Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
Proving Humanity and Ai Autonomy
Recorded: March 21, 2025 Duration: 0:58:01Space Recording
Short Summary
In a dynamic discussion, industry leaders from Worldcoin and Tools for Humanity explored groundbreaking projects in digital identity, the role of AI in future economies, and the importance of privacy-preserving technologies, signaling a transformative era in the crypto landscape.
Full Transcription
Great. I'm very, very excited for this one as well. We've got two great friends here.
So let me just start with quick introductions. My name is Vangelis. I'm with the research
arm at CELabs, working on scaling the EVM alongside my colleague Ben, who's a cryptographer, mathematician, computer scientist, amongst other things,
also at, say, Labs on the research arm.
And today, we're very, very lucky to have a couple of friends
and very, very sharp people on our space.
We've got Anastasios.
Let me start with Anastasios' intro. I'm going to try to do justice
to both and then we can get started. So Anastasios has a PhD from Imperial College London specializing
in software reliability, fuzzing and binary rewriting. Graduated from the computer science
department of Aristotle University of Thessaloniki and holds a master's degree in advanced computing from Imperial
because a PhD wasn't enough. So he started at CERN in Switzerland, worked
about four years on high-performance computing, cloud and grid infra, also
worked at different companies like Box, Microsoft Research, Google, Facebook,
and focused on cloud security, AI in medical imaging, distributed systems, and debugger development.
And he just dropped, he's going to be back on, I'm carrying on.
So in 2018, Anastasios also founded a startup called Experiment with the goal of democratizing
drug discovery research.
And today he's a senior staff software engineer at Worldcoin, well, World now, where he leads
the company's technological readiness, focusing on anonymous digital identities that are CBO
resistant, utilizing ZKPs, multi-party computation, and biometrics.
There we go.
And he's back on.
Yeah, and apart from his work at World,
he's also an angel investor in more than 20 startups,
and also an LP in a couple of funds,
a very, very supportive of the startup ecosystem.
Yeah, and then we've got Dominic, also known as DC Builder. I'm very certain that he is a polymath,
generally self-taught for what I understand, very good in mathematics, computer
science, cryptography, and many, many other things.
Started as a researcher at Morales,
worked alongside, then joined Tools for Humanity
as a research engineer, then moved to World Foundation,
working on the World Foundation grants program, Protocol R&D,
decentralizing the world ecosystem and then of course supporting the world mission.
He's also an angel in many crypto projects and not only loves biology as well,
biohacking space amongst other things.
And yeah, and super, super happy to have you guys here.
Feel free to, I mean,
Yeah, I think it would make a lot of sense
to start a little bit about, you know,
yourselves, why you joined World Tools for Humanity,
and then tell us a little more about the vision
and the goal and the problems that you are solving.
Can we start? Yeah, go first, yeah. Okay, I can start. So, I mean, the reason I joined TfH
early on, I was at TfH before I moved to the foundation, but that's a detail. So the reason I joined was I was looking for a place
where I can grow as a person and also just work on something
impactful and meaningful.
And I was also looking on a specific focus
where I got really interested in cryptography, more specifically
zero-knowledge cryptography.
And after looking for a while, the only place
that I felt like I had a lot of really
talented people and individuals that I met that were working on a mission that is so driven and
ambitious is like the only place that has any, like the only place that I found that has the
level of ambition that I had at the time was World. And so it was the most exciting place.
Like when I met the people at the event,
and you meet people like Anastasios,
and they have like 30 Anastasios next to each other,
it's really something.
So I was super excited to join and learn
from a lot of these people.
Because when I joined, I was 20 years old.
No, mind you, I'm 23 now.
But yeah, when I was 20, it was the best learning opportunity for me. And yeah, that's one I'm 23 now. But yeah, like when I was 20, like it was the best learning opportunity
for me. And yeah, that's like one of the major reasons, right? Like you can work on something
that's impactful, meaning like distributing zero knowledge access to everyone through like
digital identity and finance. And I just could work on everything I was interested in at the time,
Ethereum Core, cryptography, blockchain design,
so I could work with Solidity and Rust which were things that I was interested in. Just all around
was the perfect fit for me at the time. My story is a little bit more funny. I was working at
Facebook before I joined Wallcoin and Tools for Humanity. It was actually a funny story because
and Tools for Humanity. It was actually a funny story because I was just bored there,
bored quite a lot in my previous job and I was looking for something new and exciting.
And I mean, Vagelis knows my story. I attempted to start my own company before my tenure
actually at Facebook. So I kind of had this bugging me, like I wanted, you know, to feel
the rush of a new startup.
I wanted, you know, to feel the rush and the, the hustle that requires, you know, to up
bring something very, very early on.
And I was looking, you know, for something new.
So I started interviewing with random companies around and somewhere around you, these random,
this, this, this random set of, of companies I was interviewing, there was one email from
a company that had a very funny name and they were giving me a list of all the hedge funds.
Like, you know, do you want to interview for Citadel? Do you want to interview, you know,
for all of these standard big names in the industry? Personally, I don't care about this
kind of business, to be honest. I'm not into the FinTech and funds and stuff like that.
I was mainly looking at something in biology,
very well aligned with what DC also likes.
But for some reason, because I knew that
Sam was actually directly involved into WorldCoin,
I said, okay, well, this is actually interesting.
And WorldCoin, I think was like number 10 or number 11
in this list of, you know, I can introduce you to all these startups and big companies and Wallcoin was like number 11
or something, completely, you know, not very highlighted or anything like that.
And I said, yeah, sure.
Let's try.
I had 11 interviews and it was a very nice, it was a very fun experience because the
first two were canceled literally the moment I joined the interview.
So it was very funny, you know, to see this very, very early on situation that everything is on fire.
Sometimes you even forget your own candidates because I joined quite early in the company.
And eventually, actually, the more I start talking to everyone inside there,
I realized the opportunity there is humongous.
I mean, the mission is huge. The people there are awesome. I can learn and I can grow so much. And what really caught my eye,
and I think this is, you know, one of the most important aspects, at least for me, with respect
to what we are building here, is that I think for humanity right now, there are two imminent, very, very important objectives that we need to do as a civilization.
One is AGI or ASI, whatever you want.
It's just some kind of super or very, very advanced artificial intelligence.
And then the second, almost equally important, I mean, depending on the first, of course, but equally important for humans, is the infrastructure for humans for us to coexist with AI next to.
So what we do here is that we are building exactly this fundamental infrastructure
on how humanity can coexist with very, very advanced AI.
And that was literally the thing that really switched my mind regarding what fintech
and what we are doing here, why it's so important and what is the big vision behind all this?
Because you can do so much. You can do identity, you can do UBI.
There are a lot of interesting applications,
but the fundamental thing that we are striving to provide here is the next
infrastructure required for humans to coexist with AGI.
Got you, got you. So what are the assumptions?
What's the premise here?
Is that AI is going to be so much more productive and efficient that most of content is going
to be created, most of the economic activity is going to be carried out by AI, so we have
to know whether, you know, what's AI and what's human?
What are the assumptions here?
Please, yeah. Anyone?
One of the assumptions you make is that as AIs get better and better at reasoning, they're able to essentially do anything that a human can virtually do.
And the more APIs or like the more things you give them access to do, like whether it's a robot API
and they can move in the real world,
or whether they can interact with any of the services,
social media, et cetera.
At one point, if you're a human using the internet,
for you it's literally impossible to teleport
what is human, what is from a bot,
because the bots are so intelligent,
they're able to do anything.
And so there's a lot of these theories, like the dead internet theory, for example,
where if you're a human, at the end, all you're going to see
is just bots, because no one can discern what is one
from another.
And so if you want to prevent that,
and you want to allow humans to still exist in a way where
they can use the internet just as they did up until now.
You need a way to prove you're human.
Not just that, but there's a lot of other verticals to it.
You have, for example, the UBI experiment where if you're.
I think we might have lost DC for a sec here. Our connectivity is today's.
Yeah, I think we're kind of spread across the globe and there's a lot of, I mean, I'm
not sure if this is kind of a VPN, probably not. So there's some random hop somewhere.
I can take over from you.
Yes. Yeah, yeah.
But thank you.
Yeah, my personal opinion in this is I think it's very well aligned with what this is.
I think, yeah, the big idea is that any kind of economic activity and not only, I think is going to
be far, far deeper on what AI is coming to replace or enhance or fundamentally alter.
Oh, this is back.
So yeah, sorry.
This thing, go on, man.
Yeah, sorry.
Apparently, I just need to be on my mobile data instead of the hotel Wi-Fi.
It's just auto-connected to the Wi-Fi.
So essentially when AIs do also mostly economic activity in the world, it's really hard for
you to create an effort system for humans in that environment.
And that's sort of like the post-car city assumption.
In a world where you assume that all the price of goods and services goes to zero because their availability just skyrockets exponentially. It's really hard for essentially
giving those services to everyone unless you have a way to prove you're human. Because
if the economic activity cannot be distributed to any human, be that because they don't have
access to a self-souven means of finance or self-souven means of digital identity and
proof of personhood,
then it's really impossible to coexist or to build a system that actually solves the problems of
tomorrow. In that case scenario where you have AGI. Okay, and now kind of asking some
in need questions just to get them out of the way? And why won't government passwords do?
What's the issue with the system that we've got so far?
Either.
I think one of the premises again,
I think on Tools for Humanity is that we want to be inclusive.
The current situation with government passports or government IDs, even though I would like to argue that on the 21st century and the current world as it is right now, I think government issued paperwork kind of exists
to the world is not that is that widespread to be honest,
but I mean, it's not everywhere, but it's quite widespread.
There's still a lot of issues with respect,
what is a real ID, what is a fake ID,
what is a real passport,
how easy you can fraud and create fake ones,
which actually they do create,
you know, some kind of fundamental question, how much you can rely on that kind of paperwork.
But there's also another interesting aspect, which I think we highlight a lot in tools
for humanity, which is the privacy aspect, like, why do I need to give my government
ID or some kind of, you know, important personal details to anybody to give me some
kind of digital service.
So I think this is the kind of sweet spot that we are trying to strive for here, which
is we want to create digital IDs that are anonymous.
Well, actually not exactly anonymous, but they are very private preserving.
And on the other side, unique and very hard to fraudulent.
So yeah, this is why we are based on biometrics.
So we can make sure to a very, very high extent
that you cannot create a second identity
because changing your biometrics is something hard,
painful and expensive.
The other is that we are very privacy concerned.
So we actually, we don't give any information about our users,
no their activity, no their names, nothing about them, other than the fact that they're just a unique user. And the third part is
inclusive enough, because yeah, you don't need to bring us, I mean, Vagelis, you're quite familiar with the processes we have in Greece, if you want to update, for example, your bank account, you need to bring like 20 papers from 20 different Greek government departments or something.
And then maybe, maybe if you're lucky enough and you don't have some 21st hidden secret
level paperwork that you need to find, they might allow you to change your phone number
or your name or something. So all of these things, they're just very inefficient. And for at least the way that
I personally believe, and I think a lot of people in our context, they do believe the same,
AGI is going to come very, very fast and it's going to hit very hard. So it's not like you can
wait for passports or IDs to catch up with all these technological
advancements that we're going to see in the next couple years, if not months as they come.
So, yeah, this is why we chose this kind of direction and we think the added value that
we bring to the world is how easy it is to do these kind of technological advancements. Yeah, absolutely. And then with regards to UBI, universal basic
income, so are you guys in the view that this is going to be
absolutely necessary? Or where do you stand this?
They say I want to hear your opinion about this.
This year, I want to hear your opinion about this.
I'm very curious as well.
So I mean, for me, I mean, maybe I'm not the right person to ask this because I have a
different opinion than most people.
Yeah, like my personal opinion is that I think it's an important experiment to run, right?
Like which economic model will work post-AGI is really hard to know. It's not going to be an economic
model per se, it's more going to be like a baseline social security model in my way,
like at least this way, right? If AGI makes all goods and services barely free and the
purchasing power of humans is like 100x, 1000x, and they can buy anything literally, they
can buy a trip to Mars Mars they can do whatever they want
Build a mansion for themselves. Everyone can have everything kind of thing because the price of goods and services is free
Then money stops meaning anything when it comes to survival or just like a good life
So this is like what's called a post-car city world
so for me
Then money is just mostly going to be used for resource allocation.
There's still going to be some resources that are going to be valuable and sought after by humans,
be that like land, be that real estate or some other whatever like fundamental physical reality
thingy that's going to be scarce for people. So then it's going to be used just for that. But
UBI is still important because it feels like something that, at least having the tools
for building such a thing in a way that is privacy preserving
and self-sovereign and is not built on the existing rails,
like the rails that governments control,
or the rails that are very easy to just manipulate
or fraudulate or just violate the rights of the person.
I feel like that's really important, right?
Like something where a person can always receive at any point in time globally 24-7, they have
access to this thing, it's privacy preserving, it's cyber resistant, it has access to this
just because of being a unique human is an important tool to have. But overall, I believe
that the identity part is a lot more important than the UBI part.
And if I may add something on top of what this says, yeah, it's very important to think
UBI not only as necessarily one monetary thing, it could be like arbitrary resources, exactly
as he says.
One of the things that I kind of, you know, heard as an idea and I really like is this idea that eventually you need somehow,
you need somehow a system that kind of resuffles
some kind of resources, because eventually if you have humans or humans with AGI
or I don't know what setup we're going to have
in the future, somehow you need to still have a way
to vote indirectly for what is it a good idea and useful, a useful new service
or a new useful resource in a society or not. And in the future, if you have something like UBI,
you can start thinking about the money that you get as a way to sustain some kind of services or
products in the future by giving this amount of money back into the
system for further development. So it's like, it's a very, very interesting kind of circular
economy that, you know, people kind of, you know, allocate the resources they get fundamentally
just by being humans to what they believe to be useful. So it's kind of more like a direct
democracy in when it comes into businesses and to society, which
I personally find it's a very, very interesting idea.
And from what I understand, you are of the view that AGI is absolutely coming and it
seems to be coming fast. Do you guys have any of of course, like you can't predict the future, but do we have any timeline with probabilities of how soon we're going to have AGI?
My personal view, yeah, go DC.
Yeah, you can go first.
My personal view is that I'm unreasonably optimistic in those things. I do believe that we're going to...
Sorry, let me start the other way around. The problem with exponential growth is that literally
the next step is if you look back, you see a linear increase and when you look in front of you,
you literally cannot see the next step. So it's like so up high in front of you that you cannot see where actually is going to land. While the previous
step is like, oh, yeah, I just came from there. So it's easy. So I believe this is literally
what is happening right now to some extent, of course, it's not day by day, but because
we also kind of got spoiled from all the cool things that are happening literally every
day.
But I think in the next five years, we're going to see like humongous differences on all of these things. Just think about, you know, how technology and all other aspects have been increased the last 10 years.
I'm 34, I'm a little bit older than this, even though I look younger.
And the thing is that we have seen enormous changes in our way of life and problems also, right?
I mean, all of these social networks and all of the plastics, I mean, there are hazards everywhere,
but it's an amazing lesson for society how we can compensate and how we can fix those things.
But the life we live today is vastly different from life 100 years ago, even 10 years ago.
And I think this trend is just going to continue on unless we hit like a really, really bad wall
or firewall for the whole civilization. So yeah, and I'm on the optimistic side. I think things are
just going to go great. We just need to put our heads down and continue working.
Yeah, I share the same view.
If I had to put on a time zone, anywhere between five, less than five to 10 years
in that area until you have general intelligence in most domains that humans care about.
And then we were having this conversation with Ben previously, whether R is the Transformers architecture going to be the
one that is going to get us to AGI or are you expecting something else?
DC? I'm not deep enough to be able to comment on this topic. I know just how
it's really hard to discuss these architectures. There's like a lot
of different ones. I'm also not really up to date with all the different research directions in the
AI world. So I don't think I'm able to predict that one. I can share some thoughts on this.
I'm not an AI expert myself either, but I've been doing a little bit of reading regarding those things.
I don't know if LLMs are going to be the answer to the future. I think the way we're going to progress and this aligns a little bit with some more, at least to me, more interesting ideas what AGI and ASI means, because we kind of think AGI in a very anthropic
way, in a very, in a very, in a very mirror to human intelligent way. And this is not
necessarily what super intelligent means necessarily the universe, you know, can have like a lot
of, you know, secrets that yet to be unlocked about what intelligence means. So the thing
that I find very interesting
is that there are a lot of different architectures out
there that it seems to be performing well
under specific circumstances.
So I don't know if intelligence itself is
going to stay in an LLM way, but I see more and more
this idea of having multiple kinds of architectures, multiple experts per
specific kind of task that somehow together they create emergent properties. And this is a trend
already happening, right? When it comes to very, very new AI architectures, you see these kind of
expert models that they have multiple experts inside them. You see, for example, that diffusion is working well on specific kind of tasks like, you know,
imaging and videos, while on the other side, the lamps are better, you know, on language.
So you have exactly, you know, this kind of and also the human brain works in some way in a similar
way. We live in a world that we need experts per domain, but somehow you
need to be able to connect those things. This is why I'm so excited about very recent, for
example, advancements in technology that we are able to interface LLMs or any kind of
other AIs with actual more concrete tools. So, you know, recent, this recent protocol from Anthropic
is called MCP, if I remember correctly.
I think it's a very, very interesting turn of events
because also if you think it from a point of view
of languages, programming languages,
is like, why did we move away from assembly
and see towards more constrained languages like Rust?
It's because, you know, humans, we do mistakes as we write code.
So you need to confine the infinite search space
for a human when we want to do our job.
So you should do actually exactly the same with AI.
You need to confine the options that it has.
And by doing this, you can actually increase the quality,
you increase the accuracy,
you even reduce hallucinations, which is like a very, very big problem. So by doing such kind
of advancements that they're like the environment of these AIs, you can actually maybe achieve a lot.
This is why I'm very, very bullish on the short term to actually see more such kind of innovations.
So existing AI that's not a perfect AI, but better integration
with our environments. I think this is going to add far more value than waiting for the big leap
moment that, I don't know, the next chain of thought, big advancement that somehow
sends the new AI to the moon or something.
to the moon or something. Ben, do you have any thoughts to add?
Not especially. So I'm not a pure AI guy either, but from speaking to those that I know who
are, there seems to be some hesitation around the idea that LLM is going to get us to where
we want to go. So I think I broadly agree with everything that Anastasios said.
What do you believe, Aghelli? Do you think LLMs will get us there?
I think, I mean, they could get us to a certain point,
but I'm just very, very curious to see what we can do with like
a probably different type of hardware with neuromorphic computers. I've seen some trials
happening I think in Australia amongst other places where researchers are trying to create
biological computers were kind of hybrid ones and I'm kind of very, very curious to see
hybrid ones, and I'm very, very curious to see
what we can get from that.
Honestly, I'm not an expert either,
so I can't add anything cool here.
Nice for reversing the question back to me, Anastasios.
Well played.
Cool.
So I mean, there are a few other things
that we'd like to talk about, but also open to you guys,
like DC, Anastasios, are there any interesting things
that you've been working on recently that excite you,
that you'd like to share?
Anything cool that you saw recently
or that you worked on that caught your attention?
I personally would like to highlight a lot of the work that actually the foundation,
that DC is doing and a lot of other things that we do also on the TF8 side with respect
to cryptography things.
Something that I mean for me gets me very, very passionate is like, yes, we are probably one step away from proper AGI.
That is going to change pretty much everything and is already changing everything.
But something that is still a thing that is not going to change is a lot of, or I want to believe there's not going to change,
is a lot of our mathematics and a lot of fundamental things that we have already
discovered as humanity that has to do with cryptography and other areas of mathematics
like that. So for example, right now, there is a lot of research that we do also internally with
respect to zero-knowledge proofs, what are the capabilities there, how to optimize them, make them faster, make them practical, because a lot of those things that came
as completely abstract theories of the past, computationally they were almost infeasible,
and there's a lot of research on how to enhance them to make them practical and provide actual
value to our everyday lives, and at the same time we also push the frontier forward with new ideas, new
capabilities that we didn't even know those things exist. And an amazing thing to me is that even if
we have like AGI, AGI would still play with the same rules that we play as humans. I mean NP
complete problems probably they will never go away. So you know the fundamentals of cryptography,
even you know the quantum resistance one is
never going to change. Again, based on the things that we know today. But I find it just a very
interesting and fascinating problem that AGI is not the only thing that is cool, because as humanity,
we are also discovering the boundaries that even AGI itself is going to be bounded in. It's the same case for all of us.
Yeah, 100%. Yeah, I don't know what to add to this. Is this true?
It's really, really cool stuff and a bit trippy as well. I've got some stats that I'd like to share.
We've been doing some research.
And what we're seeing is that the performance on GPQA diamond
benchmark shows that essentially, LLMs are surpassing
humans with PhD using Google in their field as well.
We also see that, well, I mean,
we saw what happened with DeepSeq
and now with Baidu's Ernie 4.5,
which matched or exceeded GPT 4.5's capabilities
at 99% kind of cheaper cost.
And so given that, we kind of cheaper cost. And so given that we kind of expect that humans
will all want to be using their own LLMs
and probably going to get their own specialized agents
that are, I'm sorry, specialized and personalized agents
to kind of act on their behalf
with varying degrees of autonomy.
And so given that, we expect that we're gonna see
probably more agents than humans very, very soon
kind of doing stuff.
And so the question is, is proof of AI autonomy, proof of AI agent, if you
will, more important than proof of humanity? And if so, what can we do? The question is,
how important is it going to be that you know that you got the inference from the AI agent
that you requested from.
And you can provably know that if we assume that AI agents are
going to be doing most of the stuff.
So yeah, do you agree with that or not?
And also, what could be potential solutions for this?
Can we get proof of AI agent? So I've spent a fair amount of time in
the intersection of cryptography and AI. There's been a lot of developments
over the last three years where it was just a theoretical concept and nowadays
there's some like practical demos of like how this technology can work. The
problem here is that right now, it's
not practical economically, because the economic overhead
of proving an LLM inference is already
like 1,000x more computationally expensive,
and thus mostly like 1,000x as expensive to compute.
So if I want to run an inference run of, I don't know,
like, Lama 3, 70 billion, it's cheap.
It's a few cents, probably less per inference
if you do, I don't know, like, a million tokens or 10 million
token output kind of thing.
But if you want to ZK proof of that,
then it's obviously a lot more expensive.
And there are some other limitations
where the prover will run out of memory
because it's just such a big program
that you need to put.
And then you have an overhead blow up
where you need a lot of RAM per each bit, essentially,
that you're proving about.
And so these things are going to become more
feasible over time, of course, as the state
of the art of computing becomes a lot better,
as these models become more, maybe smaller as well.
You have the same amount of intelligence
in a smaller model with less bits.
Maybe that's another reason why it can improve.
But it's definitely feasible to get proofs of inference.
The most cost effective right now
are those based on secure hardware.
So you have a trusted execution environment.
You put an LLM into it.
And then you're able to
have that attestation from the secure hardware. But of course, it's not secured by mathematics,
it's secured by hardware, which has its own assumptions. You have the entire production
assumptions of the trusted execution environment where anyone can attack any part of the assembly
or manufacturing part, or then you also can just attack the firmware.
There's people who use electro microscopes to read the electrons as they go through a chip,
and there's side channel attacks of all kinds. So of course, there's those kinds of assumptions,
but it's currently the most practical way of giving a sort of guarantee that something has
been run somewhere in an economic sense that actually has some product impact at the moment because nobody is willing to pay for a zk proof of a model
and you still cannot zk proof the models that are state of the art like if i have i just bought a
new macbook i have an m4 max and even for me it's really hard to run lama 70 billion on my own
computer imagine if i had to do a zk proof of that. I would have to have a supercomputer just to do a ZK proof of a single inference run.
And so it's currently not economically nor practically feasible to do that.
But yeah, that's currently the state of the art.
Anastasios, any thoughts?
This is covering me completely.
Yeah, of course. Makes a lot of sense.
And then, what are your views on kind of open versus kind of closed source models?
Where would you bet? Like, are you kind of pro open source or closed source?
would you bet? Like are you kind of pro open source or closed source?
So you want to take it, Anastasius?
No, no, no, go for it.
Yeah. I've invested a bunch in startups that are doing decentralized AI and open
source AI. One of the ones that I really like is Prime Intellect.
They're essentially doing like distributed training over lots of compute
across the entire world, so over the network.
The thing is, the economics here are really tricky
to make them open source for a lot of reasons.
And the ones who do, they usually do it at a loss
or try and have subsidy by other business models.
If you look at Meta or Google, Meta, for example,
the reason why they can afford to open source a lot of the models
is because they have the business revenue of just
the other stuff, like all the Facebook ad revenue
and other things.
So they're able to do that.
And they're trying to establish a standard where everyone
builds tooling around their tools,
because everything that they have is open source
They even open source their schematics for how to build a data center
So a lot of data centers nowadays are built after metas data center standards
It's a question right like if you want to generate something that has an economic mode where you're able to sell something at a premium
then you need to capture value somewhere.
And it's not at the point where a lot of people
are able to capture that value in a sustainable way.
Maybe you can argue, oh, maybe DeepSeek
is able to do these things because they're
able to do it at 2% the cost of major LLMs that are state
of the art nowadays.
But then that question is kind of tricky.
Then you have state attackers or state actors
being able to subsidize these activities.
They don't necessarily need to care about revenue there.
It's a lot of really interesting dynamics.
And if you go through the depths of it,
it's really, really competitive environments
where even state actors are really interested in it.
You see even things like Taiwan
and other things like becoming really geopolitically
interesting or like just geopolitically sort of important
to the world just because they make chips
and for other reasons, right?
Like this is a really hard conversation to go into
in a short amount of time, but overall,
like the centralized AI is important
because you need to give access to everyone
to run these things locally.
You need distributed access to AI,
and you should not have a few parties that
gatekeep the access to it.
But at the same time, I understand
why some don't open source and they
want to create products that actually allow users to use
them through some API.
There's going to be a bit of both.
And I think long term, open source is going to win,
mostly because that's the trend we've seen so far
with operating systems.
Linux is the one that won when it comes to at least servers,
not consumer.
That was won, unfortunately, by Windows and Mac OS.
And for phones, Android also is one.
It's also open source.
It's based off of Linux.
So I hope that this trend continues, right?
Like you see a lot of these companies really push hard
for it, especially like Meta,
which open source Laman is trying to push that a lot.
And it's becoming like one of the default models
in a lot of the worlds for a lot of the applications
that people want to run locally.
So yeah, it's just, it's hard to tell what's gonna win,
but I hope the open source ones are going to win. Something I just want to add on top of what this
says is like, for me, it's very, very also important to notice that sometimes the best ideas
with respect to quality or functionality might not win because of economics. I want to highlight
this aspect that DC also said many times,
that the economics can be very, very tricky.
We have seen a lot of times in history,
like second-grade solutions might eventually win out
just because they're cheaper.
So if eventually open source manage to provide some kind of chip
and economically make sense,
it will prevail.
But if we don't manage to solve exactly these problems,
we might, I don't know,
end up in only centralized solutions.
Yeah, yeah.
I 100% agree on the kind of open source approach.
And I think that what we've seen historically
is that what matters the most is the community
and how big an open source community there
is for a project that usually wins.
Like we've seen that with programming languages
that were arguably inferior when they started,
like JavaScript, for example.
But then because there was so many people using it
and a very big community, they kind of improved it a lot with just in time compilations
and all these kind of improvements.
And yeah, so it's not always like the best tech
that wins in the beginning.
It's the strongest community that wins in the long run
from what I've seen historically.
So yeah.
Ben, do you have any thoughts on the mini-convo we had previously on attempting
to put LMs on chain so that you know that they're immutable and essentially you can
be certain?
Do you think in what sense?
Do I think it's a good idea or do I think it's possible?
Yeah, like would it solve the problem of having verifiability of an LLM's response, inference response?
So then, yeah.
Sure, you could argue that it kind of solves that problem on the basis that it's one, publicly verifiable, and
secondly, everything is immutable. But actually pulling that off is going to be something quite
special at the moment. LLMs are big and blockchains aren't great at doing any kind of big computation.
So I don't think that will happen anytime soon, but I'd like to say yeah. But do I think I'm going to say it?
Probably not, not in the near future anyway.
Yeah, I mean, it would not be really smart also
because you like the basis of blockchain
is the way that you get crypto economic security
over the output of some validator or something
is that the entire validator set
is gonna re-execute the same thing.
So if the validator is going to be like, hey,
here's the computation I'm doing and it's an LLM inference,
then you have to do that same LLM inference everywhere.
And you have to fix some randomness,
because of course, models are not deterministic.
So you have to fix some randomness.
So that constrains the output and the solution.
So it's just not practical, because you're already
doing one inference is expensive.
You have to do 10,000 of them at the same time for just one
output is also not impractical.
I don't think that alarms are ever going to be on chain.
It just doesn't make any sense in that sense.
It makes sense for agents to interact with blockchains.
That's, of course, true, as is just another API for humans or for AIs to just
have access to. But the way that you get verifiability for me would be cryptographic,
either through a TE or ZK proof or maybe some MPC setup where you're able to collaboratively
prove that something is actually running. But yeah, definitely not on-chain. The overhead of
that is orders of magnitude even bigger than cryptography
And it's economically just completely feasible in my view
Hmm going back into a point that you mentioned earlier
And I wanted to kind of go back to it is the the biometrics that that that you mentioned. So you are focusing on the iris
of the eye, right? So the question is like how kind of unique is it in humans and what's
the rate of potential hash collisions? Yeah, so actually we have a lot of research on this and currently, I mean, we are on the
second iteration.
We are on V2 as we call it on our Iris code.
Maybe just to give some context for everyone here, what we actually do is we are trying to get as much entropy as possible from the surface
of the iris.
First of all, when we take photos of the iris, it's not always the whole iris.
There are parts of your eye that are occluded because of your eyelids or because you have maybe, you know, have narrow eyes
or you might have like, I don't know,
your eyelids that might be in front of it.
So it creates like a lot of issues.
So you're looking, you know, for surfaces
that you can clearly, you know, see and get
and extract information from there.
And then what you do is you pass some filters
on top of this image. I kind of stuck right now, don't remember the name of the filter, which is...
Gabor wavelets.
Yes, sorry.
Yeah, the Gabor wavelets.
And we are currently on the second iteration on how, these filters, they kind of get us the information
we care about. To the point we are today, as far as I know, we can scale up to 10 billion
users. Of course, this is like very, very ambitious. So this is why now we are deploying
actually a third version of these filters. and it's a very, very interesting
task because the way, and I can explain a little bit more how we do it, but the way we are going
to do it is by doubling down on zero-know-less proofs to prove that the computation we do is
correct. And I'll come back to this in a bit. But something I just want to highlight is, again, even though with version three,
again, we expect, you know, the collision site to be, again, quite large. What we understand as,
as, you know, more and more research is happening right now that there are a lot of secondary
factors that they kind of create a lot of noise out of your,
on this kind of encoding that we do.
And this noise is actually, again,
going back to all of this AI discussion we had,
is a discussion like, okay,
can we do something useful with less intelligent AIs,
but having like better integration
with our existing tooling
or should we double down to the best possible AI
that is going to come a little bit further on?
And the reason I'm saying this is that
the algorithm itself doesn't change much,
but we manage to clear out the information there
more and more, which is much, much more important
apparently to the things that we do today.
And one of the amazing things that
they're just coming with this new upgrade that I personally feel very passionate about is that
currently our users, they stand in front of the orb and what do they do is they just capture
their biometrics. And then end-to-end encrypted, we send all
this information to your phone and you have full custody of all of this information yourself.
And then what do you do is you submit on our SMPC cluster, which is like a distributed
database, but none of these, it's a multi-party computation database and none of these charts
they have like the full information is only partially existing there.
You send, you know, parts of this encoding
that you created based of your biometrics.
Now, all of these, we have around 11 million users
already created, you know, this Iris codes.
So the question is, how do you make 11 million people
upgraded to the new version of this encoding
without bringing everyone back to the, to the orb to recreate
this kind of Iris codes. And that is, that was like a magical
moment for me, because, first of all, you cannot trust the
phone, right? I mean, you have, you have full custody of your,
of your, of your information. And that is a trustworthy
situation for you, but not for us that we would like to upgrade our algorithm.
Because if we ask you to re-evaluate your information and send us a new code,
how do we know you didn't cheat on your computation? How do we know you didn't use
a different setup or different information on the algorithm that we want you to run.
The way we do this is actually with zero-knowledge proof.
You need to run our algorithm,
the new algorithm, the version 3.1,
and at the same time submit a proof that you run
the algorithm that we send you correctly.
That was like a magical moment for me because now we have
a system that we can upgrade information around without actually asking people to re-scan themselves in the orbs.
So the question is like, is the phone equipped enough to do the same job that the orb does
in this case?
I can talk about this.
So essentially, the reason why, so as Anastasio's mentioned,
you have this iris code.
And the iris code is currently generated in the Orb.
And then it's stored locally.
The Gaber wavelets are not computationally expensive things
to run.
The problem here comes like running it inside
of the like trusted environment, right?
Like how do I, the reason why the orb was important before
is because it's a place,
it's a computational environment we can trust
because it's a trusted execution environment.
There's a lot of sort of anti-tamper prevention,
security, OS operating system that like helps you a lot
with making sure that whatever is happening on
the Orp is trustworthy. And you have an embedded graphics card that is able to do all the computational
intensive parts. A lot of the computationally intensive parts especially don't come from
creating this Iris code by applying these Gabriel filters on top of the Iris image,
but by actually doing the liveness verification. So verifying there's a real person and doing a lot of checks
where, OK, this person is not wearing contact lenses, glasses.
They have a 3D face.
They have a heat map, et cetera.
Those are the more computationally expensive ones.
And through modern advancements in the application
of zero-knowledge cryptography to machine learning
and creating zero-knowledge proofs of inference,
this is what I was talking earlier. It's really infeasible to do Z knowledge proofs of inference. This is like what I was talking earlier.
It's like really infeasible to do ZK proofs of LLMs, but it's already ZK
feasible to do proofs of things like Gabor wavelets, which are just like
convolutional filters that are applied sequentially over and over.
It's like 27 or 28 of them.
I think it's 28 of them.
Yeah.
They apply and it's not, it's not, it's not too expensive.
And so, um, last year we Tools for Humanity acquired a team called Modulus Labs,
which is now the Tools for Humanity Applied Research Department team.
They are experts at zero knowledge machine learning.
Their startup previously was just purely focused on ZKML,
and we started working with them on the basis that they would allow us to do
this specific thing, which essentially,
for those who are technical, it's
using their custom proving system called Remainder.
And it's using a SNARK protocol called GKR that
allows you to essentially create efficient proofs
of linear operations, such as the Gabor filters.
The Gabor filters don't have a lot of non-linearities.
So GKR is really well made for that.
And then there's some specific techniques that allow you lot of non-linearities. So GKR is really well made for that. And then there's some specific techniques that
allow you to handle non-linearities,
specifically logup.
But what we're talking about here
is that you're able to create a proof now
with a really performant prover client side.
All that we're about right now, we
care about client side proving and improving
the state of the art of client side proving.
We're able to create a zero-knowledge proof
in the user's device, and in a fully information theoretic
secure way.
I'm able to create a proof that my iris code was generated
correctly in my phone in an efficient time,
so that the user experience is fine.
It was much better to just wait with my phone
idle for two minutes, rather than having to commute
to somewhere where there's an ORP
and re-verify.
And the good thing is that now the ORP will never
have to actually have the iris code.
It would just be a fancy camera that will attest to the images.
The user will store that inside of their personal data
custody package.
And then they're able to then interact with the MPC database
by just doing all of the like splitting of the data
into these like linear secret shares essentially like encrypted shares and sending those to the
MPC setup so they're verifying that they're unique with their new iris code that uses the new kinds
of of of biometric algorithms and so so yeah it's literally like a soup of a lot of technologies and all of them just perfectly work together and give you like full privacy, which is insane.
Like this is much better than any other identity system we've had before and you have much stronger guarantees about like if you're unique or not. So yeah, it's actually insane that all of these things work together.
When I joined World, seeing the interplay of like, oh, now we have AI biometrics, custom
hardware, you have this and the other custom lenses, a lot of optic work, there's a lot
of work on so many different topics, there's cryptography, blockchains.
It's kind of insane that all of these technologies just became feasible over the last five years
at the same time for these things to happen. So yeah, this is like sort of like more introspection into
that specific problem. Yeah.
And something I would like to add on top of what this says is like, I feel also very proud
that we're pushing the boundaries forward. We're literally inventing new levels of technology
here to do exactly this.
Make sure our protocol and the way that we are trying to do identity is the most private
in the world and at the same time the most feasible one.
Because also if you want to do like the most private of things, it can be completely unusable
because yeah, the most private thing is that you share it with nobody.
But the question is, can you share, can you make like the maximum impact to you and the
services you would like to access
without giving any information away? And this is what we are actually achieving here.
So are users going to be able to kind of select what pieces of information to give or maybe like
a range of the information so instead of giving the actual age, they can say, I'm like above 18, so I can drink, et cetera.
That's possible, right?
Yeah.
So we released this product called World ID Credentials.
World ID Credentials allows you to sort of scan,
let's say, like your digital passport.
The NFC chip has all the data about you,
and you're able to scan it locally,
and you have it inside of your phone.
And what we allow you to do currently
is just to prove that you own such a document.
But you're not able to prove specific parts of it.
And in part, the reason why that doesn't work is because,
or it's not worked historically, is
because if you look at the distribution of devices
that our users have, on the tail end of them,
they're low-performing devices, like a 32-bit Android
phone with 2 gigs of RAM, no bandwidth.
And so if you work in such a constrained environment,
it's really hard to create these zero-knowledge proofs
of selective disclosure.
And so right now, we're working on a new proving system
that will allow you to do that exactly.
So over the coming months, you'll be able to do selective disclosure of any given field
within that password in a fully privacy preserving way.
So, I can prove to you that, hey, instead of me saying that I'm Czech, I can prove to
you that I'm in the set of European countries by doing a ZK proof of that.
Or I am, for example, over 18. Or my gender is male, right? Like these sorts of things are
things that are going to be like really feasible at scale for everyone that has a world app, essentially.
But yeah, like it's not going to be just about that. Of course, like the KYC identity is like
important because you can do now selective disclosure of any part of that. But the way
that I see it in the future is you're going to be able to do predicates or statements
about any part of your identity.
It can be your financial identity.
It can be your personal identity.
It can be your internet identity.
I can ZK proof that my bank balance is over X.
I can ZK proof that I've used this app on, say,
I can ZK proof that I've interacted with uber.com.
I've done more than 100 trips.
And then I can claim a discount, right?
And all of that while it being civil resistant.
And you have this just general environment
that allows you to create a proof about anything
as long as you have a root of trust.
The root of trust in our case is the org.
The root of trust in the KYC case
is whatever institution that actually issues the digital signatures on the passport.
So these root of trusts have different security assumptions,
of course.
Some of them are more robust, some of them are less.
But it's still important that it allows users to finally
unshackle their identity from the cages of SQL databases
in some corporate database.
But it gives the right to the user
to do predicates and statements about their identity
without anyone blocking them and without the third party having
to get a lot of information about them
before they can trust that information.
They can just consume the proof and the data and the statement,
and they have high guarantees that that is the case.
That information is true.
Awesome. Awesome, awesome guys.
I mean, we've got a couple of minutes, but I also understand that you've got to hop
on other calls and do a lot of stuff.
So we can wrap here unless there's anything else.
Thank you so much both for hopping on.
Thank you so much both for hopping on.
It's been an absolute pleasure.
It's been an absolute pleasure.
And yeah, looking forward for a potential next one in the future.
Thank you so much, Vageli.
Awesome, guys. Enjoy.
You too. Take care. Bye-bye.
