Podcast
Join Viktor, a proud nerd and seasoned entrepreneur, whose academic journey at Santa Clara University in Silicon Valley sparked a career marked by innovation and foresight. From his college days, Viktor embarked on an entrepreneurial path, beginning with YippieMove, a groundbreaking email migration service, and continuing with a series of bootstrapped ventures.
Links
Follow Me
Follow Me
Join Viktor, a proud nerd and seasoned entrepreneur, whose academic journey at Santa Clara University in Silicon Valley sparked a career marked by innovation and foresight. From his college days, Viktor embarked on an entrepreneurial path, beginning with YippieMove, a groundbreaking email migration service, and continuing with a series of bootstrapped ventures.
The Future of Personal AI and Privacy: A Deep Dive with Kin Co-Founder Simon Westh Henriksen
Play On
07 APR • 2024
59 mins
Share:
In this episode, I’m joined by Simon Westh Henriksen, co-founder of Kin (formerly Hyphen), to explore the fascinating intersection of AI and privacy. Simon’s journey from software engineer to Web3 innovator offers unique insights into how we can build AI systems that respect user privacy while remaining highly functional.
We start with Simon’s path to founding Kin, which took an interesting turn during the COVID-19 pandemic when he dove deep into Web3. What particularly caught my attention was his vision for a privacy-first personal AI assistant. Simon’s explanation of how they handle data locally on devices while maintaining AI functionality reveals the practical challenges of balancing convenience with security.
The conversation gets especially interesting when we explore the future of AI interaction. Simon shares his thoughts on AI agents communicating with each other and the potential of decentralized web nodes for data portability. His insights into managing AI hallucinations and the ethical considerations in AI development highlight the complexities of building responsible AI systems.
I was particularly intrigued by our discussion of data sovereignty in an AI-centric world. Simon’s approach to giving users control over their digital data while still enabling powerful AI capabilities shows how privacy and functionality don’t have to be mutually exclusive. We also explore Kin’s beta program and their vision for the future of personal AI assistants.
If you’re interested in AI, privacy, Web3, or the future of personal computing, you’ll find plenty of practical insights here. Simon brings both technical depth and entrepreneurial experience to the discussion, making complex concepts around AI and privacy accessible while maintaining their technical sophistication.
Transcript
Show/Hide Transcript
Hello and welcome to this episode of Nerding up with Victor, where we delve into the intricate world of technology and its implication on our lives.
I'm Victor, your host, and today we have a particular insightful discussion planned for you.
Last year, my interest was piqued by the evolving landscape of Web3, particularly around the concept of data sovereignty.
The idea that individuals could have ownership and control of their digital data is not just transformative, but increasingly critical in an AI centric world.
Our guest today, Simon from the company Kin, is at the heart of this.
Kin is a company at the forefront of privacy in the world of artificial intelligence and machine learning.
Whether you're deeply involved in the tech industry, interested in the ethics of AI, or simply curious about the future of digital privacy, today's episode promises to be a compelling exploration of these critical issues.
Stay with us as we engage in a discussion with Simon, unpacking the challenges and opportunities that lies at the intersection of AI, machine learning and privacy.
Welcome, Simon.
Yeah, thanks a lot, Victor, and thanks for having me on your podcast.
We're excited to talk a little bit about the personal AI and privacy and Kin and everything that's, you know, around that about myself, I'm your, like, classical software engineering background from like IT University in Copenhagen and always been doing different tech startups, trying to bootstrap my own companies.
A handful, definitely.
Most of them failed.
Right.
A few that are still alive.
But then I realized, I think during my bachelor on ITU actually that I have been coding since I was, I don't know, sev 7, 8, 9, 10, doing websites in Microsoft front page and all the stuff.
And, you know, I thought this was, you know, the perfect thing for me in the first year at ITU was great, you know, I, you know, I was ahead of most people.
There were of course, others like me for the first time.
Right.
But then during the second year where we started to have like really complicated, like mathematics and discrete math, really, you know, the deep computer science stuff, I saw that, okay, maybe it's not this, so this.
Maybe it's not the like nitty gritty details of the science that gets me turned on.
Maybe it's building something with software.
And I think that also came together with seeing people that started, you know, together with me completely blank, no experience.
Right.
And then by the end of those three years, actually overtaking me in there, I would say like programming skills.
Right.
So during those years, I started at Danish travel company with my cousin actually, who's been in the business for a long time about Basically Dynamic Packaging, it's called.
So you put together like any hotel that you get from APIs, flights that you get, you know, the same flights from Mondo Packaging, it, selling it online.
It was kind of new back then.
We kind of started to get that off the ground and I worked on that in the year.
I took a year between my bachelor's and my master's and worked on that and got that up and running.
And I decided to take a master's in a business IT mix from Copenhagen Business School, which was way less about actually programming and much more about how do you manage software projects, how do you manage like creative people?
Which was super interesting.
And I learned some new skills.
And then after that I was like, what am I gonna do now?
In that time, the Danish travel company had actually merged with a Swedish company and split up again.
And I had another company that was doing festival management software that was, you know, I could also go in and do full time in.
But I decided to go all in on the Swedish travel company and build a team, got it off the ground, went really well, grew to quite a big company.
You know, it's pretty big numbers in the travel industry because you need a lot of revenue and small profit margins.
And it was working really good until Covid hit and were like talking to investment bankers and you know, about to finally cash in like a 10 years journey, right?
And then suddenly from one day to the next it was like, should we just, you know, go bankrupt and call it a day?
Right, because we had to pay back hundreds of millions of krona, Swedish krona to customers, money that we already paid to, you know, our suppliers, right?
Because were in the end just a middleman.
So that was kind of tough.
And I think if we knew that Covid was going to last like two years at least in this industry, we probably not would have fought through.
But you know, luckily were optimistic and we did and all is well and it's alive now.
But that just made me realize that it's time to try something new.
And that was then about the time where the whole web3 space with NFTs and stuff were taking off.
And I've always been interested in blockchain technology more from the technical side than from the investment side of it.
But I've always found that it was very much like a closed society in a way.
Or, you know, it was always, you know, crypto products for crypto people defi, you know, NFTs and stuff like that.
It wasn't really Something that my sister would be interested in.
Right, right.
But the technology is so amazing and it enables so many things that we need.
So when I saw that Casper, my co founder and CEO in kin, started to wanted to basically build a company together with MOF Capital about Web3, they wasn't exactly sure, but it was like, let's take this new technology and see if we can build something like a real consumer product.
And that was two years ago, basically.
And were on this journey to figure out started with the technology, basically.
Right.
And then trying to find a problem that we could solve with it, which is not what you're supposed to do when you're.
The other way around.
Yeah.
But okay, so it was very research ish and we did a lot of things and we just, were really excited about the data ownership part of it.
Yeah.
Because what we saw back then and also see is that the Internet is really centered around these giants that hold your data and basically do whatever they want with it.
Right.
And it's like that's the exchange that we're used to.
You get this product for free, but the reason is free.
We all know if you're not paying.
You are the product.
Right, exactly.
So how can we turn that around?
And from our point of view, that required that you were actually able to hold your own data and decide who gets accessed and have the incentives of sharing your data yourself so that you would get that benefit of sharing your data instead of someone else getting the benefit of sharing the data.
And we're exploring many different ways.
A wallet that has a private key, like a blockchain crypto wallet, but could then also store data in different formats.
So about this, around that time, the World Wide Web Consortium was coming out with new standards around decentralized identifiers and verifiable credentials, which was kind of big because it was the first time I think in like 10 years that the consortium was actually bringing standards from a draft into a real finalized version.
So it was like, okay, maybe there is something here that's actually going to be part of the Internet going forward.
Right.
It's not just an idea anymore.
So we built identity, what's called an identity wallet, where you could store verifiable credentials and stuff like that.
But it turned out that first of all, no one knows what verifiable credentials are.
And to get people to understand and businesses to understand that first of all, we need someone that can issue the verifiable credentials.
And why should a company issue verifiable credential what's the benefit for them then?
We need the users that need to hold them and present them to some verifiers and the whole ecosystem.
Sorry to interrupt.
Do you want to take a few seconds and explain what's VCS and just explain the core concept just for people not familiar with what that is?
Yeah, So a verifiable credential is.
You can kind of think of it as like a credential in the real world, or think of it as like a card that's in your wallet that has some data on it, like your driver's license, for example, tell something about you.
It could be your name and your birthday and stuff like that.
But it can also be, you know, some access levels or something.
It could be like an employee card, but it could also be your loyalty card from the coffee shop or something like that.
And then the idea is that you can have this wallet with these different credentials in it and you can then go and present it to someone.
So, for example, I could have my employer credential and I could go and present it in the taxi because I know that the taxi company has some kind of discount thing with my company.
And the special thing here is that the taxi company would be able to verify that my company actually issued it without going to my company through some kind of API or something because of the cryptographic signature that is, you know, embedded into the verifiable credential.
So now we're back to, you know, private keys, but that my company that issued the employer card would have a private key and they would do a signature embedded into my verifiable credential so that I can, when I show my verifiable credential, the verifiers can just check that signature and see if that is a valid signature, that is a valid employer credential, that I have no other checks needed.
And I believe there's also a concept of being able to prove who you are without disclosing the actual information of that payload as well.
Yes.
Yeah.
So that's kind of the next step.
And what is also supported by at least let's say modern verifiable credentials is zero knowledge proofs.
Right.
So that is being able to prove something about you without actually showing the underlying data.
It's like kyc, for instance.
Well, that's maybe a bad example, but some.
Yeah, the classic example is like you're walking up, you know, to a bar and they ask to see your ID because they want us to know that you're above 18.
But instead of showing them your actual birth date, you can generate a proof that shows that you are above 18.
That's all they want to know.
And they can then verify that proof without getting the actual underlying data.
That's cool.
Yes, very cool.
So that takes you into a hyphen.
Yes, yes.
Long journey.
So we're trying to figure out can we use this kind of thing.
For example, take your Strava data and then generate some credentials about how good you are at running.
Like you can run a 5k in this than whatever.
And then you could take that and you can go to the shoe store and you could prove that you're a good runner.
And because you're a good runner, you might be a potential really good future customer.
So they might want to offer you some discount.
But like this almost like three sided marketplace that were trying to establish very hard when it's, you know, it's both like new business, let's say model concept that are very different from what they're used to.
This new technology, it just didn't take off at all.
But we learned a lot about the technology and what you can do with cryptography and all this kind of stuff.
And then ChatGPT came up and it took us a few months of basically using ChatGPT every day like everyone else to realize that what we learned over in what we can call the self summoned identity space that encapsulates all of the verifiable credentials, decentralized identifiers and all this stuff solves so many of the problems that were seeing in the AI world around privacy, around verifiability and just like trust in general and the LLM or the AI, it gave us basically the opportunity to build a product directly with the user that they would find super interesting because now we could build an AI that they could engage with directly with their data.
And that was kind of for us like the perfect combination.
We did a demo and everyone was just, that was a demo uso.
Everyone was really excited, investors, everything.
And then it's just like that was April last year and then it just took off.
We got an investment in August of $1.1 million and hired some amazing people to our team.
And just yeah, it has been like a race since then to get something out and now we have the, since December we have like a closed beta.
I think we have around 250 users.
You are one of them in the beta that are testing and giving us feedback.
And it's been an amazing journey and it's Just incredible to see how people are engaging and how excited people are about this personal AI future.
Absolutely.
So we chatted sometime in the fall about this and one thing that I got super excited about this was this concept of an AI assistant.
Right.
Where well, I should say privacy first AI assistant that back then we spoke by having something running locally that.
I mean my Pipe team is essentially having a system that runs locally on my hardware that has access to my contacts, my calendar, my messages on all platforms, my email, my bank account statements, credit card statements, but like unlimited access essentially because it never leaves my whole data set and then be able to query that just like I would do chatgpt.
That that was like I.
That vision I think is still the long term vision for.
For.
For kin.
Right.
I think that is super fascinating because once you have that data set and that database or ll call it, but that enables so many interesting use cases.
Do you want to speak a bit more that how what, how you're thinking on the future?
Looks like that, Yeah.
I would say, you know, when we speak to people about personal AI, you can speak to 10 people and they have 10 different ideas and mission of you know, how this should be.
And everyone immediately jumps to this crazy future that you also just described.
Right.
Where you just have this AGI like thing that can interact with all your data and other people's agents and you can plan your hours.
Yet Holiday can ask when are Victor available?
And go and find hotels and flights and book everything.
It's gonna happen 100%.
It's just not where it starts.
Yeah, of course.
So a big challenge has been to kind of stop talking about that future vision because we really want to get there and we believe in it, but take it down to some use cases and something that we can build with today that people can get value of today.
So that's what we have been trying to do.
Yeah.
And I presume having things that runs locally on your hardware requires pretty beefy hardware to start with.
So you would cut out like 99.9% of the user base.
So I presume from a pragmatic standpoint you went down the mobile first route instead because that allows you attract a much bigger user base, obviously.
And.
Yeah.
Yes.
So.
So maybe I could just explain what we're trying to build now and then why we took that specific local first approach.
So kind of what we're trying to build now, the product is this combination of imagine a life coach that you have confidentiality with.
He's got some mental frameworks, you trust that he is knowledgeable in different areas can help you.
And also maybe because you went there a few times, he kind of understands how you're thinking and stuff like that.
Then combine it with a close friend that is always there to support you, no matter how stupid you are, knows you really well because 10 years of relationship or whatever and it's easy for you guys to relate to each other.
And then combined with this new technology that are the generative AI models that are incredibly knowledgeable, they are non judgmental and they are just always available and there when you need it.
We're trying to combine this into a product that can help you in the everyday.
Because I think that's a huge difference between ChatGPT, for instance, where you have ephemeral chat essential, you have threads, I guess they have context, but you have no shared context between them unless you do a custom knowledge base in ChatGPT.
Right.
Which I've used it quite heavily as well, which I think is great.
But it doesn't have the shared context that you would have in something like kin.
Right.
Yeah.
So if you want to build personal AI as we see it, there's kind of two pillars for us.
There's the privacy part and we can talk, you know, we'll talk much more about that.
Yes.
And there's the personal side of it.
How do we make it personal?
And a big part of the personal is what people are now calling long term memory.
If you think about ChatGPT, it has memory within that single thread that you're in, Right.
You start a new thread and then it's gone.
We call it short term memory, if you want to put some fancy words to it.
How do we build long term memory where it actually gets to know you more and more as you go?
And that's a really big part of it.
And if we're talking about, you know, the local first approach, then why do we do this?
We, we're going back then to looking at the whole privacy angle.
So we believe that first of all, in future personal AIs are going to be the most valuable digital asset that everyone will own.
And because of that, it must transcend ecosystem.
It should not be locked into walled gardens.
That's really like our manifest in Kin, because like Elon Musk said a few weeks ago on an interview, he said, don't take away, my friend.
Like imagine that you have built up this, let's call it relationship because that's basically what it is over many years and it knows you really well, you use it every day.
Maybe your life is 10% better because you have this helper that can help you and guide you.
And then because you preach some terms that you didn't even read, Google or whoever is providing the air just takes it away from you.
Right, right.
And then suddenly you are, you know, your life is worse.
Right?
Yeah.
And all that investment is just gone.
And we have seen it, we know it happens.
Right.
It happens with Twitter accounts.
It happens all kind of things for Gmail.
As far as a good example, accounts.
Yeah, yeah, it's.
It's so scary.
And they don't give a right.
Yeah.
So we need an alternative to that.
And that's why we need privacy and we need data ownership.
And the privacy is really.
Why do we need privacy?
Do people even care about privacy?
I think that's a big whole discussion and I think privacy in itself is not a product.
But we also believe, and we can see that with our users, that it becomes very tangible.
When you're sitting there with your AI and you can see everything it knows about you get a little bit freaked out thinking about someone else might access that when it's Facebook and Meta and whatever.
In a way it's a little bit more abstract.
You're like, I don't really care and I'm just one out of a million.
And now they're actually really going to use it for anything because you don't see all the data that they have on you, they don't surface it to you in a view.
But we do that.
And we can see, we have asked some of our users, if something went wrong, can we do an export of your data and debug?
And then they're like, no, it's so private.
And I think that's a good point.
Because the reason why, I guess I haven't really been too concerned with the privacy element of ChatGPT is because there is no long term memory.
So the way you put it, right, it's like, it's just, yeah, sure, there's context about this particular thread, but it's not that sensitive.
But if I were to start chucking a lot of personal data into an aggregate in a long term memory, I'll be a lot more wary about where that data is processed, how it's processed and so on.
And even where are the data centers sitting, what jurisdiction are the data centers in?
All those things become really important.
Let's talk a bit about that.
Let's talk about data sovereignty and data in kin.
So it's a Local first approach.
So you have an iOS and android app.
So I presume that's using iCloud storage, presumably on iOS, I would imagine for the local data set.
Actually not.
It's completely local.
So let's see, like where to begin?
I think it's important maybe to understand if you want to build anything that handles private data and do computations on private data, that is challenging, right?
We have, you know, if you're working with private or encrypted data, you know, you have to encrypt it at rest.
That is a problem we know you know how to solve.
You have to encrypt it in transit.
It's also, you know, fairly easy.
But when it comes to computation, that is really, you know, the holy grail.
And there's different approaches that you can go to.
There's like the actually the holy grail, fully homomorphic encryption which is actually doing computation without decrypting the data.
So doing computations on encrypt data, which.
Is insanely.
Cool and amazing and there's actually a lot of work on it in the past few years.
So it's really moving even on machine learning models, but it's still like a thousand times slower than without.
So it's not really feasible.
If you want to build, you know, a mainstream application, right.
It's very small use cases still and then you can do, keep everything local, like don't send it anywhere, right.
100% local.
That's also, you know, that's a great option if you can do that.
It's simple and you know, it's a, it's a really good option.
Then you have other options like zero knowledge proofs that we just talked about, which is more for these kind of one off proofs.
So it's not really a feasible way if it's live data that is updating all the time because it's too heavy to generate all these circuits and prove them and so on.
Another way is confidential computing.
And here we are talking about trusted execution environments, trusted platform modules, whatever the names are.
And to briefly explain what it is, it's like an isolated part of different chipsets.
It's been around for, on all modern CPUs for the past 10 years.
They come with electrostatic exclusion and it's like an isolated environment within the CPU that the host cannot access.
So usually if I build an application and I put it on a machine or a VM or something like that, it loads some data into memory and the Host environment can do a memory dump of the process and see the data.
Right.
Because it's staring to your text.
What you get with this trusted excuse environment is that you get this isolated environment where the host cannot actually do a memory dump and you also get away.
So they are born with a private key and public key, of course, from the manufacturer, like from Insul, for example.
So it's like put inside the chip, Amazon, whatever.
No one knows it further up the manufacturing or supply chain.
Right.
And you only know the public key.
So what you can do is you can actually encrypt data with the public key and you can send it into the trusted execution environment.
And you know, the only place that you know data can be decrypted is if it reaches inside of that environment.
Right.
And, and bringing this back into the mobile world.
And Apple's been doing quite good leaps in terms of these things on later generations iPhones, both on the cryptographic side of things, but also on the computational side of things.
What's the current state of that right now?
How much of these models can you run locally?
I believe you can run even one of the big LLMs, somebody showed they could actually run that on a modern iPhone today.
How far away are you from able to do all this locally?
I would both say quite far, but also quite close.
It kind of depends on how you think about it or how or what you need.
So we, our goal is to run as much locally as we want because it's safest, it's the fastest we have, it's cheaper, people already paid for the computation power.
Right.
They don't have to pay us as well or some cloud.
Yeah, but.
And the devices are, you know, amazing.
The chips that are in iPhones and modern Androids, you can do so much amazing stuff.
But the very big models, you still can't run there.
So we kind of had to decide, okay, well, if we want to leverage the biggest and best models, which we need because we want to make a good product, then we can't only be on your phone.
We need to have a hybrid solution that can run as much of the phone as possible, but also offload some stuff to the cloud when necessary.
But an example of what we actually can do on the device is, well, first of all, if you want to run machine learning on the device, you also need the data onto device.
Right.
And that's a whole challenge in itself because you need a data structure or database that is highly performant.
For us, we are building like a graph database.
So we wanted to support graph operations like think Neo4J or something like that.
Right.
Then we also need typical like transactional data SQL, like we need support for vector embeddings and you know, similarity search algorithms.
And then we want to run it on a phone.
Right.
And we want to be able to sync it across devices as well at some point.
So the options that you are left with those requirements are basically none.
So even like if you could use the cloud or you could pick between any database, you're limited to a few with these requirements.
But if you run and run it performantly on the phone, there's really no databases.
So we kind of had to pick a database that kind of fulfills some of our requirements and then build on top of it integrated into the mobile, build it so we can import it into react native, all these kind of things and work on the sync as well.
So that's a big part of it.
And then when it comes to the machine learning models, what we can do is.
So for example, right now we're sitting and working on an embeddings model to do vector embeddings.
And if some of you have tried to use OpenAI's ADA embeddings model, it's great.
And it's, you know, it's pretty fast.
It's like 100 milliseconds if you are, you know, on a decent network connection.
Right.
But if you're out in the Metro or something, it might be 200 or 300 milliseconds and you do a lot of vector embeddings in these kind of applications, like a lot.
So we want to bring a vector model locally and we can actually do that.
A model is like, you know, maybe 100 megabyte ish, and you can do vector embeddings in 5 milliseconds locally, no network round trips.
Right.
And that just enables so much more.
You can do so many more embeddings, you can do so many more like searches.
So here, like this is actually something that the user can really feel when it comes to the performance.
Yeah.
And it's.
And it's really important to get that latency down and get that better experience.
Absolutely.
Talk to me a bit more about device transfers.
You say data is local in the device, so it's not synced to icloud or wherever equivalent on Android.
If you were to move the data between devices, then is there a similar model to signal where you do device to device transfer rather than through a cloud service?
Or how is that anticipated?
To be working.
Assume you, I assume you haven't actually built that out yet, but it's something.
That you have planned out that is not built out yet.
But one of the technologies that were researching a lot through these two years is a technology called decentralized web nodes.
Okay.
So it's actually a standard that came out of.
So we can think of it as a personal database where you can store your information data and it's all encrypted with your private key.
So you have control of it, you decide who you want to share it to.
You can have like shared schemas and protocols.
So if, for example, think of a music application like Spotify or something like that, Spotify has all its own data and music and that's great, but it also has my user data playlists and stuff like that.
And if I go to Apple Music and I want to play some music, I sit there and I have to start over again, right?
Why should I not have my data?
Hence Spotify can have their data and then we agree on some shared protocols for the data.
That makes it easy to jump around and use the services that you want to do.
Obviously there's a lot of incentives and all kinds of things we can talk about.
Maybe Spotify would not want to do that, but that's kind of the vision of these decentralized web nodes.
So this is the web3 concept of essentially of data portability, right?
Yeah, exactly.
And then it's kind of taken to the next step because with web3 you are very focused on, you have the private key part of it.
Of course, that's what enables everything.
But then it's like NFTs or basically transactions on some blockchain, more or less private or anonymous or you know, most people kind of mix those two a little bit together.
I think it's very important to separate them.
But it's not a place to store your conversations, private conversations or your health data or whatever.
Right.
That needs to go somewhere else.
And that's what this, that's the vision of these decentralized web nodes.
So it started like, I don't know, five years ago or something like that at Microsoft Research.
And then it was kind of taking over by a group that left Microsoft and became tbd, which is a company under Block xyz, this Jack Dorsey conglomerate of companies.
And it's some very geeky people that are also super into self serve and identity and ownership and stuff like that are building out like the first implementation of the standard.
So the standard actually lives at Decentralized identity foundation.
But TBD is doing the first implementation of it and we are working with them and trying to get that off the ground.
But it's still, it's a work in progress.
But the promise is amazing right now.
So I mean I love the idea of that, right, because that's.
It's kind of the best of both worlds.
You control your data, but you can still.
There is data portability like the traditional web3 use case has always been.
You buy a gun in some game, right?
And then.
Well, or something in a game and then you start another game.
He's like, well, I want to bring my own inventory over.
I'm not very bullish on that use case, but I am bullish on the fact that I can control my data in this concept.
Right.
If that be preferences for application or even health data or chat data for that sake.
Right.
I think that's really interesting.
Okay, so the other thing I want to chat about, and we've kind of spoken about this already before, is the whole privacy element around cryptographic identities.
Because that's a tricky one, right?
We all know how to secure, properly secure your self hosted crypto with hardware.
All that stuff is kind of table stick, but.
For some people.
Yeah, correct.
I was going to get to that.
It's table stakes of people who have been around that world for a long time.
But for the vast majority there's a reason why the exchanges exist, right.
Because most people do not want to mess around with their own hardware, wallets.
And if you do not have a self governed private key for your data set, be that a some kind of sovereign data storage somewhere, then you kind of the whole concept kind of void if somebody else controls that key.
So you guys opted for a third party solution for that right now, which is very pragmatic.
How do you guys make that decision?
How do you think about that and on all those.
Because that's kind of key to the whole privacy debate.
Yes, it's a really good question and it's something that we have been thinking about a lot.
I think.
First of all, you have to understand that privacy and security in general is not a binary thing.
It's not either you have no security or you have 100% security.
It's a scale.
And you can look at Facebook messenger versus WhatsApp versus Signal, they're all on the scale.
And you have to decide between security versus convenience.
At least that's been the classic trade off.
Things are happening now that of course kind of break down some of the Barriers a little bit.
But our take is that we could go all in and create something that is as secure and private as possible.
Key generation locally on the device, seed phrases under the pillow, the classic stuff, right?
But if that lowers the convenience to a point where people are not going to use it, what do we win of that?
Right.
Then people stay with big tech because the alternatives are too cumbersome and then we win nothing.
Right.
There's no progress towards that future where we want to go.
We have to have some pragmatic approach to it.
The vast majority of people can barely remember their passwords and if they do remember them, then they probably reuse them over 20 different services, which is why I haven't pawn.
And things like that just like flare up every time there's a leak because people reuse passwords.
So if people can't really even use a password manager to generate passwords uniquely, the faith in them being able to be custodians of their own private keys is slim to nil.
Right?
Yeah.
I think there's so much work on this Also in the web3world, right.
With account abstraction and all these social recovery methods and stuff like that.
So we will have more options in the future and we're probably also going to implement some different options and probably also that you can, if you want, connect your existing wallet, metamask or whatever and then, you know, go from there.
Right.
But if I need my sister to sign up tomorrow and start using kin, that's not the way to go.
And we need to get to a place where we actually are allowed to survive.
Right?
Because if we don't get any users, then we can have the best vision in the world, but it doesn't matter.
No, I mean, I completely understand the decision that trade off you guys been making because that makes a lot of sense.
Right.
I, I personally, like, if.
If you get to a point where you do chuck all your calendar data, all your emails, all your messages or your bank statements, like I personally would not be confident unless I can provide my own private key.
Right.
But I, I do also appreciate that I'm a extreme outlier in the use.
Right.
So I wanted to bring up another thing that we spoke about some months ago, which I think this is super interesting.
Once you start to build this data set, and you kind of alluded to it before already, is this console of agents, and it kind of dawned on me when we had this conversation about calendly is a good example of this where you're like, well, you can broadcast your calendar And I can pick a schedule that works for me.
That's great, but it's kind of a halfway point, right?
It's better than two people emailing back and forth.
But you are halfway point.
What I really found fascinating was this whole concept of agents that can chat to each other without disclosing too much information.
And I think the interesting part there is we can definitely deep dive into this.
But one of the interesting part here is it doesn't matter if it takes a thousand messages back and forth to come to a consensus, because they can happen over the span of 200 milliseconds or, I don't know, a few seconds.
Right.
So let's.
I would love to, like, unpack that and dive into, like, details on how you guys see that world.
Yeah, I believe that's gonna happen, but it's not gonna happen before we have some trust and verifiability in these interactions.
Cause let's imagine that, you know, you send your AI out to book a table at a restaurant, right?
That's.
That's a use case that a lot of people bring up in podcasts, stuff like that.
And that's great.
But how does the restaurant know that it's coming from a real human?
It could be like an army of earbuds that are just out to book all the tables and whatever.
There's so much that can go wrong.
Right?
And right now you call and they kind of verify you have a human voice.
That's also changing.
We have AIs calling now.
Google can prove this.
Prove IO like what, five years ago?
But yes.
Yeah, exactly.
But even booking online, you are required to sit there, and there's maybe a captcha that checks that maybe there's actually something that looks like a human at the computer.
Right, but if it's AI is communicating to each other.
Exactly.
So we need a way for the AIs to prove that they're acting on behalf of a human.
And that is something that we believe in a lot.
And.
And that's where the technologies around decentralized identifiers and verified credentials and signatures and stuff come into play.
Right.
And then there's the question about how do they communicate to each other?
Because I think we're seeing different concepts kind of emerging.
So there's some people that are trying to do, like, agent communication protocols that are mapping out different schemas, like this is how they should talk to each other and stuff like that.
And I think some of them are taking a little bit more classical protocol approach that are mapping out a lot of schemas of how data should look and stuff like that.
And that's fine.
And that's also what we're very used to from the API world.
But now we have this different entity, the LLM, that can actually communicate in plain text.
Right.
You don't actually need schemas to do these kind of things.
So that's, I think that's really interesting.
And I think it's both going to be agents that are interacting with each other.
But one of the use cases that we're also looking at is you might want to share some of the data that you have in kin with other people.
Right.
Because data really gets value when it's shared.
So we don't want to, you know, there's no point in locking it down.
That's, that's wrong.
We need to be able to share it and use our data to our advantage.
So imagine that, you know, I go to my therapist every two weeks and my therapist could ask my AI to generate a little summary of what happened in Simon's life over the past two years.
Up to two weeks.
Like what physical activities did you do?
What interesting conversations did you have together that we can maybe dive into?
And that therapist can just send a request to my AI in plain text asking what they want.
My AI can, of course, with my permission, prepare response.
I can approve it or automatically approve that it's sent back and she can decide what format she wants it in.
Right.
Well, maybe she just wants it in plain text on an email, but in another scenario where you're maybe on a website, you want to buy some clothes.
And website is like, I'd really like to know, what does Simon like?
What's his favorite style of clothing, what his size is, what did he buy last?
So we don't suggest him to buy a winter jacket because maybe he just bought one or maybe he talked about he want this.
And in that case, the websites or the web service could ask my AI to prepare some data in a JSON schema that they can interpret automatically.
And my AI can do that because they can do it in any format that you request.
And that just enables a whole new world of personalization and actually using your data for your advantage.
Yeah, I found that the whole debate about schemas being somewhat redundant is super fascinating.
Right.
Because we're so used to the web paradigm of schemas and how you define data and just throwing that out the window is.
I mean, it's a big paradigm shift in how we interface with data.
Right.
So I Find that fascinating.
And so the other thing I wanted to cover is the whole identity piece that you mentioned before.
You want to book a restaurant, you want to prove that you are who you are or whatever.
I think in the last few years in crypto in general, we've seen a big push, obviously for kyc, for instance, but you always have this massive friction between the OGs of the crypto world, where there's a massive libertarian over tone of like, privacy versus the exchanges, where they're bound by legal requirements for whatever jurisdiction they're in.
Right.
How do you see that changing?
Is the privacy dead?
Is that something that is just going to die out, or are we going to see two parallel universes inside of this agent world?
Or if we apply to that?
I definitely don't think privacy is dead.
And I think there have been taking really.
So one of the challenges with.
There's many things here, but KYC has always been really cumbersome to go through.
Right.
And then you get this kind of approval that is like very specific to that application that we're using.
And then you go to the next application and then you need to do a new kyc, which is really annoying.
But what if we could do like we do in the real world and we just rely on some identities or some, you know, credentials that are issued to us by trusted issuers.
Like the government issued my, you know, driver's license, for example, and I can use that, you know, for KYC in a lot of different places.
Right.
Because they trust that privacy.
We see that EU is actually, you know, this year starting to roll out what they called identity wallets.
So that's based on verifiable credentials and decentralized identifiers.
And we are going to have, you know, in Denmark, we already have like a digital driver's license, but you can't really use it for anything.
You can pull up the app and then you can show it.
But we actually going to have these verifiable credentials that can be issued to any wallet that supports the standard, and you can hold that credential and you can use that to prove who you are.
And that could be also in an agent interaction where you say, okay, my agent is allowed to derive a proof from my driver's license or basically like present it to that destination or that other party and they can verify that this is actually a transaction that is first of all coming from a human.
And the human is above 18 or whatever else I want to check and verify.
Right.
Yeah.
So I think again, the schemas is probably going to be.
We will need some kind of schemas and some standards for doing these kind of exchanges.
Right.
But then it's probably a combination of some lightweight schemas that handles these verifications and then.
And how to establish connections and stuff like that.
And then exchanges in text format or whatever format that is now the most relevant for that exchange.
Yeah, yeah.
Because the KYC stuff is even like LinkedIn is doing KYC now.
Like you can connect your account to your passport.
And I understand why they're doing that because obviously to counter fraud accounts or whatnot, but having to upload your passport to LinkedIn doesn't quite feel like kosher to me, just to verify my account.
Right.
So having a party like that would definitely work.
And I know you mentioned the Danish government is doing digital driver license.
I think bank ID is also a thing in Denmark, I believe, and that's been around in Sweden at least for almost like a decade, I want to say, for accessing all government services and all those things.
And do you anticipate these kind of government entities being kind of the authority for identity or how do you think they have a place to play in this kind of realm?
I think they have a big place to play because they are in the end organizations and entities that we trust and we are used to trusting.
So what we need is some standards that go cross border.
Right.
That are global and that is what EU wants to do now.
Because let's say we have experimented, especially in the Nordics, with these different digital identities and verification stuff like that.
But I had a company in Sweden and I couldn't use my Danish Nemida, as it was called back then.
Now it's called MIDI Day because it's a different system.
Right.
And I couldn't get actually a Swedish one because I was not a Swedish citizen.
Right, right.
So I had to rely on like the old little yellow device where scan a QR code and you enter some, you know, codes and stuff like that to verify.
So in order for this to take off, we need these globally adopted standards and that's what at least EU is pushing out of that.
I know that US is also doing it with driver's license in different states.
So we will have now these like call them credentials or objects that are no longer like bound to a specific ecosystem because it's an open standard and everyone can work with these credentials.
And that's just going to enable a lot of creativity, I think, and reduce a lot of friction and a lot of interactions.
Yeah, no, I Think that is key.
Finding that balance between revealing privacy and I think that's being able to do a KYC without actually revealing the private data.
That that would go a long way in a lot of these systems and standardizing it.
The last thing I want to cover was AI hallucination, which is a real thing.
Right?
What have you guys seen that?
Because it's obviously if you use it as a private personal assistant, those things can be.
Have pretty significant consequences.
Where's your head around that?
And AI safety at large.
And probably hallucination is just one of those.
Obviously.
Yeah.
I think hallucinations are quite tricky because you can go one way and limit your model so much that it's only using the data that you provided it.
But then it also becomes much less usable because you want that full knowledge that it was trained on.
And in that span is where it's very difficult to not have hallucinations.
So what people try to do is use.
It's called retrieval augmented generation.
It's like use that large model and then feed in some data from some data source into it and then try to prompt it or fine tune it to really use that.
But still being able to kind of, let's say, also pull upon its larger knowledge of what it was trained on.
Right.
So the challenge is really mixing those two together and reducing hallucinations.
And there's a lot of work on that.
And you can, there's many different approaches, but it's not a solved problem.
And I think in general, people have to learn that these LLMs are a new type of interaction that they have to understand a little bit the technology behind it and that if it tells you something, it's not necessarily the, you know, ground truth.
Right.
That you can rely on.
Just like if some guy on the Internet writes something, it's not, you can't trust it 100 right.
If it tells you to jump out of a bridge, maybe like reconsider that thing.
But that's a, that's a big challenge when you're trying to make, you know, personal AIs that become so human.
Like.
Yeah.
But are still not human.
Right.
It's something else.
So we have to understand what is this new technology and this new relationship and how do I use it and not use it.
I think that's the hard part.
Right.
Because I hardcore chatgpt user uses it every day for a lot of things.
And quite frequently you do have these halls nature.
It just gives you something that's plain wrong.
But it's so Convincing.
When it tells you it, you're like, no, I know for fact that is wrong.
And then it's like you correct it.
It's like, oh, yeah, I'm wrong.
So that's obviously a massive one to cover.
Are you doing like a B?
I know in GPT4, at least they give you two responses and you can rate that because I guess that's kind of a quasi way to kind of solve that as well, indirectly.
Yeah.
So that's another big challenge when you're working with private Data.
So unlike ChatGPT, we can't just use all the conversations that people have and their feedback on those conversations to improve our models.
And we are brainstorming different ways to kind of get feedback from the user and feed that back into the model without having to necessarily see the whole conversation.
And maybe we can incentivize people to share some part of their conversation.
Another track that we are slowly looking at is can we simulate users on our platform?
So imagine that we create 100 users that are actually AIs that have maybe learned a lot from some data that we have shared with it about how users are interacting that we can generate.
And then we just let these Users or these AIs interact with our system every day.
And based on the responses here, we kind of.
That was a good conversation.
That was a bad conversation.
That was a good answer.
That was a bad answer.
So we kind of simulate that data that we need to feed back into the model to improve it.
Right.
So generating fake data is a big thing.
So similar to what DeepMind was doing in the early days, like having AI playing against AI in games, but in a different context, I guess.
Yes, yes, exactly.
Cool.
This has been super interesting.
Simon, I'm very happy that I got you on the show here.
Do you want to do a quick shout out to.
Well, to kin and what people, working people can learn more about kin.
And when you guys are expected to go ga, I guess that might not be a hard date yet.
So people can learn more at least.
Or at least at very least join their waitlist.
Absolutely.
So you can go to our website, which is Mykin AI and read a little bit more.
You can sign up for the beta.
We are letting in people as we go.
But if you write in the like, where did I hear about your field?
That it was on the Learning out with Victor podcast.
I'll be sure to kind of let you in, fast track you.
And then I just want to say that if you are an engineer that are also considering some of these different concepts, working with them.
There's very few people that are thinking about these kind of things.
A lot of software engineers have experience with, let's say, now, implementing some LLM stuff, but usually it's like, go to OpenAI's website, check the docs, integrate an AI, integrate the API, and do a chatbot or something like that.
And we've been trying to hire people, and that's usually the experience.
And if you want to go like one level deeper is the understanding of what can the technology do.
We talked about these different abstractions of agents and agents communicating together, and it's very experimental.
There's so many research papers coming out.
So if you're one of the people that are actually really interested and really into that, reach out.
And yeah, we're looking for bright minds to join us.
If I was not busy with other things, I think I would definitely contemplate joining you guys because it is a really cool space you guys are working in.
Cool.
Thanks so much, Simon.
Have a good one.
Talk to you soon.
Thank you.
Thanks, Victor.
Bye.
Found an error or typo? File PR against this file or the transcript.