65 billion dollars. That is roughly what Nebius was worth on the stock market when we recorded this episode. Yet most people have never heard of it. So what does the company actually do, and why are scientists using its computing power to search for new treatments?

Dr Ilya Burkov, Global Head of Scientific AI and Healthcare at Nebius, joins Beyond Longevity to explain how AI is being used in research and what happens behind the scenes when a biotech company needs far more computing power than it has in its own lab. Ilya has a PhD in medicine and worked in medical research before moving into technology at Amazon Web Services. He joined Nebius in 2024.

We talk about the difference between asking ChatGPT a question and using AI to investigate a potential drug. Ilya explains why access to computing power matters, what Nebius provides to research teams and large companies, and why a calculation that takes two days on a lab’s own equipment could take just 23 minutes with more powerful computing.

We also discuss AlphaFold, the protection of research data and intellectual property, and the limits of AI in drug discovery. If AI points scientists towards a promising molecule but cannot explain why it chose it, has it made a discovery? And could the same tools help us understand something as complicated as ageing?

Nebius Cloud for Scientific AI and Healthcare:

https://nebius.com/solutions/life-sciences-and-healthcare

Dr Ilya Burkov on LinkedIn:

https://www.linkedin.com/in/ilyaburkov/

00:00 Can AI Discover Without Understanding?

00:55 Meet Dr Ilya Burkov

02:08 Can AI Speed Up Scientific Research?

03:47 Why Did Dr Ilya Move from the Lab into Technology?

06:06 How Is Scientific AI Different from ChatGPT?

08:55 What Does AI Drug Discovery Actually Mean?

11:36 Why Do Biotech Investors Ask About Computing Power?

13:04 What Does Nebius Actually Do?

16:36 Does Nebius Provide the Kitchen or Help with the Cooking?

18:02 Why Does Biotech Need a Specialist AI Cloud?

20:48 How Are Researchers Using Nebius in Practice?

32:25 How Are Data and IP Protected in the Cloud?

33:47 What Did AlphaFold Change About Protein Folding?

37:43 Does AI Understand What It Discovers?

39:30 Is Ageing Perfect for AI or a Complete Nightmare?

44:35 How Do Europe, the US and Asia Regulate AI?

47:57 Why Are Companies Reserving Computing Power Early?

49:59 Rapid-Fire Questions

52:52 Closing Takeaways

Foreign.

Welcome to Beyond Longevity, the podcast that explores not just how we age, but how we can build a longer, healthier future for ourselves.

A calculator can give you the right answer without understanding math.

AI can do something similar with biology.

It can predict the shape of a protein, narrow down which possible drugs are worth testing, and find patterns in huge amounts of medical data.

It can produce those answers without knowing or understanding what they mean inside a living body.

So can AI discover something without understanding what it has discovered?

If it finds a drug that works, but cannot explain how it found it, would you take it?

My guest today is Dr. Ilya Burkov.

He is global head of healthcare and life sciences growth at Nebius and has a PhD in medicine.

Dr. Ilya spent almost five years doing research at Edinburgh Hospital in Cambridge and then moved into technology and worked at aws, Amazon's cloud computing business.

And if you've never heard of Nebius, think of it as a company that builds and runs the vast machinery behind AI.

Nebias puts that whole system together and keeps it running.

Its customers range from university research teams and startups to big biotechs and some of the world's largest companies.

But does faster science give us medicines or only more things to test?

Can access to this technology help a small biotech company compete with the pharmaceutical giant?

And could AI help us slow aging without ever understanding what aging actually is?

Dr. Ilya and I discuss all this and so much more.

Dr. Ilya, thank you so much for joining me on Beyond Longevity today.

Now, before we speak about Nebius, I want to start with you.

You began your career in medical research at the time you were trying to find early signs of diseases, including, I think, osteoarthritis and osteoporosis.

Today, researchers have access to AI and computing power and things that simply didn't exist then.

If you went back into the lab today, what could you do now using AI that you could not do then?

I could probably do my PhD in an afternoon.

To be honest, I'm overreacting there.

I would definitely save a few years worth of work in terms of what I was doing.

In general.

The thing that we're getting with the various tools and hardware and software advances in AI is the freeing up of time, which means that you can do a lot more in the equivalent time period.

So a typical PhD is three to four years in the UK.

I think that you could essentially squeeze in two or three PhDs in that time.

They were worth the work.

As you probably know, when you're doing A lot of this research, you're limited by budget, you're limited by time.

And if you can reduce the costs associated with that and increase the performance, I think for me that's what I would have gained.

Wow, that's quite a start here.

No, really, I didn't expect such a big difference because it's not like you did your PhD 50 years ago.

So, so, wow, quite an advance in, in, in such a short time.

So you eventually moved from carrying out the research yourself to working on the technology other scientists now use for their research.

What pulled you out of the lab and into technology?

I realized that a lot of the work that I was doing was on a local level, on a community level.

Even though some of the, the data that I was, that I had collected was in Iceland and I had, I think, a cohort of about 300 people in my study, it was still a fairly local geographic in terms of the data.

What I wanted to do is go beyond that and see what kind of impact I can have within the healthcare system and healthcare industry on a global scale.

For me, the best way would be to go into the industry side, into the business, into the technology side.

And I was very fond of my time at the nhs.

I worked there for almost five years in Adambrook's hospital in Cambridge.

But I felt like I could do more.

I felt like I could have a bigger impact on society if I were to leave the lab and the closed unit that I was working with in orthopedics.

Now that you've seen both sides, what do you think?

Scientists often misunderstand about AI, and maybe what do people coming from the tech side misunderstand about medicine?

The biggest issue is that they don't always speak the same language and you don't expect somebody from the healthcare industry to fit right in into the tech industry and vice versa.

What I have always said is you need to have this, either specialists in one or the other, combining and collaborating, or to have those individuals plus somebody like myself who is that bridge between the two industries, somebody who speaks the language, somebody who translates the necessities of what needs to be done.

Quite often the engineers don't really understand or are able to explain what, what needs to be done from a technical level to a clinician or to a nurse or to a researcher.

They don't really speak the same language.

Before we dive into what Nubias does and all that, let's go to a very basic point.

Most people listening, and certainly myself included, for us, AI means chatgpt.

You open an App.

You type a question and you know you receive a clever answer.

A biotech company using AI to discover a drug is doing something far more complicated.

What is the difference between playing with ChatGPT and using AI seriously to do science?

The story people know is is the model.

You rightly said it.

You know, the GPT this AlphaFold that the story most people miss out is that none of it exists without completely re architected compute layer underneath it.

And what that layer is, that's what actually determines who wins and who is going to be following behind within healthcare and life sciences.

And when we were talking about drug discovery in particular, it's probably the area where the gap between the old world and the new world is most dramatic.

Quite often we hear these two terms of a wet lab and a dry lab.

Wet lab being the laboratory where the experimentation is done and the dry lab being the computational power to validate, test, theorize, strategize and then do the research.

The combination of the two come together.

But because the old world was so brutally slow and expensive, traditional drug discovery takes 10, 15 years, often costs billions of dollars, and has quite a high failure rate as well and cost.

So most of that time is trial and error in the wet lab.

What you need to do is incorporate AI and that's the real contribution is to narrow the search space before you even touch a lab bench, predict protein structures and binding in a simulation.

So simulating the molecular interactions on a computer before that occurs, using AI tools and various models and so on, flagging, you know, likely toxicity or failure points early instead of at year eight or year nine, which would be very expensive and why the failure rate is so high.

So I would say it's not just a simple model, like a ChatGPT model that will be able to cure or to find all of these wonderful new drugs to cure diseases.

It's a combination of these tools and materials that we have together that will then help to progress this even further.

So it's not really like one model to rule them all.

It's a combination of different tools plus a model or two here and there.

So when we hear companies say that a drug was discovered by AI, that could mean, you know, AI created the molecule, or it could mean AI helped somewhere along the way.

So when somebody says AI discovered a drug, what exactly does it mean?

And is it as straightforward as it sounds?

Yeah, I mean it's the early discovery, the candidate selection phase, not the clinical trial and not the regulatory phase.

So the clinical trial and regulatory phase will still take time because it needs to be tested and made sure that it's safe.

What they've done is they've simulated and predicted what, what the likely outcome will be using the AI tools.

They've looked at compounds, they looked at molecules, how they're interacting, how they're bonding and so on, and were able to essentially narrow down that period, which could still take a number of years to do in a wet lab.

AI hasn't shortened the biology or human ethics review that still needs to be done.

It shortened the search for what's worth testing.

That's the way that I would position it.

Because if you don't know what to test for, you would test everything.

And that would take a lot of time and money.

And this is why it takes 10, 15, 20 years to do.

If you can reduce the pool of things that needs to be testing or worth testing, that would save a lot of time and money.

And when you're compute bound, you're running large scale molecular simulations and, and training foundation models on biological data at a scale that needs serious, serious infrastructure.

And this is exactly the kind of extreme scale simulations and foundation model training that purpose built cloud solutions are now optimizing for specifically drug discovery and development and so on.

Did AI change the discovery or mainly, and I say mainly very loosely, save time and money.

The analogy I like to say is think of it like a calculator.

You can still do one plus one in your head or on a piece of paper, or you can do it on a calculator.

It'll take probably the same amount of time to do it.

But if you're doing a very complicated calculation and you're using a calculator, that tool will then save you time and money.

So my clear answer to that question would be the saving time and money.

They have discovered a way to reduce the need to spend so much time within this progression.

And I think that for me personally, from what I've seen and the customers that I've been interacting with, it's around compressing those time periods into a shorter timeframe.

I want to later on sort of come back to investors and business plans.

But do you think some companies are under pressure to present themselves as AI businesses because that is what investors currently want to fund.

So I think a lot of the companies that I spoke to, they've now said when they're raising funds, when they're raising A series, A, B, C, even D. Now part of that raise has always got a point on what, what are you doing for compute quite often, not just AI, like how Are you incorporating AI into your solution?

But how have you guaranteed compute?

Have you made a reservation?

And quite often we're hearing that companies that have a fantastic product and fantastic idea when they come to raise millions of dollars, pounds, euros, whatever, and they don't have a compute strategy or a computer agreement in place, they will miss out on, on some of the funding because the investors realize that because of the current market on on compute and the constraints on, on GPUs and so on, that even if they invest in that company now, they might not be able to get their hands on, on the GPUs or on the hardware for another six months or, or 12 months because they haven't made a reservation.

So that's something that hasn't happened before.

This is really fresh in the market that we're seeing.

A company that wants to build serious AI.

They need somewhere to store its data.

It needs enormous computing powers to process it and the technical systems that hold everything together.

Most companies cannot sensibly build all of that for themselves.

So this is where nubius comes in.

If I come to nubius with a scientific problem, what do I give you, what do you do with it?

And what do I get back?

Nobody outside the industry thinks about infrastructure because it's invisible when it works.

But it's actually the bottleneck on almost everything we've just talked about.

Training foundation models on biological data, or running large scale molecular dynamic simulations.

The entire kind of genomic pipeline, whatever they're working on at scale, requires all of this compute at a level most research institutions or even medium sized biotechs can't build or maintain.

And the key word is maintain themselves as well.

Because it's not like you just set it up and it works, you need to maintain it, you need to make sure that it's working and so on.

This is why AI cloud companies matter so much, almost as much as the model developers, because someone has to provide the engineered from silicon to API infrastructure.

And we as a company, we get the GPUs from Nvidia, they make fantastic tech and everything else to make them run and to make them function is done by us in house.

And companies come to us to then essentially either move their on prem compute that they might have, you know, on a small number of GPUs and they want to scale it to a larger number that they cannot do in house, so they come to us.

The other scenario is if they're coming from a hyperscaler, typically they would have had quite a lot of compute credits offered to them and by the time the credits run out and the realization says that they have to start paying for this compute, they often come to us because they're getting a lot more bang for their buck, as they say.

And the quality of the compute for AI specific workloads is incomparable.

And then there's the compliance, because this is an industry where we need to deal with very sensitive data, regulation is important and so on.

You can't add on HIPAA or GDPR compliance onto a generic cloud setup.

And the fact that this needs to be supported in an audit, it has to be engineered into it.

And again, this is what Nebius offers.

We offer quite a robust computationally compliant service.

There's lots of certifications and ISO certificates and HIPAA and SoC2, type 2, SoC3 and various other things.

That means that they can store the data safely and securely.

We don't use that data for anything other than providing them with secure storage.

So nebius does not take anything out of that.

It's their storage.

It's all located in a specific country or data center of their choice, both from a GPU perspective, a CPU perspective, storage perspective.

It's.

It's all safe and secure with the agreement of, of the company at hand.

Nabis position is, is a good example of this trend.

Rather than being a generalist cloud like a lot of the hyperscalers tend to be, we do have an entire team under myself and within the company who are healthcare and life science specialists and we call it scientific AI and is dedicated to this, both training and inference aspect.

To put it in real layman's terms here, if a scientist comes to you and brings you the recipe, is Nebbyus providing the kitchen, the equipment, the electricity, or are you helping to cook?

So both.

We provide the location for them to do the cooking.

Imagine they're cooking in their own kitchen and they only have a couple of pots and pans.

What we've got is a Michelin star kitchen where they can have any choice of pots and pans that they need, any type of fuel, any type of food and produce that they need.

And we've got the help at hand, we've got the sous chefs in the kitchen who are there ready to jump in and help with the cooking.

So if a company would like to modify their model and say, make it more efficient or move it to a different type of gpu, we've got the team in hand to say, okay, you started with this older generation of GPUs, let us help you migrate to the newer type of GPUs, and we would bring in the sous chefs and the solution architects to make that happen.

Thank you for continuing my kitchen analogy so extremely well.

But it, you know, it does bring it to life and it's, you know, mustn't forget we have a lot of people listening that have no clue about AI or how this whole setup works.

So that's great.

Now, you've mentioned that you in particular are at the life science sort of section of Nebius.

But Nebias works with AI companies across different industries.

But it has built a dedicated healthcare and life science platform for organizations working with areas such as, you know, proteins, potential drugs, genomes, medical images.

I don't know what else.

What is different about the work that these companies bring?

In contrast to other AI companies that Nebia serves?

When you're talking about industry in general, a verticalization approach is key to success.

When you're building aluminium for material design, there's different grades.

You can have aeronautical aluminium, you can have medical aluminium, you can have the ones that you can manufacture some pens out of and so on.

And the quality differs.

It's the same material, but the quality is different.

It's the same.

For building cloud, we have at Nebius a verticalization approach because it is what the industry expects.

Not only are we using the GPUs that Nvidia creates, but we are combining them with the right software, the right hardware, the right people in the same room to make that a success.

And that's why the way that I've built the vertical at Nebius here, with the scientific AI and healthcare, is with all of those things in mind.

Because when a company comes to us and they say we need to do this, this and this, the keywords that they use really, really are translated by myself, by my, my peers in my team to be able to understand what exactly that they want to do, even if they themselves don't know what they need to do.

So sometimes they might say, I know I need GPUs, I know I need to do this type of analysis, but I don't even know where to get started.

So we will sit down and really plan around their growth and their journey.

And even if they have a V1 of a model, they want to fine tune it.

They want to use Token factory from, from Nebius Cloud to pick a model that's out there, a llama or a Deep Seq or a Nematron model, and they want to fine tune it for their particular workload and their particular use case and their particular data.

That's when you need that specialist in place.

And without having this verticalization and without having these specialists in place, they would just be talking to a company that, that provide bare metal compute or just have some GPUs, figure it out, do whatever you want with it.

And there's a lot of those kind of companies in the market too.

Can you give us a real life example of a company and what you helped them with?

I think from top of my mind there was university, ucsf.

We worked with them on making their model even more efficient.

So we were able to increase it.

I believe it was 40x in terms of model performance.

One of our cloud solution architects sat down with them, went through the analysis and were able to basically release the next version of that model with significant gains in performance.

When it comes to specific companies, I can mention a number of different teams, like Latent Labs, for example, they were very comfortable using H100, so the Hopper series of GPUs for a lot of their model training workloads.

And I saw that there was an opportunity for them to test the Blackwell generation.

So that's the newer type of chips that came out and more powerful and so on.

And post training they were able to show a 4x increase in the performance of the training cluster that they were working on.

That to them basically saves time and money and it reduces the number of compute hours that's necessary, which again reduces the costs associated with it.

And they publish that themselves on their websites.

If you look at Latent Labs Nebius, you'll see what performance gains they've had.

We've mentioned GPUs and most people will probably have never heard of it.

Gpu, which is short for graphic processing unit, came from the gaming world and they were built to calculate all the moving images in a computer game all at once.

Gosh, I'm happy I know that much.

But my knowledge doesn't go much further.

It turns out that this ability is extremely useful for AI.

How did a chip built for computer games become powerful enough to help discover medicines?

I think again, when you look at it, it's how they're being used.

When you're talking about CPUs, so the ones that you have in your computer, nothing to do with GPUs, they're very serious based.

The core reason for GPU success is the work that can be done in parallel.

Parallelism and training in neural network is mostly matrix multiplication.

So when you multiply and then add huge grids of numbers over and over, a CPU has very few powerful Cores designed to do very complex tasks one after the other, very quickly.

But a GPU has thousands of simpler cores designed to do the same simple operation, but on a number of data simultaneously.

That's the main difference.

So since a matrix multiplication is essentially the same operation repeated across many numbers, GPUs can chew through a lot of that data far faster than any CPU or what we were familiar with beforehand.

Often they can take, you know, ten hundred times faster for these workloads to be done quickly.

And the neural networks, which again designed by the way that the brain works, and again going back to how biology is structured, every layer in the, in the matrix multiply, followed by a simple nonlinear function.

And then these GPUs were like you said, originally built to these, rendering millions of pixels and polygons in the video game.

But now researchers realize that they basically repurposed this hardware for machine learning and thinking, much to how our brain works and how our neural networks are designed in our brain.

So again, for people like myself who are not too familiar with GPUs and all that, when people are talking about buying access to GPUs, what exactly does that mean?

Are they buying a certain amount of brain power or what?

Yeah, so when, when they're buying access to GPUs, if we're talking about the cloud, they're buying access to an unlimited scalable amount of brains that they need essentially.

And how many calculations need to be done in the room?

So I think the analogy I like to use, and quite often I bring up, is in schools and how, you know, imagine a high school student or a college student.

A CPU is like a very smart person solving one hard problem at a time.

A GPU is like a whole classroom of students solving lots of math problems all at once.

So what you're giving them access to when people are accessing the GPUs in the cloud is you're increasing the size of the classroom, you're increasing the size of the brain power and the computational compute.

So a larger job may require many more GPUs.

But I think it would be too easy for a non technical person to assume that 10,000 GPUs must automatically produce an answer that's 10,000 times better.

What exactly can thousands of GPUs do that 10 cannot?

Do they improve the answer?

Do they run more tests?

And is there a point where adding more is only costly and not really helpful?

Absolutely.

Like with anything, when you're doing all of these calculations, the quality of the output is only as good as the quality of the input.

And yes, you can get the calculations done faster, but it's not always necessary.

So that's why there is a buffet table of GPUs available.

And the type of GPUs that you need to be using might vary depending on the task that you have at hand.

So if you want something that is more inference based, you'd probably go for some of the more powerful GPUs.

If you want to go for something that doesn't require a lot of foundation model training, you probably go for something that's a bit lower on the GPU scale, but doesn't mean that it's worse.

It's just that means you don't need to do as many calculations at the same time and you don't need as much RAM or whatever it is that you're trying to do.

Essentially, we would sit down and discuss what a team is aiming to do.

Quite often they come to us with a yes, I need this type of gpu.

I'm certain about it.

When they talk to us about what they're trying to do and what kind of work that they're trying to get done, we might recommend a different type of GPU and say, you can use these, but you'll be better off financially and also in terms of time by using another type of GPUs and faster computation between.

This is why we can, we can get a lot of this done.

When you're outperforming all the tasks, it's interesting, but sometimes you don't need to take a Ferrari to go shopping.

You might want to just go in a car that has the most trunk space or volume in the boot.

You need to have the right tools for the right job.

So you've mentioned that Nebias has this buffet menu style, serves very different companies.

Presuming you have one customer who might be an excellent biologist with an important question and very little knowledge of AI, and another one might be, you know, an AI company that has already built an advanced model but can't run it at scale.

What would each of these two customers use Nebbys for?

And can you give us real life examples of those two type of companies?

Absolutely.

So I would say that when you're talking about the scientists, there's different types of scientists.

So some of them that are now adopting the use of machine learning and AI tools, and others that are classically trained and don't really understand how to use the technology side.

What I would always recommend to them is to either have somebody join Their team with a technical background to decipher what they need before they, they do anything within the GPU space because it's expensive to just play around with it.

You need to know what you're doing.

Quite often the first steps are on premises.

So the scientists will probably do something on their laptop or on their desktop computer and the GPU in there will do what's necessary.

If they feel that they are hitting a blocker and that their work is taking too long because of computational power, that is the stage where they would contact a company like us and say, okay, I want to do this experiment.

But it took me two days to get the results.

If I was to scale this up from a on prem compute to a cloud compute environment and I want to reduce the time, what would you suggest?

So then we would sit down with them and understand what they're trying to do.

And we could essentially say that if you add these powerful GPUs, those two days worth of computational time will take you 23 minutes.

Now that's what we would do.

An example of that would be many of the academic groups that we worked with.

So we did a project with Professor Lecong and his team in Stanford University.

They were using internal resources that they had at hand to build a tool that's called CRISPR GPT.

They were limited by the compute power.

They came to us, we had a nice discussion and we were able to to give them H1 hundreds for the workload that they needed.

That basically produced the first result of ana Nature Biotech publication for what they were trying to achieve in terms of gene editing and text based CRISPR gene editor that they built.

That's an example of a group, when it comes to companies that are coming in from a technical perspective and then need the scientific or the medical background again, I would always recommend that they partner up with a hospital or with a clinic or with a university or research group where the data comes from, whether it's the NHS or somewhere in the US or in Europe, to complement to their technical skills and their technical abilities.

Because yes, they might be good at the technical side, but they might not have enough of the data to be able to run a justifiable experiment.

So sometimes we are also working with teams who basically want to connect various projects that they're working on into a single unified platform.

And Primamente is an example of a company that we work with in out of London who, who basically gather neurodegenerative disease data from hundreds of different sites all around Europe, NHS Scotland, Wales, England and the us, Middle east and so on.

And previously it's always been scattered for them in terms of what access they can have and what data lives where and so on and so on.

Nebius, they were able to unite all of that and essentially have specific projects based on the COMPUTE that they have reserved with us.

So a number that they of GPUs that they've reserved, some of it is being used for this project with a group in, in Italy, this project with a group in England, this project with a group in San Francisco.

But the data is all in one place.

So that gives you kind of two examples, one both academic and scientific, and the other one more on the technical side.

Great IP is obviously very important.

A company may have years of private research, sensitive patient data and an AI model containing the idea on which the whole business depends on.

How can it use Nebius without giving away the family silver?

Nebius has no access to what the data looks like.

We provide the infrastructure layer, but we don't have insight into what is being done.

That is built by design.

We do not have access to what kind of information that they store, what the model looks like and so on.

That is all done at their own locations.

So they have access to cloud compute and they have access to cloud storage and so on, and only they have access to that.

So some, some teams are working with very sensitive data and they come to us and they say, not only do we need access in your data center, we need access to this particular data center and this particular rack, a specific part of the data center that is dedicated to us.

And we can guarantee that.

We can say yes.

Not only are we offering you this kind of very secure location which an average person can't get into, but we are also guaranteeing you a specific place within that data center for your data that nobody has access to, not even Nebius.

You mentioned AlphaFold before and I guess that's one of the most famous examples of AI changing biology.

So many people know the name Alphafold without really understanding what it does.

In a nutshell, it's.

Proteins in our body begin as chains of amino acids and fold into particular three dimensional shapes.

That shape affects how each protein behaves and alphafold predicts this shape.

Why was predicting that shape such an enormous scientific problem?

I think a lot of people started to, to investigate that once they were awarded the Nobel Prize.

A lot of people didn't realize what it was.

Even so, the, the the problem, as you, as you rightly said, the problem that it solves is that proteins are chain of amino acids, kind of like beads on a string.

But a protein doesn't work as, as a flat string.

It folds up into specific, very complicated 3D shapes.

And that shape, not just the string and the type of amino acids there, that shape itself determines what the protein does in the body, how it fights diseases or how it builds muscles, how it catalyzes reactions and so on.

So figuring out that 3D shape really used to require very slow, expensive lab techniques like x ray crystallography, and sometimes it would take years per protein.

Scientists had been trying to predict the shape computationally for decades of an individual protein.

And since the theory and sequences would determine the shape, it's a very, very complicated and hard puzzle to solve.

When you're talking about AlphaFold 2 and what it does, you essentially give it a sequence of amino acids, just like a list of letters in a text string.

It then predicts and really looks at the coordinates of every single atom in that folded protein and essentially it hands you the finished 3D structure.

Looking at something that was done before AlphaFold.

I think it was about 10% of all protein predictions were done and it took 60 years to do that.

And then once AlphaFold was released, it did the other 90% in the timeframe that it was released.

And that kind of acceleration is a huge, huge deal.

That's why they won the Nobel Prize.

It looks for the relatives inside of these proteins.

It searches huge databases, something that a human cannot do quickly.

It searches these huge databases for very similar protein sequences from other species and so on.

It uses attention like a language model.

It uses transformer based architecture, so similar to, you know, what powers models like ChatGPT and others to reason about why the relationship between every single pair of amino acids at once gives this kind of sequence and this kind of fold, something that again, a human brain cannot do very quickly or very easily.

It refines the structure and it looks at this in an iterative way.

So it doesn't just guess the shape, it builds this structure and it really simulates it.

And only then it outputs a 3D structure and a confidence score.

And a confidence score is quite important because that's the bit that tells the scientist which part of the prediction it's sure about and which part is shaky.

So then they need to rejiggle a few things.

And that's the biggest impact that it's had.

Without going into too technical details.

That's perfect.

So AlphaFold can make remarkably accurate predictions.

But predictions, discovery and understanding are very different things.

A calculator can produce the right answer without understanding maths in a way that a person does.

Can AI discover something without understanding what it has discovered?

I think that's a very good question and probably not one for me to answer.

I think it's a series of commands, it's a series of trainings.

Until we have AGI, I think it is just regurgitating what it's been taught.

So if you're asking it to do something, it'll find the best solution for what it thinks is the right solution.

And it's very convincing at it as well.

So even if it's wrong, it can be very convincing that it sounds like it's the right thing.

So I would be cautious.

But at the same time, it's like teaching a person.

If you teach them the wrong thing, then are they wrong in repeating what you've taught them?

You just taught them what they know.

So that's why it takes a lot of time to also validate.

Like when you are writing a scientific paper, even before AI days, if you have a statement, you would typically find one or two sources to back you up.

That's what you need to do with all of these tools as well.

If you're a lawyer and you want to bring in a case that is similar to your case that you're solving, you say that case in the year that it was done.

So then that's precedence for whatever has been done.

It's the same with AI.

So when you're trying to back it up, you need to say where that source of information is coming from.

Bringing this whole thing back to longevity.

Alphafod had a clearly defined task, protein folding.

Aging is not a clearly defined task.

It involves many biological system, decades of lifestyle and environment exposure.

And still there's no single agreed test showing that aging itself has slowed.

Does that make aging perfect for AI or a complete nightmare?

I think it's complementary when you're talking about longevity.

For me it's about not prolonging your lifetime to live to, I don't know, 150 years old, but the quality of a lifetime, you know, having disease free 130 years rather than surviving for 130 years.

It's very important longevity and is often talking about the length of time that you're alive.

What we are going to be incorporating with AI in general is the multiple aspects of it.

So you're going to be looking at, yes, living longer, but also living better quality of life.

So it deserves its own dedicated kind of teams.

And we've got entire companies, entire VC backed and even VC firms now looking at specifically just longevity when it comes to it.

And when you're talking about AI use cases within, within the longevity field, it's looking at proteins, it's looking at measuring biological age clocks, it's looking at multi omic risk scores and, and, and, and various other things.

It's that combined effort of all of those things that will provide this success.

And, and for me it's, you know, when you're looking at shifting from reactive medicine to predictive and preventative medicine, that's what's key.

And that shift is only possible because of everything we've walked through today.

Prevention is earlier, more precise risk detection.

If you want to live longer, you find and prevent any potential diseases before it occurs.

That means, you know, looking at genomics, looking at biomarkers, looking at imaging.

It means intervening before the disease progression rather than after a diagnosis because it's much easier to treat it before it becomes a diagnosis.

This is the single biggest structural change AI enables in healthcare economics because prevention is cheaper and much more effective than treatment, especially when you're scaling that up.

When it comes to aging, I would say longevity research is increasingly treating aging itself as a modifiable biological process with measurable markers rather than inevitability.

AI's role is making it feasible to actually measure and track those markers across large populations and much longer timeframes, which was previously impractical to do.

And then when you're talking about longevity in healthcare systems in general, the implications, personalized screenings, having those scheduled in instead of one size fits all is key.

AI assisted diagnostics is gonna be able to reduce a lot of specialist bottlenecks, especially in imaging.

But you know, in various other modalities and you know, we touched drug discovery, the drug development pipeline that can respond to rare conditions that are specific to a patient can only be done with AI.

Otherwise it's, economically, it doesn't make sense because the discovery cost curve is really complicated and none of this replaces the fundamentals.

As much AI you want to throw at it as possible.

You're still going to need sleep, you're going to need exercise, you're going to need diet, you're going to need social connection.

There's so many aspects of it, not just health that you can test in a lab, but more, much more than that.

And that to me remained the highest leverage of longevity interventions we have.

AI's contributions is compressing the time it takes to discover the next tier of interventions and personalizing how existing ones are applied.

But it's an accelerant on a foundation.

It's not a replacement for it.

It's accelerating it.

That's it.

And there's still much discussion about what really measures lifespan, health span and all that.

Do you think AI itself could help us to discover which measurements really predict longer, healthier life?

I think it can push us in the right direction, but it's not going to discover it for us.

It can definitely help us to rule certain things out quicker, but it won't say, look here, only here.

And this is the source of truth.

At least it shouldn't.

If people are looking at it for that, they're using it in the wrong way.

AI is obviously being developed globally, but it's not governed by one global system.

Europe, the US and individual Asian countries, they're all taking very different approaches.

Who is actually responsible, or is anybody actually responsible for deciding what AI should and should not be allowed to do?

And also, are the different rules in Europe, the US And Asia keeping us safer, or are they creating confusion while the technology sort of races ahead?

I think I can spend an hour just talking about this separately.

You need to come back, obviously.

I think the quick answer would be everybody's doing it differently because that's, that's what they think is right.

I wouldn't say one system is better than the other.

They've all got their own pluses and minuses.

Ultimately, I think we're heading in the right direction.

There's going to be a lot more regulation coming in as time progresses.

The automobile was built first, and then I think 50 or 60 years later, we got the airbags, we got the seat belts, we got all of these things.

It's not like they didn't want to do that.

It's just that they didn't realize that they needed to do until people started having accidents and started having situations that they were going too fast and, and couldn't prevent it.

So I think it's a natural progression.

The one thing that a lot of these regulators and countries should be using is a lot of the tools that they're regulating.

So if they can speed up the process, everybody would agree that would be good.

But in terms of where we are at the moment and who's got a better.

I wouldn't say that, you know, anyone is, is leap years ahead of anybody else.

Europe is often said to be overregulated, but at the same time, they're much more sensitive with the data and how that Data is being used and mined and so on.

US tends to be less sensitive about the data but is more quick in terms of achieving the results that they need at the in a shorter time frame because they, they don't look at it in the same way that Europe does.

APJ and Asia in general have a very different way of looking at it and for their own reasons.

A lot of the data that Chat GPT is trained on or Western models are trained on or European models are trained on, American models are trained on does not incorporate data from Asia or Africa or the Middle East.

They need to do it their own way because it's their population data and I'm talking about from a healthcare perspective but you can apply this to other aspects as well.

There are certain diseases that are more prevalent in certain geographies in this world and various intolerances and allergies and so on.

We don't have a bigger picture if there's only one country or one area of the geography that is doing this research.

It's not enough to be widely used across the world.

I believe that everybody should be doing their own models and their own safety and security and how they feel around their culture, around their society, around what's right and wrong within their environment.

And I don't think that one area or geography has the right morally or otherwise to dictate what needs to be done in another geography.

I think where we're heading is the right approach.

Everybody knows what needs to be done.

I think it's just a matter of time.

We'll get there.

So before I ask you the five rapid fire questions that I ask all my guests, I want to ask you one last question.

What have you noticed have customers started asking Nebias for that they were not asking for two years ago?

And what do you think if you can make a prediction they will be asking for in two years time that they're not asking for now?

Absolutely.

Reservations.

So access to compute well in advance to the time that they actually need to start using it.

Before it was always ad hoc, a few reservations here and there but a lot of the time now they come to us and they say we need this kind of compute for this time of period and that is the only way that we can get them the compute that they require because the way that the market is so oversaturated that we have very limited capacity as a market, not as nebbyous but as a market that there's a lot more demand than there is supply.

So we're seeing a lot more Companies coming to us and saying this is our long term vision, how can you help us build it out?

And we've seen this, I'd say in the last six months, much more than ever before, planning together, reserving capacity and getting access for the future build out.

And what do you think your prediction is?

That people will be coming to you in two years time asking you for what?

For a lot more computer, even more, even more computes.

But they would have a roadmap of this is what my next six months looks like, this is what my next year looks like, this is what my next two years look like, this is what I think I need.

Let's sit down and really strategize what we've got going for us.

And it's really sitting down with us and building out their compute strategy together with us.

And I think even in two years time, this is going to be even much, much more.

Sounds great.

Well listen, Dr. Elia, thank you so much for joining me on Beyond Longevity today.

Before I let you go, here are the five rapid fire questions I ask everybody.

What's the single best piece of advice you would give your younger self?

Be patient.

That's an advice I still give myself now.

Never mind when I was younger.

Name one habit everyone should adopt for a longer, healthier life.

Naps.

Daytime naps.

So siestas, Southern Europe have got it right.

Even if it's a 10 minute cat nap, it goes a long way.

They're definitely onto something.

If you weren't in longevity science, longevity tech, what career would you have chosen?

I would have been an astronaut, but I'm 6 foot 5 so I think that's going to be difficult.

What microdose habit, sort of five minute routine or small daily action yields outsized longevity benefits?

I know it doesn't take five minutes, but I would say at least twice a year to do your bloods, to do all the testing essential to keep an eye on things.

And yeah, the actual blood taking would take five minutes, the analysis would take longer.

Yes.

But doing that and looking at it in a preventative manner, no matter how much money you spend, that's the best advice that I would give.

Keep an eye on your health, regularly check in on your bloods and what's.

The craziest longevity myth you've encountered and is there any truth to it?

Ah, I think a lot of these holistic approaches eat this, drink this, this type of tea, this kind of mushroom, this kind of thing.

There's just so much out there.

I hear it all the time and for me, that's, that's, that's funny.

I mean, there, there's a lot of holistic medicines and, and things that, that people approach, I would always caution and, and say, first of all, look at the studies.

And second of all, if you've got a couple of people saying it doesn't mean that it's true.

So there's, there's so many myths and I, I, I, I, I can't, I can spend half the afternoon, you know, this type of tree mushroom is, is good for this brain activity and function and so on.

Just balance, keep a balance of everything in your life and don't do anything extreme.

Very true.

Thank you so much for joining us today, Dr. Ilya.

It was so fascinating and so different to what we usually discuss on Beyond Longevity, but very much needed, I think, because so many people that come on the podcast talk about AI in one shape or another, you know, within their work, if it's clinicians, if it's researchers or whatever.

So it was really good to hear it from the horse's mouth how it all works.

Thank you.

My pleasure to be here.

We are only beginning to understand what happens when the power of AI meets some of the biggest unanswered questions in medicine.

AI could change not only how quickly we discover new medicines, but what we are able to discover.

Scientists can explore thousands of possibilities, spot connections they might otherwise miss, and tackle questions previously beyond their reach.

But this requires enormous computing power, and that is where companies like Nebias come in, providing the technology that makes this research possible for longevity.

The possibilities go far beyond developing new drugs.

AI could help us understand why we age, identify diseases earlier, and find entirely new ways to prevent them.

Of course, AI cannot replace clinical trials or guarantee that a discovery will become a successful treatment, but it could fundamentally change how quickly medicine advances and what we can achieve.

Many thanks to Dr. Ilya Berkov for joining me today on Beyond Longevity.

And thank you for listening.

Please follow and subscribe.

SA.