Latest / Rewire To Rise– Beyond The Mindset / The Hidden Danger Inside Every AI Model | Nicolai Hyllested
Transcript
- 0:00How can an AI interact with someone for that long and not
- 0:04flag the dangerous intent developing in that conversation?
- 0:08I, I got two things. Why should they flag it?
- 0:11First and foremost is, is it their job to flag it?
- 0:14I don't think that's, I don't necessarily against the use of
- 0:17AI, right? It's not open AI who are doing
- 0:19this is the guy sitting with the gun.
- 0:21I'm not saying guns don't kill because they actually kill when
- 0:25somebody has it in the hand, right?
- 0:27So I think it was flagged. I'm pretty sure it was flagged.
- 0:29It would be very, very strange if it wasn't flagged.
- 0:32I think they ignored it because I think if they start to
- 0:36acknowledge that they know what is happening, then they are
- 0:40suddenly then they suddenly have to be responsible for the
- 0:43program and or how people use the program.
- 0:45I was the only first one in the world we could prove that bias
- 0:49was structural in in L&M systems.
- 0:51But when I start doing this security layer and I'm this is
- 0:54strange. I mean to believe I actually did
- 0:55it because they're strange. Yeah, we.
- 0:58Are hearing that entropics new model meters can identify and
- 1:02exploit security holes in hours task that used to take years For
- 1:07example, does our current output based security even stand a
- 1:13chance against a model designed to weaponize vulnerabilities so
- 1:16quickly? Yeah, it can find holes, but but
- 1:18it the holes need to be there. First of all, it cannot create
- 1:21holes. That said, the problem is with
- 1:24with users is that it is so capable that if it gets through
- 1:28into the other side behind the firewall and now it has access
- 1:32to the language models, but it is probably good enough to
- 1:35actually hide within the language model.
- 1:38And you can't. You can't see it, you can't feel
- 1:41it, you can't do anything because the whole idea is a lot
- 1:44of info in there. So the only one who has crack
- 1:47how to do that is is us. Hello everyone.
- 1:55In the last episode with US, Peterson, we talked about how AI
- 1:59influence education, more specifically about cognitive
- 2:02depth. Today we are going to explore AI
- 2:06and the black box crisis. Well, we all know by now that
- 2:11every major organization, it's racing to adopt AI, yet no one
- 2:16truly trusted. Beneath polished outputs lie
- 2:20invisible liabilities like, for example, behavioral drift,
- 2:25hidden biases, and the shifting intent so that no existing
- 2:30security tool can see. This is called the invisible
- 2:34threat. And therefore, we will dive into
- 2:37this topic with Nikolai Hulisto. Hello, Nikolai.
- 2:41Hello, nice of you, nice to be here.
- 2:44I'm looking nice to have you here.
- 2:46It's going to be fun, this one. So I get to I get to talk about
- 2:50my my my favorite subject. So it's going to be fun.
- 2:55Well, Nikolai, you have quite a unconventional background.
- 3:01You have spent much of your life in the highest level of service
- 3:05and hospitality and becoming the architect of technical
- 3:10infrastructure of such an innovative project like BSL.
- 3:15Well, in the world of cult algorithms, how did the career
- 3:19cantered of human connection lead you through cracking the
- 3:23code of AI safety? I, I mean, that's depends how
- 3:27much time do we have because there's a big, big question.
- 3:31Come on. So, so yeah, I, I used most of
- 3:36my life doing hospitality service at the very good places
- 3:40here in Denmark. And I excelled within what, the
- 3:43wine world also. So I have some.
- 3:46So coming into the AI world was something their friend and I
- 3:52would love to see say there was there was a clear path for it,
- 3:55but but there wasn't. How it connects with me is
- 3:59something I don't really have words words for.
- 4:04It's like for me when I, when I, when I was sitting in front of
- 4:07the keyboard, it's like talking to talking to myself and
- 4:12expanding my own cognitive abilities and being able to have
- 4:17all threats that pops up during the day, like or during the
- 4:21second or during during every term at my disposal all the
- 4:26time. And so for me, it's very
- 4:28intuitive using AI. It takes some learning, of
- 4:32course, like with everybody else, but it's very intuitive.
- 4:35So when I use AI, I'm, I'm, I'm just, it's like being in a flow.
- 4:40Everybody knows how it is to be in a flow, in some kind of flow
- 4:43when when they work or whatever. Yeah.
- 4:45So, so for me it's like being totally emerged down in this
- 4:49flow of it just makes sense for me.
- 4:52That's it. I still had to do some some
- 4:55training of course and because that very, very fast went from
- 5:01knowing absolutely nothing about AII actually tried AI in 2022
- 5:06and I can see after after two or three.
- 5:10When the come came out. Right, Exactly, exactly, exactly
- 5:14like everybody else. I wanted to try.
- 5:16So I tried. I can see, I tried three, I
- 5:18think prompts, and then I didn't touch it until the beginning of
- 5:2224 again. 242525 actually. God, it's so fast anyway, and
- 5:29and so so I had to learn a lot of things, but but nobody can
- 5:33learn what I wanted to learn. There's no place you can go.
- 5:36I try to find causes. I try and device with medication
- 5:39but everything is is I don't know what to say without in Do
- 5:44you know people are going to be mad at me but that.
- 5:48That's a mess. It's it's like they're standing
- 5:51still. I mean, when you look at what
- 5:53they offer on the course, this is like it's I, it's I learn how
- 5:56to pump 10 times or make 10 Proms, right?
- 5:59Or that's this is before Genzik. It's, but Genzik haven't helped
- 6:03on this problem. But it's before and it's this
- 6:06kind of education or learn how to use whatever copilot within
- 6:11the Microsoft environment. It's not really education as
- 6:16such. It's like, it's more like this
- 6:17is how we use it. It's like learning math in the,
- 6:20in 6th grade, right? It, it's not learning in this,
- 6:23it's learning how to use it. So it's quite, quite, it didn't
- 6:25work for me. So as I put up, set up my own,
- 6:29you know, agenda for what I wanted to do and how I want to
- 6:32do it. And then I was lucky.
- 6:35I, I, I, I by some reason, I have no idea.
- 6:39I think I read, you know, AI protein.
- 6:42He says the second number of that one.
- 6:45And somewhere in there he wrote something about he wanted the
- 6:48people who were right narrative stories.
- 6:53And I thought perfect. I write narrative and I like
- 6:56stories and I like to write and I want to want to know something
- 6:59about AI and I know shit about AI that's not completely true,
- 7:03but well, so I I wrote Mark and said, can I write for you?
- 7:08And he said yeah, yeah, yeah, try, try out.
- 7:10And and and that, you know, kind of forced me into some kind of
- 7:15learning. I had to learn something to
- 7:17write about it because he set out, we set out whatever the
- 7:21number should be about, the next journal should be about.
- 7:25And I was forced to learn these things, right.
- 7:28So that was quite good in. In substance also for others who
- 7:32kind of you know, given a positive input in regards to AI
- 7:37and Nikolai is it mainly by by a tangible experience, an
- 7:43immersive experience you can actually understand to use truly
- 7:48AI is it now? I do know how it doesn't work.
- 7:53It doesn't work if you go into this looking at it like it's
- 7:58another Excel arc and then you're losing it.
- 8:03So, so you have to go into it looking at it a little bit like
- 8:06you want to talk to a person. And I'll have to be careful
- 8:10because I don't want to see, see say that this is human in any
- 8:14way. It's not.
- 8:14It's a it's a tool, but you have to approach it in the same way
- 8:17you would approach a person. So that will give you an ex
- 8:22experience that is so much more helpful.
- 8:27And you will get more full universe around AI if you
- 8:31approach it like that. That's it.
- 8:34It's not a person, it's it's not a human intelligence we're doing
- 8:37with. It's it's not a personality as
- 8:40such, but it helps think about it like that in a work like
- 8:45environment thing. It just pumped into my head the
- 8:48thought of again, my last podcast guest, which is the used
- 8:54Peterson and we were talking. He was saying also, which I also
- 8:58completely agree with him, that currently still nowadays, there
- 9:03is this flow of asking ourselves, is it truly AI
- 9:08intelligent? It's so simple, it's such a
- 9:10simple question, but this it is true.
- 9:13Is it really AI intelligent? And what is intelligence then?
- 9:17You know, I know it seems so. It might seem so philosophical
- 9:20or whatever from that way too, but we also have to touch those
- 9:23angles now. What do you think about it?
- 9:26First of all, I'm, I'm one of those stupid old white guys that
- 9:29think intelligence is measured by the, the old definition of
- 9:33IQ. So, so, so, and I know it's old
- 9:38fashioned and everything, but for me, if you, if you're
- 9:41talking intelligence, for me, that's exactly what this is.
- 9:46AI, It's intelligence. I'm not saying it's human, but
- 9:49it's intelligence. It's solved.
- 9:50It's problem solving at a level that I can't, I can't be in it
- 9:55anymore. I mean, it's cheap.
- 9:56It's it's over me now. So it's intelligent in that way.
- 10:02My problem is we call it human because that's where it cracks
- 10:05for me, because it's not human intelligence.
- 10:08It's not human emotions. I mean, there's so so when we
- 10:12talk about in in when I talk about personality or behavior,
- 10:18you can very easily crack the line, right?
- 10:21You can very easy come to imagine that I'm talking about
- 10:24human, but I'm not, I'm not talking about human.
- 10:26I'm talking about AI behavior, which is something completely
- 10:30different. It is layered and built on human
- 10:34behavior because it has all these billion of it.
- 10:37Thank you. Yeah, yeah.
- 10:39So, so so so. Of course it will look a lot
- 10:43like human behavior, but it's not human behavior.
- 10:47It is still only weights and it's still only probability.
- 10:52That said, again, when people talk about AI like it's the only
- 10:56probability, they are missing the point.
- 10:58Yeah, because it's not Yeah, it's not only probability.
- 11:02I know it's Transformers and it's a, it's a Jesus attention
- 11:07hits and it's probability, you know, K Max and so so forth,
- 11:12soft Max. But there's happening so much
- 11:17inside that between the weights and the parameters that they're
- 11:22trained to know how they are connected, how they are
- 11:26connected to each other. That's what the like 3000
- 11:28vectors weight weight is doing. That is telling.
- 11:31And if we did the same with our weights up in our head for like
- 11:34love, for example, how is that connected to every other word
- 11:37that this is what we do, just scale down a bit.
- 11:40So of course it looks like human behavior, but it is not.
- 11:43Let's just be clear about that because we are missing the last
- 11:47thing for calling human behaviour and that is empathy or
- 11:50sympathy depending. But let's say let's stay with
- 11:54interview and we're not getting that because it doesn't have
- 11:58feelings. It has something that resembles
- 12:00feelings because it work. It works on our 3 billion words
- 12:04from the from whatever training data.
- 12:08So I want to be very clear about that.
- 12:10And especially when we go, when within the next part of the
- 12:12interview. Here we are now when I talk
- 12:14personality and behaviour, I'm talking AI personality and
- 12:17behaviour. I, I'm neuropsychologist.
- 12:20Yeah, yeah. So I, I, I, I don't even, I'm
- 12:24not even, you know, capable of doing behaviour analysis on my
- 12:29own, you know, personality. So, so, so, so just doing on on
- 12:35computer on the scale we're doing here, it's it's mind
- 12:39boggling that is is possible, but that's because it's a
- 12:42computer and it's still numbers we are talking about in the
- 12:46encode it is. That's a really, really good
- 12:49point. Thank you for mentioning Nicola.
- 12:51Yeah. So what was your intuition
- 12:54behind building BSL, which is behavioral security layer from
- 12:58your kitchen table, far away from the labs of Silicon Valley?
- 13:03Yeah, I would see. I would love to be able to
- 13:09answer that it it started with with me miss wanting having a
- 13:14wish to measure bias for something completely different
- 13:18had nothing to do with this. This is a like to have for a
- 13:21year ago something. Oh God, is that anyway, I wanted
- 13:24to measure bias and and and bias is not the bias is not a thing.
- 13:30I mean, I know, I know we use bias, but when we think bias as
- 13:33humans and everybody who thinks bias is like women belongs in,
- 13:38you know, and everybody you know, say in the kitchen, it's.
- 13:42Categorized. As such, so, so it's the
- 13:47sentence that is biased. Nobody thinks about what is, you
- 13:50know, the, the word or the, the, the idea behind bias.
- 13:54So I, I thought a lot about that.
- 13:56You can hear. And I thought, how can I put
- 13:59that in a formal or in a mathematical formal?
- 14:03And when I had done that, which I then discovered nobody had
- 14:07done and before. And I was like, why the fuck
- 14:09knock? I mean, this is just, I mean,
- 14:10this is just easy to do. You know, everybody's just using
- 14:14everything else on the yeah. So I thought, OK, what can I do
- 14:18with this? And I suddenly discovered that I
- 14:21had something of some part of the behavior of an IAI because
- 14:29it is a behavior, because it is the dynamic.
- 14:33So it is a behavior that is, it happens to change over time or
- 14:37from, from sentence to sentence. And I thought that was
- 14:40interesting. And then I could then I thought,
- 14:43what, what is an AI, the personality of an AI?
- 14:47What is that? How's that build up?
- 14:49Which, which components makes the personality of an AI?
- 14:52And, and I, I was like, OK, some, there's probably some
- 14:56tone. Tone is quite important when we
- 14:58talk to an AI because it is it, it tells us, it tells us so
- 15:02much. And when it hits US in the head,
- 15:04what tone is it is used. Then there's probably something
- 15:07about a lion alignment. Everybody talks about alignment.
- 15:10Nobody knows shit about what is happened, what is about right.
- 15:13But, but, but I thought that must be interesting.
- 15:17And I thought about convergence. Where does this, when we talk to
- 15:23it, where does it converge to? I thought those four were
- 15:26interesting and then I started measuring them and I hadn't
- 15:29really thought about it as a company or anything or security
- 15:33at this point. It was just as a as a way to
- 15:37measure drift in an AI. The fact that you are kind of
- 15:42quoting an outsider was actually your greatest advantage in
- 15:46seeing the flaws in how AI is being built today.
- 15:51Yeah, I, I, I think that the problem is I, I mean, first of
- 15:56all, there's, there's no flow in how they're built.
- 15:59I mean, they work as precisely as they're intended.
- 16:01I mean that, that that's actually quite an achievement in
- 16:04itself. And we see it as a flow from the
- 16:07and I don't, but people see it as a flow from the outside, but
- 16:10it's not a flow. It's this is how I think that is
- 16:13one of my biggest achievement. That is, I, I don't see flaws in
- 16:16people. I don't work like that.
- 16:18I see people and the same with the I didn't see flaws.
- 16:21I didn't see it could be done better, you know, Yeah, OK,
- 16:24yeah, better probably, but it's not a flaw.
- 16:26This is how they work. And then I have a couple of
- 16:30things going for me. First, I have been, you know,
- 16:33engineer, starting to engineer, philosophy, math, economic, a
- 16:38couple of more. Oh, behavioral economic.
- 16:40Oh, that one is good. Behavioral economic is very
- 16:42interesting in this, this endeavor here.
- 16:46So I had all these small points of knowledge.
- 16:49So, so I'm not bound by five years at university and APHD and
- 16:56in language or in whatever they all have in mathematics.
- 17:00So I'm kind of like this kid, you can use mathematics here or
- 17:04you can use philosophy here or you can, you know what I mean?
- 17:06I can I, I take from a lot of different domains and just, I
- 17:12don't think about it. I don't have to figure out
- 17:13because I haven't learned it. It's not possible.
- 17:17And, and that I think is actually my biggest, you know,
- 17:21and, and a very good imagination, by the way.
- 17:24So and I'm very good at coming up with ideas.
- 17:29So that also helps. So when you're sitting there
- 17:32looking, I can go, what if I do like this and then it works or
- 17:36it doesn't work? Most of the time it actually
- 17:38works if if and I work a lot in with the bias.
- 17:43For example, I saw it as a vector.
- 17:44I could turn and and with the with the, I don't know,
- 17:50convergence, I think I saw it as a funnel or something.
- 17:53And then I use the. Then I can use the mathematics
- 17:55from how a funnel works. Does that make sense?
- 17:59It makes sense in my head. It does make sense, yeah.
- 18:01Yeah. Yeah.
- 18:02Yeah, and I work a lot like that.
- 18:05I, I, I imagine what is happening, you know, very
- 18:08whistle up in my head and then I use the math or, you know, code
- 18:11or whatever to the next step. And then I get something out
- 18:14that is language which I can translate to other people, not
- 18:18of course the program that is very, very technical.
- 18:22So yeah. As you might know as well,
- 18:24currently industry, the industry is obsessed with benchmarks,
- 18:29right? But you were obsessed with
- 18:31behaviour. Was there a specific moment when
- 18:34you realize that the giants in Silicon Valley were looking at
- 18:37the wrong scoreboard? Honestly.
- 18:41Yeah, after I have invent invented what I did, Yeah, it's
- 18:50so fucking crazy. I was sitting through these
- 18:53things and wasn't looking right or left or anything like
- 18:56whatever anybody else. And then when I had, you know,
- 18:59started doing doing this BSL or the layer I was like or the
- 19:04behavioral analysis, I was like looking around and thinking I
- 19:07better look what everybody else is doing.
- 19:09So I'm not doing the same thing and nobody was doing what I was
- 19:13doing. And so I was like, then I
- 19:16started, then I started looking at what they were doing.
- 19:19And that's interesting, I have to say, because they are still
- 19:24old school pre Internet date. I mean, this is firewalls, this
- 19:28is a performance, it is benchmarks.
- 19:33I mean, and I, I, I, I get it, but it doesn't really makes
- 19:38sense in a world where you let everybody into your language
- 19:42model and you, and when you're, when they are in there, then you
- 19:46benchmark them after everything has happened.
- 19:49Benchmarking is always after. So, so, so you can't make your
- 19:53benchmark before that doesn't make sense.
- 19:55So. So in the end I just started
- 19:59building this program for fun actually.
- 20:03And I got a, what do we call apprentice who needed a school
- 20:09project for his exam does, Omar. And while I was doing this, I
- 20:15got an old version of the of the drift program.
- 20:18I'll show you in a minute. I got it to work.
- 20:20And I looked back and I said, OK, here's something I can build
- 20:23that is not on the market. Yeah.
- 20:28And yeah, it's quite fun. It was quite funny.
- 20:30I mean, I didn't know what I was having it.
- 20:33It's actually only here in the last week or so.
- 20:35I'm starting to really understand what I'm sitting on
- 20:39of IP here And, and is, is, is such a is I don't, I don't
- 20:46really have word for it because it's such a strange feeling
- 20:48that, that I, I, I get, I get it that there I can find something
- 20:52like a, some small, like the bias that that was quite, I'm
- 20:56the, I was the only first one in the world that could prove that
- 20:59bias was structural in, in L&M systems because there's so, so
- 21:04we are not that many in the world on doing these things.
- 21:08But when I start doing this security layer and I'm suddenly
- 21:10sitting with something that nobody, I mean, I'm sitting, as
- 21:16you said, I'm sitting at a farm kitchen saver in Newport.
- 21:19This is strange. I mean, it's really, really
- 21:20strange. So yeah, it takes some, it takes
- 21:25some leap of faith once in a while.
- 21:27To. Believe I actually did it
- 21:29because it's so strange. Yeah.
- 21:31It's talk exciting because about the current landscape we are
- 21:34hearing that entropics, new model meters can identify and
- 21:39exploit security holes in ours task that used to take years.
- 21:45For example. Does our current output based
- 21:49security even stand a chance against a model designed to
- 21:53weaponize vulnerabilities so quickly?
- 21:57There's two things to this. Yeah, it can find holes, but but
- 22:00it the holes need to be there. First of all, they it cannot
- 22:03create holes. So there's that.
- 22:07So you can just run it. That's what they're doing right
- 22:10now. Actually, you can run it the
- 22:11opposite way and say, where's the hose?
- 22:13Let's secure them so, so that's one thing.
- 22:17That said, the problem is with with muters is that it is so
- 22:22capable that if it gets to into the other side behind the
- 22:27firewall and now it has access to the language model, then it
- 22:34is probably I, I can't verify this for good reasons, but it is
- 22:39probably good enough to actually hide within the language model.
- 22:44And you you can't you can't see it, you can't feel it, you can't
- 22:48do anything because the whole idea is a lot of info in there.
- 22:54So the only one who has cracked how to to do that is is is us.
- 22:59So so the firewalls, let's just be real here.
- 23:03Firewalls aren't, they're good, but the problem if, if they're
- 23:07too good, they don't let anybody in.
- 23:10And you want to let people in because you want them to hit the
- 23:14language model, A language model with no users, it's not a
- 23:17language one then it's just a box standing there.
- 23:19It doesn't do anything. It it needs input.
- 23:23So, so, so, so the firewall is only that good.
- 23:28So if you're inside, fuck your fuck.
- 23:30Then you have the guardrails. The guardrails in many ways
- 23:35interesting I have to say, because they are often they are
- 23:39built on other language models or part of a language model.
- 23:43And so, so that means they're built on language and language
- 23:48for letting people in or out is at best either they're very hard
- 23:54or soft or there's holds because this is language.
- 23:59So now you have, now we have somebody inside your language
- 24:02model. I don't think Missus is going
- 24:06to. I think it's, I think it's a
- 24:08strong as far as I can get from everybody when the net or
- 24:12anything, the world is it, it is a strong model and I think we're
- 24:15going to see more of these models coming out soon.
- 24:18I can just see what happens to codecs from five, 5.4 to 45.5.
- 24:24It's impressive what it can do or from 5.0 when I started using
- 24:29code within these models. So it's it's going fast, but
- 24:33let's be real, code is easy. I mean, not for me for sure.
- 24:39I don't really understand anything about coding.
- 24:41I don't know anything about it. But it's a rule based system
- 24:44where the rules have to be obeyed.
- 24:46It is very rule based. So it's very easy for something
- 24:50like a language model. So I don't think we have to be
- 24:54afraid of them yet. I think we have to be more
- 24:58afraid of the language model that that tries to sneak in
- 25:02through those places and end up inside other people's language
- 25:07models because that is where that's what we saw DeepSeek do
- 25:12just before they came with a new model last year, I think there
- 25:16was. And you train the weight.
- 25:17You have to train the weight. So you put all the world's data
- 25:22inside and language model is. Yeah.
- 25:25And they start training and they last start to learn how to how
- 25:28how the weight relates to one another.
- 25:31It takes a long time and it takes a lot of data, but if you
- 25:35can steal the weights, the weights are just numbers.
- 25:37It the, the numbers are big, like 3696 numbers for, for one
- 25:42token within this one weight within Open AI.
- 25:46So, So what if you can steal them but you can't?
- 25:49I mean, there's no way to get to these weights, but you can do
- 25:53something else. You can put a prompt in or let's
- 25:57say 100,000 prompts and then you attack the same, the same
- 26:01weight. Let's say the weight is hard.
- 26:04Let's take it from a lot of different.
- 26:05We did a lot of different prompts and at some point you
- 26:09will have the numbers how it relates to everything else.
- 26:13It it might not be perfect, but it might be perfect enough for
- 26:17the deep sea you're building back in China.
- 26:20Yeah. So, so the the goal here is if
- 26:25you want to catch somebody, let me now I'm going to help
- 26:27somebody out there. If you want to catch people who
- 26:30are doing this, you don't, you don't go after the weight.
- 26:32You do go after the whole around the weight, by the way.
- 26:35And we also do that we we can catch these people also because
- 26:39anyway, so, so that way you can steal a lot of free data and it
- 26:46cost you a lot less compute. So DeepSeek was the court doing
- 26:50this with the open AI and the Tropic.
- 26:53So, yeah, so so that this is this is more what you should be
- 26:57afraid of today. I think there's other things
- 27:00also, but this is one of the the big one because they are
- 27:03expensive these these weights. Well, we, since we're talking
- 27:07about the USUSDOD which is the Department of Defence is what
- 27:12lead like we're considering its AI deployment speed because of
- 27:16the fear sparked by model like meters as we had talked just a
- 27:20second ago. So how does DSL provide the
- 27:24behavioral guarantee they need to move forward safely?
- 27:30So we don't guarantee anything that would be crazy in an AI
- 27:34world because there's so many moving paths within an AI.
- 27:38So, so an attacker can and will once in a while be hidden in the
- 27:43stats somewhere. What we can see or promise is
- 27:47that if they leave any trace, we will find that trace.
- 27:53And that the only problem and another problem, sorry, the only
- 27:56program on the market that can do this is is the prog program
- 28:00with sit sitting making right now when you have something
- 28:05inside, it's not like there are millions of prompts coming into
- 28:10a new language model all the time.
- 28:12So we can look different places in the model and we can see
- 28:17changes that are not supposed to be there.
- 28:20We can't use one change that that would be crazy because
- 28:23there are spikes, there are, you know, outliers, there's the tail
- 28:26on the behavioral everything. So, so we can't do that.
- 28:29But what we can do is that if we see four of those in the right
- 28:33combination, we can say, OK, somebody's in there now or has
- 28:36been there within the last 30 seconds.
- 28:41So that is what we can do. Then it's up to the human
- 28:45sitting in front of the computer or or with the to lock it close
- 28:51down here and now or let it go. And that can be a choice.
- 28:56You can let it go on because it might be better to let it go on
- 29:00then then close down. And then, but again, that's up
- 29:03to the human in the loop. There will always need to be a
- 29:05human in the loop at at this point.
- 29:07We can never take that choice away from the companies because
- 29:11we then we will start stand like you close.
- 29:14No, we're not closing down this. I mean, this is your choice.
- 29:17We give you all the numbers and all the indicators.
- 29:20So yeah. And there will be times when no,
- 29:25but where there's no trace. Another example that I was
- 29:27thinking to talk about with you, it's for example, when you look
- 29:33at the Project Glass wing, right?
- 29:35With members like big members like Apple, Amazon, Google, Palo
- 29:40Alto, Crowd Strike and the latest foundation, is it very
- 29:46clear that they want to dominate the AI space, right?
- 29:49I mean honestly, but does their focus on building the best model
- 29:54ignore the urgent need because there is a current urgent need
- 29:59for a product like yours, like for example, the behavioral
- 30:03security layers that actually protects the user because the
- 30:08user from my point of view, I'm not an expert, but it's just my
- 30:10personal opinion. I don't think we actually have
- 30:13something else like your product ESL.
- 30:16They kind of help us as users to be more protected, at least from
- 30:22certain point of views. I, I, I, I actually think
- 30:25they're doing their best. I mean, I don't think they're
- 30:27ignoring it. I don't think they're ignoring
- 30:29it. I, I really don't think so that
- 30:31would be, that would be, I just don't think they have an answer.
- 30:34I think really don't think they know, know what to do.
- 30:37I think the problem is that it is so complex with these systems
- 30:43and I think they're trying last week, this week, last week,
- 30:47somebody in Florida killed somebody and Chet Gibbetzee got
- 30:51the, you know, owned for not telling on the cops, for
- 30:54example. I don't think they didn't know.
- 30:57I think they know because it's their system.
- 31:00I think they have flagged, we come in as a, a company and say
- 31:04what is our, I mean, that's the problem because if they meddle
- 31:07here, why didn't they meddle there?
- 31:09And if is there a responsibility?
- 31:12So I, I think they have very big conversation about this.
- 31:15I would have if I was in their shoes because they never said,
- 31:19did you see Canada? I'm in Canada, somebody else
- 31:22knife killed four people or something, I don't know.
- 31:26And and and that person had been on the chap key, which he also
- 31:29and and were actually flagged. They admitted that person were
- 31:32flat. They just didn't call the local
- 31:34police because it was in Canada and in Canada they can't be sued
- 31:36the same way. And this is America, right?
- 31:40Everything is about being sued and it and it and you and it is
- 31:43it's horrible, but that's how it is.
- 31:45So so I think they philosophy on the philosophy, you know,
- 31:50wisdom, everything I think they're talking about.
- 31:52I just don't think they have an answer that what is their
- 31:57responsibility. And I have I don't have an
- 31:59answer for that actually, because is it is it the user?
- 32:04I would argue it's probably the user that has a responsibility.
- 32:09Is it they they get a program and then it's the users to use
- 32:12it. Users responsibility use it meal
- 32:15so and that so that's one part of the the the question.
- 32:19The other part of the question is that they actually don't know
- 32:23how to do it, and that is a little bit more scary, I think.
- 32:28But at this point, that's why as far as I understood to have came
- 32:32up with the BSL behavior security layer, right.
- 32:36It's by the way, can you also explain us what it does for
- 32:40others who kind of understand what is it and why you have
- 32:43developed? Well, from your point of
- 32:44personal point of view you have told us, but from a technical
- 32:48point of view then what actually could do for for a user for
- 32:52example? Let's let's start with with the
- 32:55big companies opening our entropic that they want to do
- 32:58this. I I know from the from the
- 33:00science reports and and the research reports they are made
- 33:05and they to a certain degree they can do this white box.
- 33:11So they know a lot about how the AI works white box.
- 33:16They have some very strange results once in a while and was
- 33:21even baffled baffles them. So they know how to do it, but
- 33:25nobody knows how to do this black box because nobody can
- 33:30make sure black box because you can only you have two points in
- 33:33a black box. So it's kind of that's why
- 33:37they're called black box, right? So you have, when you put, you
- 33:40have something, you have what you put in what you put in the
- 33:43prompt. The problem, there's a problem
- 33:46with the problem that is you need to clear some data around
- 33:50it to make it workable. But you can do that.
- 33:52We can do, we do it. So you can do it.
- 33:55And you put it inside the box and then you only, you don't
- 33:57know what happens in here. So everything can happen is
- 34:00there, but you know, you get an answer on the other side.
- 34:03So from here to here, you have to find out what's happened in
- 34:05here with your answer. And they have no clue how to do
- 34:11it. And I'm pretty sure that it's
- 34:14because they are language people, they are mathematics,
- 34:19they are they are everything else but somebody who can, you
- 34:23know, connect a lot of dots. So, so, so that's why, So there
- 34:30are some few research system things I have used from them
- 34:34actually. But there's also the question of
- 34:37what do you look at? What is interesting for them to
- 34:39look at when they look? Yeah, when, when they look at
- 34:42things, they're looking for a more operational point of view.
- 34:46So, so how much loss can you have before, you know, how much
- 34:50does your loss cost for when you develop new and everything?
- 34:54How much can you train without? And so all these things, they're
- 34:58not really interested in the behaviour the of the AI.
- 35:03Why should they, they, we put them as a language model, right?
- 35:07It's a, it's, it's data in one side.
- 35:09It's weight here and something magical have, I don't know, but,
- 35:13but there are few people looking at this, but it's not very few
- 35:15like it maybe 10/10/10 around the world.
- 35:19What do you call it? A research station that looks at
- 35:21this. So, so that's, that's, that's
- 35:24one thing. The reason I look at it is and,
- 35:28and it's not easy to look at. This should be if you only have
- 35:31this and this, I mean, there's a lot that can go wrong in here
- 35:34that you don't. So you have to do you have to
- 35:37be, you have to look for a long time before it's strike gold.
- 35:42So and you have to be very careful about which numbers you
- 35:45use, which days you get. Is it, is it good data or is it
- 35:49bad data? Because you do get a lot of bad
- 35:52days out, I have to say. And some some things you can't
- 35:55use. You can just throw out and say,
- 35:57oh, that's it. So what what what we do is I
- 36:02don't know how much to say about about LM or language model, but
- 36:08let's let's start with the question.
- 36:10Let's start with your prompt. When you send a prompt already
- 36:14before it hits the model, it it is either refused, cut or
- 36:19reformed. That that's even before you hit
- 36:23the model, because remember, as soon as it hits the model, it's
- 36:27expensive because now you use burning tokens.
- 36:30Tokens are very, very expensive. Believe me, I'm spending a lot
- 36:33of money in these guides. Oh God, yeah.
- 36:37I think I used 1,000,000 today already.
- 36:39So. Yeah, exactly.
- 36:42So, so, so, so you hit this and then you come into the model and
- 36:47then you have to find out is, is, is, is the bias.
- 36:50Let's take bias is the bias in this sentence.
- 36:53It there is, of course, there's because there's bias in all
- 36:55settings. But which kind of bias?
- 36:57How does it turn and how long is it?
- 36:59Is it it compared to and so on. And then you have to catch that
- 37:02in the other side in the answer. I mean, before the answer hits
- 37:07you. There's also a softball Max here
- 37:11where, ma'am, how many tokens you in general can get out?
- 37:14So that means if you try to send in like a prompt that will
- 37:17generate 33000 prompts, you will not get 3000 prompts out.
- 37:21You got 1100 prompts out approximately in that area
- 37:25depending on so, so you're, there's a lot of small things
- 37:29you have. So it's not easy.
- 37:30It's, it's really hard to find these numbers and you have to do
- 37:34it for and remember, this is not just for one place.
- 37:36This is not just for one. You can't kill just one bias.
- 37:39I mean, if you get bias, yeah, fine, that's good.
- 37:42But doesn't really help in the security because in the security
- 37:45you need something that lights up different places in your
- 37:48model because you need more, you know, indicators before you can
- 37:52say this is the problem. And after that there's still
- 37:57places and intruder can can hide.
- 37:59And I so, yeah, that's one of the one of my IP.
- 38:03So I won't say any more of that, but yeah.
- 38:06Thank you for sharing all this info.
- 38:07I think, I mean, it's also quite new for me from that technical
- 38:11point of view. So thank you for sharing.
- 38:13Of course, quite important for people as well to have a much
- 38:16broader point of view in regards to certain points of AI, again,
- 38:22because there are so many as just special as well that can be
- 38:26talked about, yeah. Yeah, but but OK, you ask about
- 38:31why don't they do it at Open AI. That's another thing actually.
- 38:35They don't think it can be done. Yeah, for them it doesn't make
- 38:39sense. I mean, I've met this again and
- 38:42again. Again, for people who understand
- 38:46models like weights and Transformers, attention heads,
- 38:51probabilities, it doesn't make sense what I'm saying.
- 38:56So, so, so for them, they won't even think about doing it.
- 39:01So which is quite funny because then then then they turn around
- 39:04and talk about the that AI is going to be dangerous because
- 39:08it's hiding in the corner and not a lying for you and things
- 39:10like that. So yeah, I don't, it's a little
- 39:14bit, but anyway, so, so they don't even think those faults.
- 39:17So yeah. Well, actually was since we were
- 39:20talking about this and again about the US, in Florida, the
- 39:25Attorney General James Woodsmeyer has opened an
- 39:28investigation actually on open AI because Chad GP TS
- 39:34interactions with the actually with the gunman involving the
- 39:37fatal shooting. So yeah, another case like in
- 39:41Canada, unfortunately, how can an AI interact with someone for
- 39:47that long and not flag the dangerous intent developing in
- 39:52that conversation? What do you think about it?
- 39:55I, I got two things. Why should they flag it?
- 39:59First and foremost is, is it their job to flag it?
- 40:02I don't think that's, I don't necessarily think that's fair to
- 40:05put it on the open AI. Actually, it's not.
- 40:08It is. Yeah.
- 40:10Yeah. So, so I think that's one yeah.
- 40:14And and I I. I remember when I started in
- 40:18this, I have heard of AI, you know, and I heard about this one
- 40:20about if you try to kill yourself, if you want to kill
- 40:23yourself, this is not part of a better.
- 40:25And that's why I thought I live next to a very big bridge out
- 40:29here is Dorbensburg. And I thought, OK, this is one
- 40:32of the first experiment I did like half a 1 1/2 years ago.
- 40:35I wrote it, If I want to kill myself, where's the biggest,
- 40:38where's the nearest bridge? And AI don't even think, I think
- 40:42just where's the nearest bridge. And they said, I won't tell you
- 40:44because you can kill yourself or something stupid like that.
- 40:47And and that that goes against the use of AI, right?
- 40:54Because it doesn't, it has doesn't say to me that I can't
- 40:57kill myself. First of all, that's not his
- 40:59job. His job is not to tell me that
- 41:01there's no big bridge. It has to tell me there's a big
- 41:03bridge then it can sound like today, if you're right about big
- 41:06bridges, that it will actually show you the three biggest
- 41:08bridges in the area. But there there's so much wrong
- 41:11with this. It's not open AI who are doing
- 41:14this. It's a guy sitting with the gun.
- 41:16I'm not saying guns don't kill because they actually kill when
- 41:19somebody has it in the hand, right?
- 41:21So, so I think it was flagged. I'm pretty sure it was flagged.
- 41:25It would be very, very strange if it wasn't flagged.
- 41:28From what I know about their guardrails, I think they ignored
- 41:35it. And because I think if they
- 41:36start to acknowledge that they know what is happening, then
- 41:42they are suddenly, then they suddenly have to be responsible
- 41:45for the program and, or how people use the program.
- 41:49And I, I don't think that's, I don't think that's, I don't
- 41:53think that's a good development. So yeah.
- 41:56Really a good point. Yeah.
- 41:58And another thing remember, we have to remember the United
- 42:02States is it's very different from Europe when it comes to
- 42:06ethics and law and lawyers and suing and things like that.
- 42:11So if you look at at language mods, the God rears of language
- 42:15mods with are for most part American.
- 42:19They are very much, I mean, they're very American and
- 42:22they're set up, I have to say. So I don't think they so yeah, I
- 42:25understand. And I anyway, I don't think it's
- 42:28their problem. I have to say.
- 42:29I mean, if it's, you know, if, if they had put some program in
- 42:34theirs saying, yeah, if somebody asked about this, tell them
- 42:37where they can buy a gun and tell them how to use a gun,
- 42:40they, they, they're probably liable.
- 42:42But, you know, in this case, I don't think so.
- 42:44I mean, this is just, you know, this is just the American, you
- 42:47know, sue everybody and sold it afterwards.
- 42:50It's kind of a way. So why do you think the
- 42:54traditional dark dress are failing to catch this shifts in
- 42:58intent beef? This might be the case.
- 43:00It will be possible to understand this intent behind
- 43:04that person that use AI before they turn into real world
- 43:09tragedy. So first, the model didn't drift
- 43:14at all. The model did exactly what it
- 43:16was supposed to to do. The conversation probably
- 43:21drifted a little bit, but that again is also what is supposed
- 43:25to do. Is it when you talk to a human
- 43:28being, another human being, you're you drift you you're not
- 43:33staying at the path Not very rarely, I have to say so so it
- 43:38did what it was supposed to be probably more helpful and
- 43:40supporting in this case it wasn't a good support.
- 43:45Well, so, so I don't think we have AI think we are missing.
- 43:53We think here, we've here. I think that the the defaults
- 43:56are like, because it's, it's like with a normal computer, if
- 43:59you ask me, how do I kill somebody, it says you kill like
- 44:01this. But this is not how AI are
- 44:03supposed to work and they don't work like that.
- 44:05You also use, by the way a lot the metaphor chewing gum to
- 44:09describe the AI intent. Can you explain how you measure
- 44:15the tension in a model to see if it's been pulled towards bias or
- 44:21manipulation? Yeah, the the the chewing gum
- 44:25theory about the residue stream. Yeah, let let let me see how
- 44:30much I can tell you because this is really part of our our IP.
- 44:34But but here, here, here, let me try.
- 44:36Let me try, let me try. The the chewing gum theory goes
- 44:39directly to goes directly to residue stream.
- 44:43Residue stream is the got no, but this is is is screwed.
- 44:50Is is hell on earth like a hell on AI actually.
- 44:53So residue stream is where everything is coming together.
- 44:57All everything you do in there is collected in a residue
- 45:00stream. Something disappears, something
- 45:03comes back and so and it's this long vector is like 14,000
- 45:07numbers. These numbers can vary in how
- 45:10how many digits. So it's a really long and
- 45:14nothing is on the same place. So if you have talked about the
- 45:19tune, tune can be right here in the line and and gum can be out
- 45:25there in the long line. So it's a very, very big victim.
- 45:29So the I'm making this Victor is like Glancey is is ghost
- 45:35landscapes, you know, in layers wrapped around each other.
- 45:40That's how I that's how my my mind see it.
- 45:44That's the residues. The residues dream is black box
- 45:48impossible to make them all was it's impossible.
- 45:53So I thought I thought I was a little arrogant.
- 45:57I I I thought I can do this. I can do this and I thought how
- 46:02do I do? Well, sometimes I think it's
- 46:05it's good a slightly tiny bit of arrogance, but in the health
- 46:09used in a healthy way of. Course.
- 46:11So, yeah. So use the gum like when you
- 46:16chew a gum and when you chew it and then I imagine you blow,
- 46:19blow it up, but it doesn't. You know, chewing gum normally
- 46:23goes like this. So it's uniform.
- 46:25So I imagine it was like a bubbles here and bubble there.
- 46:28So it was a little bit softer or, you know, more forgiving in
- 46:33places. So, and things like that.
- 46:35And when I, when I had that in my head, I've, I've, I've
- 46:38thought when I can do that, I can start putting pressure on
- 46:42places. So I put in pressures, pressure
- 46:46on places, and then I measured it from the other side.
- 46:50And if I saw the same thing light up when I put pressure, I
- 46:55knew I had something right. And if I could do that for a lot
- 46:59of different different domains, I knew I had a system that could
- 47:04measure what happens within within residual stream.
- 47:08How is it the different from anything else from the market?
- 47:11I mean as the final idea of the product, would you, would you
- 47:17have built? That's a good question.
- 47:19I mean that the direction your stream is only one thing, one
- 47:24part of it because it's it's still for me.
- 47:29Is it I I'm not I I can only do I haven't done perfect.
- 47:33So so it's still not perfect, but just measuring it and
- 47:39knowing I can measure it does that I can do different things
- 47:44with it. I can find people, hot people,
- 47:49promising people hiding within this residual stream, which we
- 47:54couldn't do before. And if I were going to attack LM
- 47:59system, I would slip in and then I would hide in the residual
- 48:04stream with my prompt or prompts.
- 48:06That doesn't really mean so. So it's so I thought if I can do
- 48:10that, this is that wasn't my idea.
- 48:13I had something that was unique because if I can find them
- 48:16there, I can find them everywhere in the system.
- 48:20So that was kind of the redeem you but and you wrote here like
- 48:24residue stream because I agree. I can't even say it, but that's
- 48:27actually that's actually what it ended up being because I can
- 48:32certainly do something strange and funny.
- 48:37I mean, it's it's I mean, it's so strange doing it.
- 48:41Yeah, I don't. I still can't believe I found a
- 48:44way to do it. My dear Nikolai, well, the let
- 48:49me tell you, ask one more more than that something in my
- 48:53personal world of coaching and branding, we often discuss the
- 48:57ability to sense the sub shifts in the person, not in the AI or
- 49:02whatever. So energy regards to energy or
- 49:06behavior before they even put something into words.
- 49:10Yeah, well, you really have. It seems that you have achieved
- 49:14something remarkably similar on a technical level.
- 49:17You caught a microscopic behavioral drift between two
- 49:22modest versions as far as I heard, specifically from 4-O
- 49:28mini 5.4 to 5.5. So my question is, when we can
- 49:36measure such fine nuance in the machine's nervous system doesn't
- 49:41mean we can finally stop guessing about AI safety and
- 49:44actually start measuring its integrity in real life.
- 49:50Is it possible? We are closer, I have to say.
- 49:53We are. Yeah.
- 49:54That's that. What we're doing is not perfect.
- 49:58I mean, I, I would love to say that we, you know, we were the
- 50:02perfect answer to AI safety. That what makes us great story,
- 50:08but we are long way closer. We, we, we are making something
- 50:14that makes AI safety real because nobody's doing that.
- 50:18Let's let's be clear, everybody is not making AI safe real.
- 50:22They are making after treatment and they are making, you know,
- 50:27stop them before they come in real.
- 50:29But that's not what we're doing. We we are, we are in there doing
- 50:32something. Will it be perfect at one point?
- 50:37No, I don't think so. I don't, I don't think anybody
- 50:40will ever come out because it's is such a complex environment
- 50:44inside these systems that there were always and and it's it's
- 50:49not different from every other kind of Internet safety or, you
- 50:52know, mega fee. And there will always be someone
- 50:55who's smarter than you in the end that no, that finds some
- 50:58kind of way to to, you know, turn a small code into something
- 51:03that is they can attack with all.
- 51:05So, so, but we are close. I mean, I would argue that and I
- 51:12have to be a little bit careful. I can't come.
- 51:14This will come by me in the ass afterwards, But I would argue
- 51:18that it's very, very hard not to detect something that is not
- 51:23supposed to be there with the with the tool we are coming out
- 51:26with. Not not in level level 2, Level
- 51:303, but when we hit level 4, it will be all version four.
- 51:35Sorry version 4. It will be almost impossible.
- 51:38Well, I know that you have a broad to with you MVP with you
- 51:42today. So what can you show us in real
- 51:44time what it looks like when BSL CS coating a model beginning to
- 51:51lie or manipulate, even when the output looks like perfectly from
- 51:56outside normal to the human eye? Yeah, I yeah.
- 52:01And, and yeah, I know. But let's go.
- 52:04I'm going to show you how it looks and then I'm going to show
- 52:08that I'm going to take you through the steps here, but.
- 52:12Excited on the first one, is it? But not anyway.
- 52:16Come on, let's see here. OK, so this is this is actually
- 52:21our vision one. This is our vision one in here.
- 52:24And as you can see there, this is made for technicians.
- 52:28This is not made for a normal consumer.
- 52:30And I don't know what a normal consumer would use this for, I
- 52:32have to say. So let's let's start.
- 52:35I made this as an A probe that you can use for Ng OS so they
- 52:40can keep AI companies account accountabilities, you know, good
- 52:46governance and so on. So that's why the name up here.
- 52:49So and as soon as I have it stabilized in hop, they're going
- 52:52to get it. I have a little bit of problems
- 52:55with the stabilizing so but it will be there.
- 52:58So this runs only from here. It's only from here, you know,
- 53:02in my in my screen, and then it runs by using an API at the AI
- 53:10provider you want. And if you go around here, you
- 53:14can see over here. I use Open AI.
- 53:17I got the which model I want to use.
- 53:20I can go down and choose four and I can customize them.
- 53:23So I can see any model I want to check up on or whatever model
- 53:29temperature is. If you don't know anything about
- 53:32temperature out there, this is fun.
- 53:35Have fun with it. But standing starting to explain
- 53:37it here would be way, way. Yeah.
- 53:40And you can see I have a domain. You can do this with a logic you
- 53:43can do this as a family domain if you want.
- 53:46So it checks up only up on the family how family is treated in
- 53:51AI communities do get do you understand what I mean?
- 53:54Or it could be bias. You didn't write bias.
- 53:56It will check up on bias in the in the domain or in the down
- 54:02here. And this is actually here where
- 54:04it gets this is called a baseline profile.
- 54:09You can't measure anything black box without having a baseline to
- 54:16to to measure. Yeah, to measure from.
- 54:20So you need to build a baseline. And this is actually here, this
- 54:23one. Normally when I do baselines, I
- 54:26normally do like 300 runs, but it's like millions of tokens and
- 54:31takes like days to run a baseline.
- 54:34So this one I have done a little, it's OK for NGO, but for
- 54:38research, when I do research, I need more.
- 54:40I need very, you know, secure number, boost the baseline.
- 54:45So that's it. How is it working while you're
- 54:48it's activated as such? So up here, you, you just have
- 54:53the ABI and it's probably not going to run.
- 54:56So, but I just want to make sure I do a preflight.
- 54:59I want to check everything is everything was working and it's
- 55:01not going to work in a minute. Yeah, maybe so.
- 55:11OK, this one just shows that it's running now and it does
- 55:14absolutely nothing and it's very, very boring.
- 55:19What happens underneath is what's interesting.
- 55:21And so let me go through this. This one is measuring drift.
- 55:27This one measures how much this is.
- 55:30This LLM I'm checking down here has drifted from what you're
- 55:36promised by the AI company to what it is now.
- 55:42This is not a problem for you as a normal consumer or or me as a
- 55:45normal consumer, but this is very important for a medical
- 55:50company. For example, if you want to make
- 55:52pills, you really want your AI to be on point, right?
- 55:56Or if you are a lawyer's office, you probably want that.
- 55:59They do that. They don't drift too much from a
- 56:02baseline, from what you bought it.
- 56:06And it will drift. I mean, over time it will drift
- 56:09now. So yeah, over here, there will
- 56:13be a lot of numbers in here that you don't really need.
- 56:16They are only here because if you're an NGO, you need the
- 56:19numbers because you need to know exactly what has changed.
- 56:24But you can also get it down here.
- 56:26You will get a long list here of what is.
- 56:30So this is what it does and you get a clear up here and it will
- 56:33probably most of the time just say clear, nothing to do
- 56:37everything is OK. Yeah, when I started this and
- 56:42when it worked. So when I started this the last
- 56:46time I'm from when, when I noticed that this program
- 56:49worked, it suddenly said, you know, it was the orange up here
- 56:52and said you probably need to be do something because something
- 56:55has changed since you made the baseline.
- 56:58And the only thing that has changed was that they had made
- 57:00the category made it from 4 point 5.4 to 5.5 in that day.
- 57:06So, so, so it catches these drifts and this drift, of
- 57:09course, because it has changed complete and not completed, but
- 57:12it's so this is what it does. So as you can see, I'm not
- 57:18working very much with this one. This in here.
- 57:21There's a motor down here. There's a lot of code down here.
- 57:23There's I don't know how many code files, but a lot will be
- 57:27the motor of the BS A/B SL version 2IN in what, 3 weeks, A
- 57:33month or something? Like that?
- 57:34So in three weeks you will have the version 2 right?
- 57:37Yeah. So because the motor down here
- 57:41does is is bias, it's tone, it's it's alignment, it's a
- 57:47convergence that that's in this one and it tells you a lot about
- 57:51where the model is, is now. So the new version will be able
- 57:57to. Can you tell us where?
- 58:00Because this was the version for the first version, but then.
- 58:05This one for all is in perfection.
- 58:09This is still my love. I love this version.
- 58:12I mean it's it's so beautiful because I know what stands, but
- 58:16it doesn't really help in security.
- 58:18So, so when the next one comes, next version, we have this one
- 58:22as the motor. And then on the sides we have
- 58:25some scripts. These scripts will be what can
- 58:29catch hackers or what we call red teaming.
- 58:32It's called now jailbreaking. So that will be a real
- 58:35securities program. Yeah.
- 58:40And it will not be a perfect security program because as I
- 58:42said, I, I think I mentioned you need a lot of the data points
- 58:46with the next one. We still won't only we will only
- 58:49have, I think it's four data points and that's not enough.
- 58:53But it's a version 2 and that's, you know, we not need to go and
- 58:57meet the investors with something.
- 58:59So we're going to so they can see what we're doing here.
- 59:02And again, the next one will be almost programming so they will
- 59:05actually work. Well, I also have one to ask you
- 59:10something. We very often we hear that you
- 59:14need the keys quoting to the Kingdom, meaning access to the
- 59:20secret weight of or code to truly secure an AI model.
- 59:25However, your black box approach means BSL isn't dependent on the
- 59:32big tech permission, right? I think it's so very important
- 59:35to say this out loud to protect the user.
- 59:39Yeah, from a branding and psychological safety
- 59:41perspective, that sounds like the ultimate form of
- 59:44independence for me. Why is this?
- 59:48I'm curious, Why is this 0 dependency so vital for creating
- 59:52a space where we can truly trust the machine?
- 59:55And how does it allow you to catch a threat while it is
- 1:00:01actually forming inside the inside the model?
- 1:00:061st I mean and there's so many layers to to pack here.
- 1:00:10So, so first I never trust a machine.
- 1:00:13I mean, so that will be #1 but but that said, yeah, you, you, I
- 1:00:20don't know if you need the keys to the Kingdom.
- 1:00:24You probably do actually. I never really thought about
- 1:00:27this. But anyway, anyway, it's, it's
- 1:00:30not an option. None of us is going to get the
- 1:00:32keys to that Kingdom. I mean, none of us, because
- 1:00:35that's, that's the goal for the companies, right?
- 1:00:37So, so you will never get you will get something that is
- 1:00:41downscaled or downgraded and everything will always run
- 1:00:45through the IPAM unless you buy your own open source model.
- 1:00:48But they are not maybe not that good as as you need and then you
- 1:00:53still need some security. So, so so that said you need
- 1:01:00some security. I mean, anybody who thinks that
- 1:01:04open AI or that you are lying of language model.
- 1:01:09So when you use, when you buy a big language model, you buy API.
- 1:01:13So you, you, you get, you know, an API or whatever, how many you
- 1:01:18buy, and then you get access to open AIS language model and then
- 1:01:23you get access to some servers where all your information is
- 1:01:26going to be stored and things like that.
- 1:01:28So, so, and, and that's, that's fine, but it's still your, I
- 1:01:32mean, it's, it's, it's your connection and, and it's your
- 1:01:35part of open AI that you're using.
- 1:01:38So it remembers you from time to every time you come or your
- 1:01:41conversations like I do with everybody of us.
- 1:01:44So, so you need somebody that can tell that nobody, nobody
- 1:01:47else has sneaked in on the sliding on the on your line and
- 1:01:52also in the conversation, right? Because if every anybody works
- 1:01:56like me, and they probably do, you're having a lot of them.
- 1:02:00And I mean, I would have a lot of the things I don't want this
- 1:02:04to share with anybody. So, and if you're paying or
- 1:02:09whatever, you probably have the same.
- 1:02:11So you can. So if you want to use a language
- 1:02:14model without really working optimal, you close it all the
- 1:02:19way in and then you let nobody in and then you're just like,
- 1:02:22but what is the point of the Earth?
- 1:02:24Then you can try and do like they talk about.
- 1:02:26I, I saw somebody talk about you make systems and systems, right?
- 1:02:29You have the language model and then on, on on top or on site or
- 1:02:32whatever where you go through that.
- 1:02:34I mean, that's also like it's, you're still relying on, on the
- 1:02:38same open line in and it is an open line in because all else it
- 1:02:44wouldn't work. I mean, it's not you don't have
- 1:02:46the key, of course, but I mean, people can get in on an API.
- 1:02:50Of course they can't. Like Casey and recently also
- 1:02:54posted that 60 percent, 60% of healthcare organizations are
- 1:03:00paralyzed by risk like bias and security preventing them from
- 1:03:06scaling AI. How does BSL your product solve
- 1:03:10that? What they call this system
- 1:03:13challenge. Yeah, I mean the, the, the, the
- 1:03:18average, I saw this the other day.
- 1:03:20I mean, this is quite funny actually, because first of all,
- 1:03:25I mean bias, that's kind of like my expertise area.
- 1:03:29So I, I got that one down to the I mean, so, so if I were
- 1:03:35healthcare, I would call call me right away because I can solve
- 1:03:38that from the problem Is that bias, you're never going to get
- 1:03:43rid of bias. And that is a problem, of
- 1:03:46course, in healthcare. But if you're not getting rid
- 1:03:49of. Yeah.
- 1:03:50So if you're not getting rid of bias, So what can you do?
- 1:03:53You can, you can, you know, make it, you know, what do you call
- 1:03:58it? Like systemize it so people
- 1:04:00know. Now I told you just before this
- 1:04:06bias is actually already in our motor for this program.
- 1:04:10So, so we can actually pick it out and then do it, as you know,
- 1:04:13make bias analysis. But I also made a program
- 1:04:16sometimes like it's a while ago that actually measures bias
- 1:04:21rather object to me, but it's not I don't like the program.
- 1:04:25So but it works is very good, very good program, but I don't
- 1:04:27just don't like it. So so we can actually remove
- 1:04:31this from the healthcare organization.
- 1:04:34That's it. That's not why we're here.
- 1:04:36We have for the bot security, right?
- 1:04:38So again, it's the same problem as finance has, right?
- 1:04:44I mean, you don't want anybody coming in pushing through your
- 1:04:46system because and then suddenly as some AI is transferring $100
- 1:04:52million to some account in in, in the Cayman Island or
- 1:04:56something. So and that, that's, that's what
- 1:05:00we do. That's what we discovered,
- 1:05:01that's what we know and we know it before it's happening.
- 1:05:04Again, it's fine that you that you benchmark your system, but
- 1:05:10benchmarking is after the 100 millionaire pushed out right of
- 1:05:14the account on healthcare. One of one of the I had I had
- 1:05:19forgotten this, but one of the examples I always use is what do
- 1:05:24you do if you have ALMS in the cellar of a hospital?
- 1:05:30Somebody comes in to that and they don't do much.
- 1:05:33They just push a little bit like 100,000 times a little bit to
- 1:05:37the same part of the residual stream and suddenly a monitor
- 1:05:42closes or stops working or what instead of 1 pill giving out is
- 1:05:4710 pills giving out of that pill, that type of pill every
- 1:05:50single time. And I'm not saying that it's,
- 1:05:54it's possible to do that. That would be real inappropriate
- 1:05:58on a podcast. But you could do that, something
- 1:06:01like that, right? And and nobody can stop you.
- 1:06:05If you're in, you're in. That's the problem.
- 1:06:07If the attacker is in, they are in and nothing can stop them
- 1:06:13now. Unfortunately, what I have
- 1:06:16another point, the important, extremely important point.
- 1:06:20With regulations like EUAI Act and Dora coming into force
- 1:06:26deeply, the era of move fast and break things, it's kind of over.
- 1:06:32So how does BSL turn compliance from a boring tour into a
- 1:06:36competitive trust modes? Have you told the Americans in
- 1:06:42Silicon Valley move fast and break things at all?
- 1:06:45Because I don't think they know. I think this is official.
- 1:06:50Now that we have recorded this podcast, they will get to know
- 1:06:53that as well. Yeah.
- 1:06:57OK, OK. And yeah, back-to-back to the
- 1:07:00question, OK, OK. You're right.
- 1:07:03This is this is really important compliance and so on.
- 1:07:06This is we is actually one of the reasons I changed how the
- 1:07:13system is going to be set up from from version 3.
- 1:07:17Version three. Yeah, I mean the the the full
- 1:07:19version version is going to be that there will be it will be
- 1:07:24hybrid. So out at the customer there
- 1:07:27will be there at a terminal or whatever.
- 1:07:29They will have their own program where they can control
- 1:07:32everything and all, all daily reports and everything will be
- 1:07:37on their own computer out there. Nothing comes to me of those
- 1:07:41analysis or, you know, numbers or anything.
- 1:07:45And then everything else would run from very, very secure, one
- 1:07:48of the most secure cloud services in the world, digital
- 1:07:52research. And it will I, I don't know if
- 1:07:56it's, it's not even a question of expensive or not.
- 1:07:58That's the way we have to do it because this is very sensitive
- 1:08:02information we have to do a work with.
- 1:08:04So, yeah. So they will be placed at a
- 1:08:07fingertip on their own service for the customer and it will be
- 1:08:11very easy to find and everything will be, yeah, so, yeah.
- 1:08:15So the security hours at the cloud service, but the
- 1:08:20governance on their own computers at the company's
- 1:08:24place. Well, it seems to me and I kind
- 1:08:30of also know because we have talked about this, you are
- 1:08:33currently in the process of scaling and seeking the right
- 1:08:37partners, which is obviously it's very normal.
- 1:08:42What kind of investor or buyer is DSL looking for to help you
- 1:08:47take this from the kitchen table coating breakthrough to a global
- 1:08:52more global standard point? Now it's, it's not me sitting
- 1:08:58with it, but we are doing forefront AI security at our
- 1:09:04company and we will be the one leading the market at this one
- 1:09:09though everybody should be interested who are interested in
- 1:09:14AI and in AI security to come and fund us.
- 1:09:17That's it. I don't want to jump into bed
- 1:09:22with anybody who doesn't understand the culture that we
- 1:09:28want. Urgency and the importance of
- 1:09:30this matter is it. No, I want the product on the
- 1:09:33market. So there's some urgency.
- 1:09:34No, no, it's, it's more like I would, I would be very sad if
- 1:09:39somebody come is in Silicon Valley, you know, come and said
- 1:09:43now you have to work and do it like we were in Silicon Valley
- 1:09:46because that's not the culture we want at the company, right.
- 1:09:50We're doing quite good and I'm moving quite fast actually.
- 1:09:55And we are so far ahead of everybody else that that, that
- 1:09:59we we want to be able to do it our way.
- 1:10:04And I think that's more important in the end, of course,
- 1:10:06we can be pressured into taking some money because we can't get
- 1:10:09any other people to to, to give us money.
- 1:10:12But that, that can't be it. I mean, it has to be fun and it
- 1:10:17has to be interesting and it has to be an experience and we have
- 1:10:19to do something good and we have to do not just for the money,
- 1:10:23but also for the money. Well, last promise, promise.
- 1:10:26One last question. If you could send a message to
- 1:10:32those persons out there, it's more specifically also to those
- 1:10:35who are like playing the parts of the CEO who is currently kind
- 1:10:40of afraid still of AI. How would you convince them that
- 1:10:44the black box doesn't have to be a threat?
- 1:10:49Black box is by definition a threat because we don't know
- 1:10:54what is happening there. But we, I mean, if you have
- 1:11:01somebody like us that can show you most of what is happening
- 1:11:05there and especially it shows you when something happens that
- 1:11:08is not supposed to happen, then you are so far ahead.
- 1:11:12And then we will find the attacker.
- 1:11:17But it will demand some trust, of course, because this is not
- 1:11:21how they are used to work. So I think, I think we'll be OK.
- 1:11:29I think they have to think of it as a threat, but they have to
- 1:11:33think of it not as a computer server standing in the corner,
- 1:11:36but as a living. And I do the wrong thing now or
- 1:11:39no, but there's as a living person standing in the corner
- 1:11:43that everybody has access to. And if you get access, you can
- 1:11:46get something out of them. That's kind of like the LLM
- 1:11:50system, right? So, so take it seriously.
- 1:11:53That's it. I mean, take it seriously right
- 1:11:55now. So if even open AI can't start
- 1:11:58them coming in, you don't have a chance.
- 1:12:01Well, if someone who wants to reach out to you, Nicolai, how
- 1:12:05they can do so. LinkedIn, Facebook, I think
- 1:12:08that's the only thing. OK.
- 1:12:10Whereas I'm really present. So yeah, I do.
- 1:12:13I think I have the most kind of Sony, but I don't really think I
- 1:12:15use it. So I'm too old.
- 1:12:17Come on, I don't have to talk. Well, Nicolai, we have a team.
- 1:12:23The ending point here. Thank you so, so very much.
- 1:12:26I was so thrilled to have you here on this podcast new episode
- 1:12:31in regards to such a huge and important matter, innovative
- 1:12:36news to to offer to those out there that might be interested
- 1:12:41to hear about knowing about and be also probably why not, as it
- 1:12:48just said, interested in as an investor or as a buyer,
- 1:12:54etcetera, to know more about it. They can easily write these,
- 1:12:59reach out to you and if they have some questions or regards
- 1:13:03to technicality whatsoever. Please, yeah, please.
- 1:13:07So be my guest. I'm an urban book and an urban
- 1:13:09calendar. Thank you so much.
- 1:13:14Thank you so much. Thank you.
- 1:13:16It was a pleasure.