AI and Critical Infrastructure – The Inevitability Curve EP2

This is Textron tv. Hello, my name is Chris Blas, and I'm your host today for this episode of The Inevitability Curve, where we look, take a topic, we look back in history to see how we got to where we are right now. We take a look at that and we look forward.

Uh, joining me today is a good friend and colleague, Tim Roxy. Um, Tim, if you don't know who Tim is, look up the name. It's a fascinating ride.

Uh, keep your hat, uh, on tight for it. Uh, Tim and I met, uh, in the cybersecurity realm, uh, when we are, I think, working at Nerf, the North American Electrical Liability Council, and gave the way for cybersecurity be even power red, uh, around the wall sent. Um, but as we're discussing in the green room before this, the topic of AI and infrastructure, we have fun, g so I'll let you choose exactly where we start.

Um, but where have we been in this space and how do you think you got here? Um, yeah. So get into the, to get into the space of artificial intelligence, you have to, to think about what, what the alternative to artificial intelligence is.

Of course, it's real intelligence. That's you and me and others. On good days, we're okay, smart monkeys, and on bad days, we're not so good.

We do bad things. But, um, here we are humans. And, and then in one of my career paths as a nuclear engineer, that's, that's my engineering degrees are both in nuclear engineering.

Um, we study a lot of, um, pretty arcane things, uh, mathematics, symbolic logic, blah, blah, blah, bunch of stuff, boring stuff. But the fundamental reality is that you have these nuclear power plants, uh, all over everywhere. A bunch of them used to be, uh, 104 'em in the United States on 64 sites.

So how you control nuclear power plants is very important. And the industrial control systems for these nuclear power plants, uh, fall into two or three broad categories. But what I focus on the most was nuclear safety or safety related category.

Uh, those are the ones that are really greatly important, but that's just, you know, industrial control systems. It's in all 16 critical infrastructures. It's everywhere, but where have we been in there is really still nice of fair scan rugs, but done very well, really nice dog knives and very cool drugs.

We have a, a safety analysis report for a nuclear power plant. And that safety analysis report has a collection of design base. Uh, pumps will be so big, uh, pressure vessel will be so thick.

Containment building will be so large, it will be able to do this. And based on the general design criteria, these design bases turns to engineering specifications, engineering specifications to build specifications and nuclear power. Plant code operating.

The nuclear though have to stay within certain parameters. Lots of ground real intelligence. Those systems power land down into system structures and components lumped on clumped into functions.

So, um, there are primarily fission barriers. The, the fuel pellet itself, the cylindrical tube that it sits in the, the fuel assembly, the reactor pressure vessel in the containment building. All of the nuclear power plant regulation is around how to prevent the nuclear, uh, reactor materials, the contaminated materials or the radiologic materials from escaping into the general public.

Industrial control systems control almost all of those processes. Over time. We've gotten extremely good at the industrial control systems around, gosh, I think 1980 ish.

So where have we've been 1980 ish. Um, what we would con consider today, uh, machine learning started to, to rear its head, which was kind of interesting because we had algorithms that would control flow rates and all kinds of stuff. So software stuff.

I, I grew up on writing software for different control systems that software would control the operations of, or the design of parts of the nuclear power plant. As time progressed, we had industrial control systems. The boxes, the widgets became powerful enough that you could start putting a lot of extra software into those.

So we wrote software for those boxes. All of a sudden they became very intelligent, intelligent field devices. Then along comes artificial intelligence, and this is in, um, the eighties.

It was that, that was, uh, some of the terms we, um, inference engines and rules base. So rules base, what, what is the basis of operating a nuclear power plant? Um, and then an inference engine, which can be considered to be something like a fgy fuzzy logic module, which is going to process an operation into a smaller set of, uh, tokens, which can be understood by the endpoint machines and, and a react and a and a, a physical manipulation of a device would occur.

As a result of all of this, these became quite good. Then in a completely unrelated, unrelated field, um, there was some experiments done, I think in the sixties and late and early seventies where a series of circuits were, were created. I think 40 was the magic number in one of the books I was reading, where rather than program, rather than program a computer system to give you a result based on an input, like recognize the letter E rather than try to do that, you instead, um, create a neural network that you expose a letter E two, and then you adjust some bias controls inside of this neural network mesh until you get a plus one or a, or a minus one.

And, and you teach this group of 40 circuits how to recognize the letter E in different orientations. Kind of cool. You have an emerged behavior off of a, a collection of hardware that we'd spent literally decades trying to write software around very smart, intelligent software that would be able to do this recognition pattern.

Whereas if you do it this other approach, there's no software programming per se. What we've now done is programmed those little circuits, 60 to a hundred lines of code and collected them in certain ways, and which you can now expose a letter E to a set of optical scanners on one end, and you will get a digital output of a letter e no matter which way you turn it. That's, uh, voice recognition, that's speech recognition, that's, uh, text recognition.

Those capcha that we've been trying to correct for years and years and years. Are you a human? Um, those are actually this huge archive of letters, um, and or words that have confirmed human intelligence, that has confirmed that that is indeed bicycle, or that there are so many bicycles in the walkway or whatever, whatever.

Anyways, those huge databases now go into training these artificial intelligence engines, which are now being applied to industrial control systems. And the interesting part to me is, uh, we are way above our skis, so to speak, because back to the basics again, you have a design basis of nuclear power plant. You have a design-based accident.

What's the worst that can go on? And then you have a way to protect the public from that design-based accident that's all written inside of the final safety analysis report. The reactor is built and operated so that that protection mechanism, there's always front and center.

You're always gonna protect the health and safety of the public. Fast forward now to a very deterministic approach. We know what the design basis was.

We postulate what a design-based accident is. We assess the hell out of a design-based accident, and we create a scenario in which we can protect the public very deterministic, boom, boom, boom, boom, boom. The equations, although complex, are solvable for most cases.

So it's, it's a doable deterministic thing. However, when you have machine learning, you can mimic that process very, very well. And some of the coefficients of the algorithms of that machine learning program can be adjusted every time you execute that, that cycle.

So every time an industrial control system loop goes through, you can make a final, a tiny adjustment to those coefficients in the algorithm. By doing so, machine, machine learning, uh, control systems can actually learn to adapt to the environment that's different than a generative AI generator. AI gets weird.

That's where you're going is taking some of these industrial control systems that have very deterministic things, and you're gonna bridge that gap across an emergent behavior, and you're gonna look into the generative form, which is gonna sound nice, but may have some issues like biases like hallucination. The things we've been seeing in the AI world, the question that was put to me several years back now is, where would you put a bright line criteria in the industrial control system for safety related industrial control systems like you find in nuclear power plants? Where would you put a bright line that says that you're allowed, you should be in incentivized to put in really elegant machine learning solutions because they're very fast and very strong, very, uh, very deterministic at some level.

And where would you be able to use a generative component like we've seen in chatbots and, and stuff? Where would a generative component be used? Remembering that the generative component that emerged behavior brings with it all of these weird behavior problems like hallucination just completely out of out of scope stuff.

So how do you apply the indeterminate characteristics of artificial intelligence generative to a far more deterministic, um, set of controls for industrial control space that is out by free. Right. You know, and this is a, as I thought about this one before we started recording here, I, I think it's an argument to make that we've been building artificial intelligence into infrastructure four thousands of years, right?

You know, you take the earliest, you know, water systems, you know, you find out that if this water goes beyond a certain level, that's a bad thing. You put a notch in the rim, we have an overflow valve, you know, so we're replacing what a human would've done. I would, if I'm sitting there watching the water level up, I am, you know, draining it.

I'm pulling out the log, I'll just build that in. Mm-hmm. And he can follow that arc up to, you know, the industrial revolution.

I'm always fascinated with, uh, uh, Joe Screwdriver's retro tech time machine on Ro Roku, which plays all these, you know, training films and, and documentaries of industrial processes since the early 19 hundreds. And things like World War ii, battleship artillery calculators, which are these, these analog machines. You rotate this nub and this gear turns and, and the math is worked out.

Mm-hmm. Which is taking the intelligence a person would've, would've used and just automate, you know, making it artificial. And I think you led up very well to where we are right now with what we're calling quote unquote artificial intelligence.

Right? And there's this determin, Right? And, and I, and I find fascinating as well, that as you know, you know, you and I did some work in South America, uh, a number of years ago, and just before I went down there with you, I've been working with Columbia, which is public record.

I made a 25 year national, uh, uh, infrastructure plan. And the seven and 15 year marks, you know, I felt like worthwhile calling out certain things. And this is really effectively one of them, you know, artificial intelligence information to, uh, integrity, uh, wins War town.

And I, last week on the, on the show, we're talking about where we are with, you know, uh, with, with, uh, information warfare, meta, you know, what he's determining is metamor, and it's the same sort of emergent property issues, you know, where where we have not, through all the things we can look at through human history, we haven't really actually encountered this one yet. So where we are right now, you know, in, uh, in this, um, what do you think, what do you think the, the main, uh, misunderstandings, misconceptions if you'd like, or, or the opposite of that, you know, what are we getting right, right now in thinking about this, that bright line? Or are, are we clear on how far we can go with generative AI as a managing critical systems as opposed to deterministic?

Well, oddly, oddly enough, right now we're still at the, the very, very earliest days. I mean, artificial, uh, artificial intelligence is a, is an umbrella term. That means a lot of things.

Um, one of the bigger components of it is machine learning. Machine learning is actually a subordinate component of, uh, generative ai, uh, as well. So open, open AI and others generative forms.

Um, they have these sub-components, uh, within them who do other parts. And there are, as you said, uh, decades of experience with machine learning kinds of things, like a notch and a pipe that gives you a certain level. And beyond that, it, it, you know, your toilet has one in the tank, right?

A ball valve. Um, but the things that can be done with where we are today is imagine if you will, uh, you're the system engineer for a turbine control system, and the analog computer that you use is an electro hydraulic computer. So electro some type of electrical components and hydraulic some kind of fluid.

So to operate an electrohydraulic control system, it's a lot of work. You know, you've got electronics, you've got electric, and you've got components like the viscosity of an oil, uh, uh, you know, specific gravity of the oil, the purity of the oil, blah, blah, whatever, whatever the working fluid is. And then someone comes up to you and says, Hey, I can do this whizzbang thing here with these flashing lights, right?

And, and you go, wow, that's pretty cool. So in, in the world of, of engineering, we would look at the parameters, the basic engineering, uh, design criteria for the electro hydraulic con turbine control system. And we'd say, it, it is this, these are the requirements that that thing gives to fit.

And then you have material requirements like the metals and blah, blah, all this stuff. So when you're done, you have an operating electrohydraulic turbine control system, but with the bright lights and flashy lights, um, the thing that wasn't said before was deterministic. In that EHC, it's very deterministic.

If these are your input criteria, this is your output position, your pound mass per hour into the turbine, is this right? If this is it, okay. So you can make an electronic version of that, which is very deterministic.

So for certain pound, mass per hour, coming in with a certain saturation of the, of the steam, you have a certain energy density, you have turbine blades, which are still the same. You have rotating parts, which are still the same. You have a specific output that you get to.

Very deterministic is your input. That's your output. When you get to another component of artificial intelligence, not the machine learning component, which that was working on, you did into the generative forms, all of a sudden the outputs can be generatively driven back into the inputs.

So you can dynamically take into account certain types of behavior that may emerge. So I'm get ahead of your skis on this one, but just taking the electrohydraulic control system into the electrical age, making a digital control circuit, gives you opportunities now to take all of those transducers. Imagine the 15, 20 different touch points on that turbine control system, the actual turbine and the generator, the 15 or 20 engineering touch points where you go in and sample the temperature, you sample the pressure, you sample the turbine speed, you sample a pound mass per hour coming into the turbine.

Imagine all those touch points. Now building up databases of all of that information, driving it down into the data. And then horizontal to that, you have this generative component, which is looking across all of that data, a generative component of mine.

So you can have a conversation with it, right? Not the machine learning component, which is gonna be sitting there looking at this input. Same thing, looking at the input going, oh, I hit a trip set point.

I'm gonna flip the reactor off. I'm gonna stop my input energy, I'm gonna trip the plant and go into a safe mode, because that's what my machine learning algorithm tells me. Reactor operator is gonna sit there and scratch his head and go, why?

Why did I just trip offline? I was just happy. I just got my coffee.

What the hell's going on? So you're gonna query your database you'd like to query. In the old days, this was very laborious root cause analysis, right?

You'd sit and you'd pull these logs and you'd look at the pressures and you'd go, well, we had a pressure spike right there. Say, well, that shouldn't have caused that. But in the new world where we, where are we now we're stepping out into a new place where we can have a conversation with the generative component of an ai, which is sitting on top of that network of information data associated with your turbine control system, with your turbine generator set, all those transducers feeding into all that data.

Now you can query the data and say, what were the conditions just immediately before the trip? What do you think was the cause of the trip? So now the algorithms of the machine learning are gonna be feeding information into the generative component.

And the generative component is the cool part. It's gonna create language. So it's very sophisticated, fill in the blank kind of stuff.

It's gonna create a language response back to you, tokens being bashed back and forth. And the result is, it's going to speak to you and say, the conditions were this, and it'll give you an engineering list of the conditions that were going on beforehand. The artificial intelligence generative component is feeding your real intelligence information.

And then you can ask another question, a probative question on that, and say, out of all other GE mark X, whatever turbine systems, what were the root causes of reactor trips associated with the turbine trip? And it'll search across all of the database that the vendor has across all of their installed base, trying to answer your question through the generative component. All of a sudden, you have access to the root cause analysis reports being fed back to you through a generative generator, uh, from some other facility that actually did the root cause analysis on their trip that was similar to yours.

Somewhere in the end of that is a human that is looking at the top three or four options and going, ah, I'll bet you it was this, that human did a function called command broker. It's looking to try and create a command to restart or not, but it has to broker that command across the multiple inputs that it's receiving. Look at how high up in the stack you are with all of the information that is being synthesized and analyzed to get to the point where you have a selection of four or five opportunities to restart your plan based on what you now see in this list of four or five.

You, the real intelligence can now ask the artificial intelligence to pull data associated with trying to figure out which of these five options is the specific one. In your case, you have amplified human intuition because your intuition is leading you into the brokering of these different options. You've amplified human intuition through a combination of machine learning on the transducers and a generative AI sitting on top of that data that has access to a lot of other information.

That's kind of where we are. We're at the edge of that, and we're still exploring the pieces. The pieces that are very, uh, interesting and troubling are, are what happens next.

These are all, um, I forget the terminology, but they're, they're the early stage of ai. It's, it's not, it's not the general ai. These are, uh, closed form ais, if you will, or, or, oh, that's the term narrow.

These are all, all the applications I've been talking about are narrow AI or a subset of that called machine learning. In that mix, if you put, if you put, uh, a generative form on top of where the machine learning is, then that reactor trip that it executed may or may not have happened the way you wanted it to. 'cause that component of generative is very deterministic.

You don't know. And that emerged behavior, that question. That's enough of a question for I believe, the regulators of the future to say that's not, that's not gonna be acceptable.

You cannot use generative Well, and, and that as well, you know, takes us, I think, directly into the next stage of this conversation. Because, you know, the reason, you know, I came up with this term inevitability curves back in the early nineties. 'cause I was just seeing this is the future.

In that case, everyone gets on the internet and it was so hard for people to think about. I finally had to start saying, just assume for a minute, it is, let's look at, you know, the, the, the lines that connect now to then, you know, over the, like, say 30 years from now where we are now. And in fact, everyone's here.

You have to think through these things. And when I look at the complexity of infrastructure, right, you know, where we have been in my little lifetime, you know, in our, you know, post-industrial revolution short period of time, and where are we going? Are we going to less complex systems that are less credit?

No, quite the opposite. And what bounds, you know, what I would see as an inevitability curve are the, not the possibilities, but the impossibilities. Are we going in 20, 50, a hundred, 200, a thousand, 10,000 years going to be going this direction or that direction?

Are we gonna be doing more hand work in this case? Are we gonna do more automation? So for the purposes of argument anyways, we're gonna do more automation, you know, lots and lots and lots and lots more.

And everything you said is also true. You know, so that we have, there's future versions of ourselves who will have figured these things out, or we can't have infrastructure. So, you know, I, I think you've led up to a lot of these, these things, but you know, which other way you want to take it, you know, what are we not going to do?

What's completely off the table that people are expecting, uh, to be part of this in the, the medium to distant future in 30, 50 years, you know, beyond perhaps you and I time, but enough that we should worry about it. Now, I think that, um, there's a paper I wrote a few years back on, um, the transformative initiative that Asante and I kicked around for a while. I came up with this idea and I said, we're gonna get to a point where we're gonna automate with complexity to the point where the systems themselves, or the system of systems as the case may be, uh, becomes very brittle to failure.

Um, and the failure is gonna be extremely difficult to figure out. It's gonna be trans conceptual. You're, you're not gonna be able to conceive of the failure.

So, and there are limits to what we can conceive of a failure. And sometimes we have the big button on the wall that says, you know, give it to God, let the, let the higher being sort this one out. So when you get into the position like that, the transformational initiative, um, really it takes into account kinda like the inevitability, the complexity is gonna happen.

It, it's gonna com keep getting more and more complex, and it's gonna get to the point where human conceptual abilities fail. So you're not gonna be able to actually see and understand, uh, absent the genius level that can, that can see things, um, you're not gonna be able to sue you and control. So how do you, how do you get around that?

Well, in the industrial control spaces, there is, there is connection between the logical weirdness of the world, this fantasy world in the, uh, in the virtual world, which mimics a control, which mimics a, a power plant, for instance, which the computers are operating. So there's this, this, this thing that's out there, a virtual power plant, uh, digital twin, if you will. There's this digital twin that the computers are operating, but there's a real twin too.

And it's the thing on the ground. It's the thing that's flying through the air. It's the thing navigating through your arteries with a surgical probe at the end.

It's, it's these things that real thing. It does have specific strength and materials and limitations of its design, design basis again, right? And what happens when you get design beyond the design basis is the structural integrity will probably fail.

Or the, if it doesn't fail the structural integrity, then maybe you put, um, you put an elbow joint into a position where it can't recover. So it, it breaks. One of the ways that we came up with this was something I helped design in the early days.

Um, you remember the Aurora vulnerability, right? Everybody vulnerability an asynchronous attack against the rotating AC machine. Well, if you look at the physics of what's going on at the Aurora attack, there is an industrial control system that has opening and closing breakers, but there's also this real machine out there that when you open the breakers, it starts to go faster.

When you close the breaker, it slows down dramatically right now, right? But when you open the breaker, it starts to go faster. Well, you can program the logic for preventing it from going faster into the control device that is actually being manipulated to open and close the breakers.

But that's the one the adversary already owns. That's a hyper complex system. It's a system of systems.

It's on the end of your scada, it's a circuit breaker or something. Somebody far, far away is manipulating that breaker, not you. As a matter of fact, they probably denied you activity and they're probably showing you what they want you to see.

But if you have a device that's over on the box that's spinning the turbine, the generator, if you have a box on hit outside of this digital realm, outside of that digital twin, if you have that box, which is only associated with that device, turbine generator in his case, and its only connection to the world, is the closed coil on that circuit breaker, then as soon as that circuit breaker gets opened and the design parameters are exceeded by this local box, it'll send a circuit, it'll send a signal over here, hardwired that will prevent the closed coil from Energizer. Basically, it won't close, therefore the circuit will remain open, therefore, other protections in the system will trip the turret. You've just prevented an or aurora attack by thinking very local on the industrial control device that you're monitoring that envelope.

That envelope is not associated with any of this other stuff. It's associated with the grid physics of that thing. You're trying to protect these envelopes.

We have forgotten how to make them. We don't do those anymore. It's a very simple circuit.

And when you look at the logic, uh, and you look at the, the physics of the DDTs curve, you know, the acceleration curve based on frequency, you can plot a very simple plot of where you don't want. That closed code would be energized. And if you get a signal that tries to close it during that, if the signal doesn't go where it's supposed to go, it doesn't close, it's, it's a transformational initiative to transform the way in which you approach the solution to a problem by not trying to put all the solution into one set of things, but by creating a thing, part of the machine, part of the target generator that's built up through the supply chain of the generator that comes with it.

And that that is the component which prevents the destruction of the generator. I, I was hoping you would go right there with that. 'cause I think it, uh, it, it wraps.

I think we're sort of rapidly agreeing on this one, but I wanna see, see about that. 'cause one of my favorite things in all of industrial, uh, control systems, industrial systems in the modern era were the biome metallic bearings of steam locomotives in the 18 hundreds. Oh, Absolutely.

Good for, weren't they? Oh, yeah. Because if you're, you're operating a, a train, you're going down the track, you know, and those bearings overheat, you know, Hey, you break your equipment, you're gonna die and break a lot of other things, but your nose actually works really, really well, and you can smell it because that, that biome metallic bearing will smell different as the temperature rises.

And I think there's something of an analogy to that with where we're going with AI and all this stuff, because as you just said, you know, there's, you know, Aurora, for those who out there who don't know about that, Google up, Aurora, Idaho National Labs, basically imagine a, a locomotive size generator being hacked into and smoking and burning himself to death. And to your point, and you know, you can, we can, and we know how to engineer things that just don't fail. And I think, you know, as you arc forward, you know, we're not talking a couple years, you know, but maybe over the next couple decades we start, you know, as you just said, putting that back in, you know, let's not build things that will just be, you know, fail on a mechanical basis where we can do that.

And with the com time compression of ai, you know, really enable operators to do the math crunch numbers fast enough to stay on top of the complexity of, of modern infrastructure. Yeah, pretty much, pretty much it that, so, so in a document I've written for some other people, ask me this question early on, um, uh, bright line criteria. Where can generative apply and where can machine learning go?

Machine learning is already being done, uh, very advanced. Um, software enabled, uh, systems, uh, using machine learning algorithms right now, and they're doing very, very well. And it's very fast.

Another term of art that's in involved in this is the command broker function. There's a, there's a set of, of, uh, commands that are given on SCADA systems across critical infrastructure, focusing on critical infrastructure, not non-critical infrastructure stuff that leads to extreme consequences. Billions of dollars lost, hundreds of lies lost, et cetera.

Critical infrastructure. There's a, um, a thing called a command broker where the ai, if you will, feeds the commander, whoever the commander happens to be, uh, a series of opportunities to enhance performance of the system and or save the system from destruction. And these have to be, these opera options have to be selected by the commander command broker.

You have to broker whatever that function is. Most of the software vendors have been dealing with, which are pretty much the normal cup 1520 across the globe. They understand these terms called command broker is very important to them.

Autonomous behavior of a machine learning component is, that's the electro hydraulic control system. Digital twin. I mean, that's, that's autonomous control.

It does it all by its vso. No one has to say, I think you should trip the reactor. It's gone before you can even think about, well, maybe I should trip the reactor.

Uh, so it's, it's already done, its performance. There is no command broker, it's already done by the machine learning algorithm, which is very deterministic or at least reasonably deterministic. The artificial intelligence version of that, with the generative component being far more in deterministic, it is generally felt by my supplier friends that, um, a commander in the sea, the command broker function needs to be there.

That's part of the bright line criteria. So no autonomous behavior beyond a certain point in criticality. Um, therefore, with no autonomous behavior, you have to broker the command.

Command broker is a trained operator. It's what we have today. But to your point, that command list that you will, that you're gonna have to broker across that command list can be quite a lot.

That those machine learning algorithms and that generative AI can give you a really beautiful list of all kinds of cool stuff. But as we've seen in far more trivial, um, implementations of generative ai, some of those commands opportunities that it may post on that list of things you can select, oh, do number five. Some of those commands may, may be hallucinations.

Uh, some of them could be extreme biases in an industrial control setting, you don't typically get the misogynistic behavior in your industrial control systems. But there is an, there is another kind of aberrant behavior that you can get. And some of those, those opportunities that you're supposed to select from may not be the best one.

So your command and the experience with the machine, the industrial process, that's gonna give the ability to get the best selection of that criteria, best selection off the list or the selection that's not even on the list, that's the relationship between critical infrastructure, autonomous behavior, machine learning, and generative ai, which requires a command broker. And that's, that's kind of the brighten criteria in a nutshell. Well, this is, it has always, it's been, it's, uh, it's always fascinating talking with you.

And, and if I don't say it too often, you know, I I I, it sincerely amazes me that I, I get to just have these conversations, right? And you and I have discussed this before. I, I think that's the defining characteristics of, of, of sort of our community is we're all sort of standing here going, is this real?

Are we really talking about this? And, and we are, right? Yeah.

Someone has to figure it out and, uh, and I think it's in pretty good hands, right? I have a lot of trust with our community. Um, you and I may be, you know, approaching the end of our, you know, active involvement in this stuff, you know, uh, but I see a generation coming up behind us that are native to a lot of these things that have the vested interests.

It'll keep coming up with those solutions to keep all of this stuff working indefinitely. Yeah. It's gotta be fascinating.

I hope to stick around for a bunch of it yet. Yeah, that's right. I'll, I'll go for the broke cause brain often if necessary, just so I can observe the whole thing.

Yep. But let's, uh, let's end the recording here and have more conversations later. Everyone else, thanks for coming along the ride with us.

Thank you very much, Tim, for the time today, your friendship and everything else you've done for the world and the industry. Thank you for, for being there, Chris. My pleasure.

Take care. Thanks folks.