter kind of pushed forward the whole mapping of the human genome years ago.

And then throughout that time, kind of continually found myself in the position of people being very kind with their time. And that led to really the understanding that we could actually, if we went after cancer, we'd still die of a heart attack on average like three years later.

The Genesis of Humanity

But there was two things that we could do, is one, we could go after that base thing, which is this loss of function in our bodies that leads to all these chronic diseases, we call it aging. One, we could measure it, and one, there was stuff that we could already do about it. And so that led to humanity and we created that and just kind of going straight at the problem.

What me and Pete know how to do is we know how to get millions of people to use something and to change their behavior because of it. And that was kind of the one missing piece in the space, luckily. So we found our space in it.

Driving Behavioral Change Through Feedback Loops

On a positive side, just because I know there's a lot of nihilism in health, is one, we found that almost everybody wants to be healthier. There's really no any of those groups that you would think, you know, oh, they don't care. We haven't found that.

The other is that this feedback loop of, is your biological age, is your functional age getting better or worse? Giving people a feedback loop actually does change their behavior, which is obvious. We don't argue about it when it's like, oh, Instagram addicted me. But when it's health, a lot of times we think that it can't. So we can do something about it and giving people feedback loops can do things about it.

The Role of AI in Health Tech

What can us as people that are looking at AI, what can we do about it? So I didn't invent this concept. I probably pushed for coining this term harder than others.

Learning from Language Models

So the idea is really that LLMs have actually shown us something that's quite different than the way that we do all other health algorithm modeling. Really, almost all other modeling, but all other health health model you know modeling and so i'll go into that in a second but it wasn't until if anybody has seen any andre carpathy's videos he just explains llms so well to a layperson and it wasn't until watching one of his videos just a few months ago that i i understood that oh we're missing we've basically been shown with llms that we're actually missing this major pre-training step in almost every other space, but definitely the medical space.

And so that's what I'm looking at. Just to define the problem, if you categorize it in a very simple way, 90% of the spending that happens in health really never helps a person, it's just gathering data. because we have this old method for doing algorithms and it takes a ton of data every single time. People talk about siloed data and that's a big part of the problem, but it's every single time we're kind of having to create new data or going and finding new data or buying new data.

So if you think about health tech, if you went around to all the health tech startups in London, they're probably spending at least their first $3 million, 3 million pounds of their funding just getting their algorithm and the data enough to do the algorithm. And so every single time, if you're investors in the room, every single time you're handing money to people, they're spending almost all of it just getting to that first point, whereas I think we could do it so everybody starts already past that point.

The Impact of Data Accessibility in Health

This is a pretty simple one. Lack of access to health data is killing people. Couldn't be more true.

If we talk about female health, we're just constantly talking about not having data. And because of really bad decisions in the past, we didn't even collect that data. And it wasn't until very, very recently, crazily recently, that we started almost mandating that we had to.

This one you may agree with or not, but just to suspend disbelief, I very much agree with it. 1All the data that we need to solve all health problems exists on Earth today. And so this, we don't need to find new data, we don't need to create new ways to get data, the data exists on servers today.

Couple trends in getting to the end of the contextual part. A couple trends is we can open up data in a different way than we could in the past.

In the past, synthetic data was used for augmentation, if anybody's used that for modeling. Now synthetic data, Google Cloud team and a bunch of other people have shown that synthetic data can be just as good and completely private.

LLMs, I mean, you guys know this very well. Every single time we say, oh, well, it needs to be done this way, two weeks later, someone releases some research saying, oh, no, actually, we can do it with 10 times less data or 10 times less processing power. Might not be good for NVIDIA stock if you're into strategy.

I'll skip over this for time, but basically, if anybody has a question of some of those methods for creating synthetic data, could walk through that. basically they showed that they could create, they could take a whole electronic health record, create a synthetic version of it, then they tested by training a model on the real data, the original data, and then completely separately on the synthetic data, and they came to very, very similar answers, as you can kind of see by the bars there.

Understanding the Impact of LLMs

So, the magic of LLMs, which I learned, again, by listening to Andrej Karpathy, But that's not what happened with LLMs. LLMs, they basically set up, and I'll use the word arbitrary. I don't think it's too strong a word, but it's just for effect.

So what happens when we're training models, and still to this day in humanity, until very recently, is you're basically saying, hey, we need this very, very highly curated, very carefully collected data set, because hey, we're gonna use this for people's health, so we need to be really careful there's not spurious data in there. We're spending millions on getting that data together, sometimes hundreds of millions if you're talking about human trials, and then we're training an algorithm on it,

They took an arbitrary kind of swath, large swath of the internet, and trained on that. And they got to this base parameter file, right? And so when they first tried to use that base parameter file, they kind of got what you would expect to get is, you know, hey, it kind of knows what an email looks like, but it's really random.

So it's kind of like nonsense. And everybody's like, oh, that was definitely still an interesting direction of study. And so they kept going with it.

The magic started to happen when they then one day decided, okay, well, we're going to take this finely tuned, you know, curated data set. This is what we start with in health, but they took it as a step two. We're going to take that. We're going to put it on top of that base parameter file.

And what they kind of magically found, I'm simplifying, is that it kind of clicked it into place in a way that it made the prediction power, the standard error dropped by 100x, if you want to think in health terms, but it basically made that small data set about 100x more powerful to the point where you would start getting announcements saying, oh, maybe it's alive. And obviously that's great for, page clicks, but actually some people were like, really did not expect it to have that power.

Proposing a New Approach to Health

And because they had done a lot of stuff with small data sets in the past and it never happened with that, and then they put it onto this arbitrary kind of base parameter file and magically it kind of exploded in its ability to understand. And so what am I proposing is, I think we've stumbled upon something pretty, I guess we're recording, pretty effing amazing. And so let's apply this to health.

And yes, we can apply LLMs to health, and people are doing that, and that's great. But let's go a step further.

The Challenge of Data Access and Utilization

In the past, we were spending like $100 million, and there's some great companies, again, getting this highly curated data and trying to get it together. But the problem with that is that then you need to somehow open that in a way, You know, my friend Daniel was talking about like, he sees his job as just making it actually, he needs to actually make it usable by the lawyer. The problem with having this hundred million kind of like, you know, data set is you basically no one gets access to it really.

And there's no way for like you in the audience, like raise your hand if you've trained a medical model. with any of your kind of machine learning skills, yeah. And it's, I think I saw like three.

You know, raise your hand if you've like trained on a text or like predictive text kind of thing, you know, you're gonna get, was that like 20, 25. And so we're sitting around with all these great skills and we're not allowing people to interact with health.

Exploratory Approaches to Health Modeling

And so this is an exploratory thing, but if we basically went line for line with how we created LLMs, say, oh, okay, well, it's not the same thing. Well, okay, so it's a string and you have a vector. That vector leads to another string. Okay, how's that work if it's biological data?

Okay, maybe we need a label if it's a blood set of 500,000 people's blood markers. Maybe we need to switch the string to that. Maybe we need to put an extra label there.

But we're kind of watching the world work on a problem that will actually kind of put the whole road in front of us for how we can basically rediscover how to change everything in health. That's my assumption.

Open Sourcing AI and Its Implications for Health

And so if you look at AI, I meant to look up when this paper came out, but if anybody, you know, Basically, open sourcing AI came out of, obviously it's been worked on by a lot of people, but one paper just a few years ago, not that long ago, basically open sourcing in AI has accelerated everything. If anybody knows anything about the software industry, you're not using any piece of software en masse that's not built on open sourcing, right? And yet in health, very, very little, especially if you're talking about the data.

Building Tools for the Future of Health Tech

And so in the end, just to make this very relevant to you is what we should be doing is building tools. And I think it's large biological model, LLM, but LBM, and we should make it so everybody in this room next time that's working with AI or working with machine learning raises their hand when they say, have you ever worked on a, on a health model, and I think LBMs are probably one of the fastest way to get there.

Advocacy for Net Neutrality and Closing Thoughts

I always like to show this photo to Humblebrag. I got to hang out with, in a different life when I was running that consumer VPN, me and Tim Berners-Lee, who created the World Wide Web, got to go around DC and basically fight for net neutrality.

And he gets asked by every single senator, what do you think the impact of your invention was? time and time again, 30 times a day. His main answer was, first of all, I didn't create the internet. He also needs to say that about 100 times a day. I didn't start the internet.

But I created, my team created the way that, you know, the internet before was like academies and government offices and just a few of them. And it was a great way for them to exchange information. But the other, you know, someone said 8 billion, maybe it was like 5 billion at the time. The other four or 5 billion people had nothing to do with the internet, right? They couldn't touch it.

And so the World Wide Web was the thing that actually created the impact of the internet. And so it's great if we have things in the health space that really smart and really great people are working on, but we need to unleash that to the 8 billion people if we really want to speed things up. And I'd love to talk to anybody who wants to discuss that further.

Finished reading?