YT YouTube2Text

YouTube2Text → Transcripts → Who Watches the Watcher? Understanding Privacy and Power in AI

Who Watches the Watcher? Understanding Privacy and Power in AI — Transcript

by Secret Network · 3,937 words · 571 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:07So next up we have who watches the
0:10watcher and it's moderated by Anubis and
0:13if we are discussing the parenting of AI
0:15then Anubis is the father of this
0:17conference. He's pulled it all together.
0:19So Anubis please come up and welcome on
0:22your uh panel and talk about who watches
0:24the watcher. Thank you.
0:28looking down. [music]
0:41>> Check. Check.
0:43>> All right. Thanks everybody for coming.
0:45I'm excited to dive into this panel.
0:47Bring my speakers on up to the stage,
0:49please.
0:59I am Anubis, the CMO of Secret Network
1:01Foundation.
1:02>> I'm going to be moderating this panel.
1:05Uh I think it's going to be a great
1:07conversation. We're going to dive into
1:09who watches the watcher, understanding
1:11privacy and power in artificial
1:14intelligence. I'm going to go ahead and
1:16pass it just around. Let all of the
1:18speakers go ahead and introduce
1:20themselves.
1:22>> Sure. Hey everyone. Uh, my name is
1:24Marshky.
1:25Uh, it's really good to be back in
1:27Toronto even though I live in San
1:28Francisco. Um, I'm a UFT uh, alumni. Any
1:33uft folks in the audience?
1:37No. Wow. Okay. Uh, so much for pandering
1:40to the audience. Um, yeah. So, I I work
1:44on Cella. I'm one of the co-founders of
1:45Cella, which is an EVM compatible um,
1:48L2. uh and uh recently we started
1:52self-protocol uh self-protocol is a uh
1:56zero knowledge privacy preserving uh
1:58decentralized identity protocol. Um who
2:02here has a passport? Anybody have a
2:04passport?
2:06Nobody has a passport. Okay, a few
2:08people are raising their hands. Amazing.
2:10Uh so passports nowadays have these NFC
2:13uh chips on them. uh they're called
2:16biometric passports and in there your
2:18biometric data is signed
2:20cryptographically with a certificate
2:22authority chain and so self-protocol
2:24basically verifies uh that certificate
2:28authority chain in zero knowledge so
2:30that you can prove onchain that you have
2:33a valid passport and better yet so that
2:36you can disclose uh information about
2:38yourself in a privacy preserving way uh
2:41with third party DAPs that you might
2:43want to interact with. So, think of it
2:45as a uh yeah, a really cool new uh
2:49mechanism that allows you to prove that
2:50you're a unique individual, that you're
2:52a human, that you're not an AA agent uh
2:54in a way that's very privacy preserving.
2:57Uh very relevant in this new agentic era
3:00that we're in right now. Um so, yeah,
3:02that's me.
3:04>> Hey everyone. Um I'll help Maric out. My
3:07co-founder went to uft, so at least uh
3:11we have another person. So uh anyways uh
3:14I'm Michael co-founder and CEO of uh
3:16ZeroG Labs. Uh quick background on
3:19myself. I started at uh Microsoft SAP
3:22Labs and then also worked at Bay &
3:23Company and Bridgewwater Associates and
3:25before ZeroG started web 2 unicorn
3:28called Garten that came out of Y
3:30Combinator and now ZeroG what we're
3:33building is the largest AI layer 1 uh
3:35ecosystem and as part of that we have a
3:38decentralized storage network. We've got
3:40a compute network that does inference
3:42and fine-tuning uh in a privacy
3:45preserving way using tees and we can
3:47definitely get into that. And then we
3:49have a layer one that's uh essentially
3:50infinitely scalable both on the data
3:53throughput and on the transactional
3:55level so that you can literally build
3:57anything on chain. So anything that's
3:58possible in web 2 including many AI
4:01applications you can now build uh fully
4:03on chain with our infrastructure.
4:06Hello. And uh this is Chiao and I'm the
4:09co-founder also CEO of Kopus. So uh I'm
4:12an artist and uh do you have any artists
4:14here creating you know no matter what
4:18you like? Oh yeah we we do have a few
4:21and uh we we're a content creation
4:23platform and we're also making a a
4:26content discovery engine which based on
4:29what you are viewing uh it will
4:31recommend you more related content no
4:33matter where you're at. So just like you
4:35may also like section but anywhere you
4:37can visit and uh yes we we use zerog as
4:40our storage for for our content. So
4:44that's why we're here. Awesome. Uh hi
4:47everyone. I'm Sebastian C at Ember AI.
4:50Uh we are infrastructure for execution
4:53uh focused on di uh agents. Um I guess
4:57to plug maybe this uh we'll be releasing
5:00uh a framework to build agents that do
5:02things in DeFi, automate strategies of
5:04all kinds. Uh the arbitrum vibe kit. So
5:07it's uh meant to be a vibe coding
5:09experience only. So you have your own
5:11trading strategies you want to just
5:12automate for you. Um keep an eye on
5:15social media for that.
5:17>> Awesome. Awesome. A lot of really good
5:19projects here. And at Secret Network, we
5:23build decentralized confidential
5:24computing infrastructure for web 3 and
5:27AI. Uh, and what we're currently
5:31launching is the Secret AI platform. Uh,
5:34allowing users to basically own their
5:37own data in centralized AI uh, platforms
5:42such as Chat GPT.
5:44all the data when you use chat GPT open
5:48AAI sucks up all that data uses it for
5:50their own means and to train their
5:53models. Uh we believe that users should
5:55own that data and uh at least be paid
5:58for it if uh if they want to let someone
6:01else use it. Um so on that note, we're
6:06obviously very focused on privacy. Um
6:09this panel and this event is very
6:11focused on privacy. So I wanted to just
6:13ask each one of you, how does your
6:15projects focus on privacy? How do you
6:18approach privacy with your own
6:20platforms?
6:22>> Yeah, great question. So I think um
6:26yeah, so self by the way it's self.xyz
6:28if anybody wants to check it out. It's a
6:30really nice short memorable domain name.
6:32Um we use zk snarks uh to ensure
6:36privacy. Uh and you know it's funny
6:38because a lot of people talk about ZK
6:40projects these days and really uh what
6:42they mean is the the kind of snark part
6:45of ZK snarks right the succinct part if
6:48you look at ZK EVMs are not actually
6:50offering uh typically any privacy uh but
6:54for us we're actually using um you know
6:56the ZK part so the zero knowledge part
6:59of ZK snarks so if you're not familiar
7:01with that that's a lot of u very
7:03advanced math and cryptography uh some
7:06people call it moon math math uh and um
7:09it's pretty incredible. Uh but with this
7:12moon math, you can effectively prove
7:14statements uh without revealing um the
7:18inputs to those statements. And so you
7:20can prove for example uh using self uh
7:24that you are not on the OFAC list uh or
7:27that you are you know not um that you
7:31are an actual human with a passport uh
7:33without revealing you know the
7:35information on your passport. I mean the
7:37OFAC one is pretty interesting like you
7:40don't have to reveal your name, your
7:41birthday, which which country you're
7:43from. uh but yet you can prove that
7:45you're not on this you know US sanctions
7:48list uh that u you know many banks for
7:51example enforce um to you know prevent
7:54kind of moneyaundering and all of that
7:56and so it's really really interesting I
7:58think the the age is another really
8:00interesting one you can prove that
8:01you're over 18 or I guess in in Ontario
8:03over 19 without revealing your actual
8:06age and again we can do that with with
8:08this advanced cryptography uh which is
8:11just really really cool to
8:14That's really powerful. I I've always uh
8:17wondered why it is that every time we go
8:19to a store and we want to buy alcohol or
8:20whatever, uh why we have to show the
8:23cashier exactly where we live, where we
8:25sleep, you know, it just uh it's always
8:27bugged me. So, but on that same note,
8:30what's really cool about what you're
8:31building there is proof of humanity in
8:33in the age of AI. I think that's going
8:35to become more and more relevant uh and
8:37very just very important. So,
8:40>> yeah, absolutely.
8:42All right. And Michael, go ahead.
8:44>> Yeah. For us, uh, the way we think about
8:46privacy is, uh, because we have
8:49different layers to our infrastructure,
8:51different, uh, aspects of privacy matter
8:54at these different layers. So, for
8:55example, on the layer one, it would be
8:57can you do private transactions. So, we
9:00haven't implemented that, but that's
9:01something that we're looking uh at in
9:03the future. uh what we have is on the
9:06for example storage layer you could
9:08easily encrypt your data and that allows
9:11a sense of privacy and technically we're
9:13also agnostic to what kind of privacy
9:16methodology you want to use. So if in
9:18the future FHE becomes uh compute
9:21inintensive enough then you can
9:23definitely use that from a storage
9:24perspective as well. So we don't
9:26necessarily want to prescribe what the
9:28privacy methodology is but really enable
9:30it through our infrastructure. Uh on the
9:34compute side that's where it gets kind
9:35of interesting because different
9:37approaches have actually been tried. So
9:39ZKML
9:40um SPML TML FHML are all different ways
9:46of approaching um inference and other
9:49type of compute applications through a
9:51privacy lens. But the issue that we
9:54found is that none of them are really
9:56production ready for consumer use cases
9:58besides TML because the overhead that
10:02you have between TE and kind of regular
10:04let's say inference that's not
10:05verifiable not private is almost zero
10:08and so you could use it in the same way
10:10as if you would uh you know chat GPT but
10:12with an open source model running in an
10:15TML which is stands for trusted
10:17execution environment on an H100 Nvidia
10:20card and So that's why we've chosen to
10:24support that at the moment, but we
10:26definitely can also support ZQML and
10:29FHTML and all these other verification
10:31methodologies. It's just that if you
10:34wait an hour for a token to appear,
10:36that's not a great user experience. So
10:38for certain use cases, that makes sense,
10:40but they're very very limited. And so we
10:42primarily use TE's. The way TE's work,
10:45it's a hardware based methodology. So
10:47there is actually an environment on that
10:49graphics card that's completely private.
10:51So you can run computations in that
10:53environment and then verify that
10:55something happened in that environment
10:56without outsiders knowing about it. And
10:59so it's quite powerful. It works very
11:02well. But there's a downside to it as
11:04well. You have to trust the hardware
11:05manufacturer. If the hardware
11:07manufacturer for some reason builds in a
11:09back door for a government entity, you
11:11don't necessarily know about it. And so
11:14it may not be the end solution, but it's
11:16the most practical solution at the
11:17moment that we've chosen to support just
11:20from a a way of kind of working with
11:22these systems, but definitely want to
11:24support other methodologies in the
11:25future.
11:27>> I love that answer. It's uh at secret
11:30network, we also use TEES and for much
11:32the same reason. Uh there's other very
11:35powerful encryption technologies for
11:37confidential computing and each one has
11:39its own place. uh for example ZK is very
11:42good for something like an ID
11:43verification but for uh AI inference
11:47it's just not going to work something
11:49like a TE uh which has a a faster and
11:52more efficient throughput uh is what's
11:54needed
11:55>> there are people working on that it's
11:56going to happen
11:57>> oh I think so yeah you're there is
12:00people doing some very good work
12:02>> yeah he was saying uh people are working
12:03on it will happen and I would agree
12:05there's some use cases where if it's a
12:07very small parameter model it's it works
12:10works okay especially on like edge
12:11devices where you can use some type of
12:13ZKML methodology but for larger scale
12:16models even like 7 billion parameters it
12:18already starts breaking down because
12:20it's a lot of latency you're introducing
12:22with the methodology
12:24I I've always figured that the future is
12:26going to be multi-technology not just
12:28multi-chain and so a mixture of a TE
12:31with ZK FHE uh and one of my favorites
12:36is actually MPC and and TE as a
12:39combination
12:40By combining the two or two or more, you
12:43get very powerful use cases.
12:46>> Oh, were you gonna say something?
12:52>> I'll pass it.
12:55>> All right. And ciao, same question. How
12:58how does how does your project approach
13:00privacy?
13:01>> Well, so like one great thing uh about
13:04web three is uh they can keep you off
13:06from authority. And recently you have
13:09found like a very interesting uh product
13:11market fit is the fanfiction writers in
13:14China. Uh because uh there are lots of
13:16fanfiction uh websites have been taken
13:19down and the police even go to invite
13:21those writers uh for some tea party and
13:24what we what we do is like we allow the
13:27wallet only login. So first like nobody
13:30will know like who writes those
13:31fanfictions and keep you safe from the
13:34biggest authority. And uh another thing
13:37is like we support uh crypto
13:39transactions and uh these are all like
13:42web two writers but right now they are
13:43learning to login with their wallets and
13:45also u to earn with their wallets. So
13:48it's a safer way to them uh for them to
13:51no matter where they live basically. And
13:53the second thing is how we protect the
13:54content creators right you want to you
13:56don't want your own content to be um
13:59pirated or learned by the AI very
14:00easily. So what we do is we have like a
14:03like a maze. If we detect there's a AI
14:06agent trying to visit our uh creators
14:09page like trying to scrap out some
14:11content uh we'll just generate random
14:13content for them. So it if the creators
14:16are actually writing about a Harry
14:17Potter's fanfiction, it might just
14:19generate them like a Amazon River um uh
14:23ecology study uh just totally random
14:26nonsense and uh good luck with those AI
14:28agents. Yeah, that's what we the two
14:30main thing we do to protect the privacy
14:33uh for our users.
14:36>> Um I I think everyone that's working
14:38closer to the end user and kind of on
14:39the application layer ends up with this
14:41temptation to in AI make very simple
14:44user experiences that require a lot of
14:46data and kind of a lot of personal
14:48information from people. Um it's really
14:51important when you're thinking about how
14:53to build the systems to try to separate
14:55those two things and within reason,
14:57right? Um, obviously people are looking
15:00for that automation, but there there are
15:02ways to let them own that information,
15:04right? For us specifically, I mean,
15:05we're dealing with DeFi execution. It's
15:07all about, you know, owning your own
15:08strategies. We do take a lot of, uh,
15:12data points that can give you good
15:15suggestions, help you, you know, explore
15:17the world in a more curated manner. uh
15:20but the the key of this conversation I
15:23think lies around making sure that the
15:25data that you are using is not leaving
15:27the user's possession right maybe you
15:29get permissions to access it maybe your
15:31systems built out in a way that you'd
15:33never need to touch it um so you know
15:35for us distributing this as an open
15:37framework and letting people build out
15:38whatever strategies they want on their
15:40own computers owning their own code is
15:42uh I don't know kind of the the best of
15:44both worlds in a way
15:47>> awesome awesome So,
15:50one of the things to really consider in
15:53in the AI space and privacy uh in
15:57centralized AI systems,
16:00you know who to go after if there's a
16:02large data leak that that exposes a
16:04bunch of sensitive information. You know
16:06that you can go after open AI. You know
16:09who's responsible. Now, in decentralized
16:12AI, it's a bit different because you
16:15don't know who's hosting the the
16:17servers. you don't know who owns the
16:19hardware. So my next question is really
16:21double clicking into that. Um in your
16:24opinion, who should be responsible for
16:27setting privacy standards in
16:29decentralized AI?
16:32>> Yeah, it's an interesting question. Um
16:36I mean I think ultimately it's it's it's
16:38going to have to be initially um the
16:42decentralized projects, right? I think
16:44initially this will be too complicated
16:45for regulators to step in and I think if
16:47regulators do it uh immediately I think
16:50they may also um struggle a little bit
16:52to understand all the nuances but if the
16:55if the projects themselves if the you
16:57know the the open source community
16:59doesn't do it quickly uh then you know
17:02certainly bad things could happen uh and
17:05then we might see an overreaction from
17:07from regulators uh and so in in a way
17:09that could be even worse and so I think
17:11it it it's certainly going to be um you
17:14know the people in this room, the people
17:16working on these systems um you know I
17:19think we have to be as thoughtful and
17:20and careful as possible. I mean this is
17:23a crazy transition that we're going
17:25through. Uh I mean it's really hard to
17:27even predict what's going to happen in
17:29the next 5 to 10 years. Uh if AGI really
17:32happens, you know, everything about our
17:34lives may may change just overnight. And
17:37I have a lot of conviction that, you
17:38know, decentralized systems are going to
17:40be the ones that are are going to be
17:42preventing concentrations of power,
17:44ensuring that people's data remains
17:47private. Um, I mean, just think about
17:49how much data you share with, you know,
17:51Chad GPT these days. It's crazy. I mean,
17:54the amount of information that they know
17:55about you. Um, there's there's a lot of
17:58a lot of risk there. Um and so yeah, I
18:02think it's it's going to be, you know,
18:03it's going to fall on all of us.
18:06>> Absolutely. At at Secret Network, we
18:09we're building the infrastructure, so we
18:11always come at that infrastructure
18:12level. Um and to be honest, part of that
18:15is because it's it's just a need. So
18:18many people building in the space aren't
18:19even thinking about the word privacy. So
18:21if we can make it easy for anybody to go
18:24ahead and just plug in build uh the same
18:26way they would build anything but add
18:28that privacy you know that's that's the
18:30best case in my opinion. Yeah, maybe
18:32just a quick followup. You know, the
18:34topic of world came up in the in the
18:36panel before and there was a lot of um
18:39um concern about, you know, people
18:41sharing their iris information. You
18:43know, that's a type of data that, you
18:45know, I think um a lot of people have
18:47have concerns about sharing and it's
18:50really important in this future where
18:51we're going to be, you know, thinking
18:52about um um how do we how do we give
18:56UBIS, how do we differentiate between
18:58humans and agents, right? Right. I think
19:00this is the reason why world is is
19:01solving it in this way. And I think one
19:03thing that's really nice about self is
19:05that it solves it in a way that doesn't
19:07require you to share your your Iris
19:09information. Right? You already have a
19:11biometric passport. Um EU IDs at hard
19:14systems. There's a lot of ID systems
19:16that have these cryptographic primitives
19:18that we can then bring on chain in
19:19privacy preserving ways. And we think
19:21that that's going to be a really really
19:24great way uh to protect users data as we
19:27kind of go through this crazy transition
19:29that's uh coming about.
19:33>> Yeah. So just uh our own spin on kind of
19:36the question who's to blame if something
19:38happens at OpenAI you have a centralized
19:40authority and you could then extend that
19:43analogy and say well in a decentralized
19:45way you have the protocol so shouldn't
19:47the protocol be to blame. Um, but that's
19:51kind of like saying the individual
19:53actors that are trying to attack the
19:55protocol are then the ones that should
19:58be responsible, but the protocol itself
19:59is to blame. So the analogy doesn't
20:02fully extend in that way. What the
20:04protocol can be responsible for though
20:06is to put in safeguards. So for example,
20:09we have something called AI alignment
20:11nodes that we're currently researching
20:13and in the future we want them to
20:15basically uh figure things out like
20:17model drift, uh negative behavior by AI
20:20agents,
20:22um kind of data poisoning attacks, those
20:25types of things that really influence
20:26the functioning of the protocol. So we
20:28definitely want to be responsible for
20:30that, but we can't possibly predict all
20:33attack vectors. And so there will be
20:34protocol upgrades over time, but it's
20:37really kind of nefarious actors that
20:38will try to attack the network in one
20:40way or another. And so the best that we
20:42can do is to ensure that there is
20:44slashing or economic incentive
20:45conditions so that these types of
20:47situations don't occur in the first
20:49place. And there's probably a finite set
20:51of things that can be attack vectors
20:53that we can think through and then
20:54through experience keep upgrading the
20:56protocol over time. So that's something
20:58that we can take responsibility for.
20:59Sure.
21:02>> Uh I totally agree. And if you can't if
21:06it's hard to go after uh the the like
21:09the sellers, it's probably like easy a
21:12little bit more easy to go after the
21:14buyers. And uh I I I need I think as a
21:18content creator background uh they need
21:20better regulations for um about like how
21:24AI no matter if it's decentralized or
21:26centralized AI to steal I mean kind of
21:29steal or learn from people's content
21:31because uh every law including the IP
21:34law uh works very differently from every
21:36country and uh your your content might
21:39be safe uh by law in one area doesn't
21:42mean that you content is safe
21:45uh everywhere. So we do need uh u we do
21:48need like regulations and also uh
21:50decentralized uh uh storage providers uh
21:54which basically do less KYC to their
21:57users. So they just know like even
21:59though they know the content they don't
22:00know where to come from I mean it's it's
22:03still better for users that what we can
22:05do for best for right now.
22:07>> Just to jump in real quick because I
22:09think we're we're shorting time. I I
22:11want to talk about the government
22:12approach to this. It's definitely not
22:14the most popular take in web 3
22:16especially, but there is a
22:18responsibility to I don't know give a a
22:21high level guideline that will
22:23potentially be broken by private
22:24businesses, but that exploration needs
22:26to be held back in my opinion uh to some
22:29degree by the government. Right? I I
22:30don't think we can trust uh private
22:33interest to 100% put individual uh I
22:37don't know best benefit in the first.
22:40>> Awesome. Yeah. I think I think one of
22:42the big things that we all really should
22:45think about is is how we can improve the
22:47education in the space on on these
22:49issues. Uh we are out of time, but I
22:52want to just thank each and every one of
22:54you for coming up here um and for
22:56sharing. Uh it's been a great panel.
22:59>> Thanks everyone.
23:01>> Thank you.

About this transcript

This page contains the full transcript of Who Watches the Watcher? Understanding Privacy and Power in AI by Secret Network, generated from the public captions YouTube serves with the video. The transcript has 3,937 words across 571 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.