How to Get Started Operationalizing AI | AI in Action 2023

Hello, I'm Amanda Razani. I'm the custom content editor for Techstrong. I know you've been enjoying the AI and action event, so many wonderful topics being discovered today.

And we have three amazing speakers. On our panel, we have Alan Shimel, he's the CEO, we have Mike Vizard, he's the editor in chief. And Mitch Ashley, he's the CTO and, uh, in charge of research at Textron.

Happy to have y'all on our panel today. Thanks, Amanda. It's great to be here.

Awesome day. Yes, it is. We've learned so much and we've had some great topics we've been discussing.

So to start, my first question would be co-pilots. They seem to be the big topic of discussion right now and the dominant means for leveraging ai. Um, but how are they going to be managed?

Can y'all speak to this because that seems to be an issue? Sure. Well, I'll, I'll jump in and, and kind of get the conversation started.

I think it goes to a central question of is AI gotta replace everybody? And I think most of us think no, it's, it's going to enhance and, and help us be more productive. So copilots are an easy way to say this is an adjunct assistant in some way, helping you, whether you're writing code or marketing or documentation or whatever it might be, particularly around generative ai.

But that seems the market is responding with how that's a lot of times how they're introducing generative AI as an assistant co-pilot. So I, I will tell you, oh, I'm sorry. Go ahead, Mike.

I think the challenge is gonna be not so much whether we have co-pilots, is just that, how many of 'em are there gonna be? I mean, Microsoft's got a co-pilot for every little thing, and every other vendor is building something that either looks like a co-pilot that they're gonna call it that or they're gonna call it something else. But essentially it's the same thing.

And I'm scratching my head going at some point, will all the co-pilots know about each other? It's kinda like having an assistant, but none of my assistants know. Any of your assistants.

Have your people call my people, have your ai, call my ai. So, so I, I, you know, surprised, I'm gonna take a little bit of a contrary opinion here. So first of all, I, I have to tell you that I was amazed that the amount of companies that are actually calling their AI assistant co-pilot, knowing Microsoft is du for 30 plus years, their lawyers must be sharpening their teeth and billing their hours right now saying, how can we be the co-pilot?

Right? And I think Mike pointed it out to me, we were in Las Vegas, that it's co-pilot with a small C maybe Microsoft's is a big C, but sooner or later is co-pilot become a generic term like, uh, you know, Reynolds wrap for aluminum foil or, or, you know, pick, pick a or Kleenex or something like that. But more importantly, beyond the what's in the name, you know, I think initially a lot of people use that chatbot interface for chat GPT or Bard or whatever chatbot, you know, AI you're interfacing with.

And it was, you know, you would ask it a question and, and what separated the, the pros from the amateurs, and we're all amateurs still really is how well you put your prompt in, right? Who, who's the better prompt engineer? And I used to laugh that that wasn't a real term, but over time, I've, I've, I think I've come to the conclusion that a prompt engineer is a real job.

That being said though, I think when we talk about co-pilot, what you're talking about is embedding AI functionality right into your, your app that you're working in. So whether it's Microsoft Word or, or, or, or Excel or it's, it's PagerDuty or it's GitLab. And, and a lot of 'em have different names besides copilot, but whatever, whatever the app you're working in Salesforce, the ability for it to copilot your mission, to copilot your flight is, is I think what, what's inherent here.

So it's not just answering a prompt, it's actually helping you drive your, your app or whatever your workflow is in your app. And so from that point of view, copilot becomes a term of art, a generic term. And not everyone's using copilot.

Last week at AWS or the week before at a WSI heard Davis and Duo and, and, uh, Mitchell and Mike, you probably heard a bunch of 'em yourself, Einstein and Bedrock, I think of Fred Flintstone, but you know, all, all of these names. And we heard quite a bit of companies using copilot. So I'm not sure they have to talk to each other.

I think copilot is generic term for built-in help. It's your built-in assistant in whatever app you're working in. I always think we're, we're living with the rapture of Watson, you know, IBM named their Watson, their big computer, right?

Or Alexa, right? Everybody's got a name, a person name for their, for their bot or whatever this, whatever their, whatever their interface is, their assistant. And that's sort of the really what copilots are is designed to be assistance, maybe really helpful, maybe not.

I mean, in Visual Studio, I mean, it makes a big difference. Code completion sure does. Even does write a little bit of code for you.

We'll do more, but you know, I is a chat bot as an interface into your customer service. Helpful. Well, it all depends, you know, about what it is gonna Respond With.

I'm, wait, I'm gonna disagree with my most colleague over there, Copilot, go ahead. Disagree with Alan. Co-pilots would definitely need to be integrated with each other.

If you look at any process in the enterprise, it spans a minimum of four to five different applications typically. And they're all gonna have co-pilots, and they're all gonna have to have some way of invoking each other. There'll be need to be some automation framework that sits between the copilots and these things are gonna, you know, need to become covid The same way every app integrates with each other.

We'll have APIs. Yeah, right. I, I don't think that's an AI specialty thing.

I think that's just right. I think it's just an a an API kind of thing. But, you know, Mitchell, a funny thing about what you said about whether it's Alexa or, or Siri or whatever, you know, my, my, our latest car came with the BMW help built in, and you can name it whatever you want.

So I thought I was gonna be cute. I named it Esther until we had someone in the car named Duster. Don't do it.

I just, I, that's my only, that's my only advice to you. Don't, don't name your AI some common name that, you know, we Yeah, it's, it's bad news. It's bad news.

Hmm. No, well, well it depends, you know, that depends on, you know, are you worried about it sending you out, opening the pod bayit doors or something. But, um, in any event though, but certainly, I, I, I'll give you an example.

I spoke with, uh, chief product office officer at GitLab yesterday, David, and, you know, they have this talk to code copilot, they call it duo, but it's copilot where you just tell it, Hey, I, I, I want, I want my app to do this, blah, blah, blah, blah, blah. And it generates the code. It actually could generate a, you know, an app for you just telling it what you want.

That, that kind of freaks me out a little bit, to tell you the truth. But it's amazing. Amazing.

Yeah. Other, just to that point, other folks are talking about, um, uh, Microsoft is a good example where you're just gonna describe what you're at and want your app to do in the project management app and then the code will get generated. I'm not entirely clear if that works to what degree, but it's gonna be interesting times.

I, I think one of the interesting parts of this too is when we go through another person to get who is using the co-pilot or whatever it is I was doing, going through a troubleshooting problem on some Linux stuff the other day. And I asked the asked the person that works with the, where'd you get this script? This is actually really helpful.

He says, I got it from chat GPT. I'm like, oh, okay. Oh, hopefully embedded it before we ran all this stuff.

But, but here's, here's the point though too, guys, we're having this conversation because this isn't even the beginning of the end. Yeah. Or the end of the beginning.

Excuse me. This isn't even the end of the beginning. We are at the very first leg of this marathon, and we're still to figure out what the ground rules are and what, how everybody plays nice together in the sandbox, so to speak.

But I think these things will, you know, in hindsight, they'll be obvious whether or not co-pilots should talk to each other, how we should hook them up. What are the right, right APIs here, is it a bunch of zaps or something, you know? Um, and, and I think that's why we're having this discussion because it's still kind of uncharted territory a little bit.

So have your copilot talked to my copilot. We're good now. Well, you mentioned chat, GPT, and that's a good segue into really the main AI focus seems to have been generative AI this year, uh, with OpenAI and chat GPT.

So how are we going to embed generative AI into workflows across the enterprise? Well, I kind of started that conversation already, but jumping into it a little bit deeper. Um, the issue is gonna be we need some standards.

I think we need some ways that these things can talk to each other. It may be an API, it may be something that's a little more direct, but I feel like, um, there's an opportunity for more technical standards to emerge in this space. And everybody seems to be rushing out building their own little, you know, quasi open box.

But I think we're gonna have to at least take a step back and say, what are the specifications that we can all count on to be there as a fundamental thing? I don't know, Mitch, what do you think? Well, Al and I met with the, with the, uh, tech vendor.

I don't know if I can say who they are during AWS but their approach was, I, I'm gonna create a layer between your workflow and all of the LLMs and, you know, generative AI things. 'cause we may wanna use different LMS for different things, but they actually kind of create their own prompt layers where they prebuilt the prompts of what workflow actions might look like, and then you can evoke that through their layer. So it's, it's a sort of, um, kind of a cleanup layer, I would say, or simplification.

Well, It's a trust layer too, right? I think they called it a trust Layer. That, that's a really good point.

I'm glad you brought that up. Say some more about that. 'cause it was a very interesting conversation.

Can I name that company? 'cause they're kind of, Yeah, go ahead. Salesforce.

Are they paying us? Okay. No, they're not sponsoring this event, I don't believe.

Okay. Take, take that. Alright.

Name them anyway, Mike. Hey, it's the CEO in me. I had to mention it, but go ahead.

So, you know, at their base level and then gross oversimplification, but to Mitch's point, they are creating a trust layer that says, um, we will validate the LLM for toxicity and governance and guardrails and all these policies. And the idea is gonna be that you're gonna have multiple L LMS optimized for specific tasks rather than one single general purpose chat G-P-T-L-L-M. And we're gonna mix and match these things.

And you as an end user may never know what LLM they're using at any given time. They're just gonna swap those things out. And we could be looking at the commoditization of LMSs already.

Absolutely. And it, what, by the way, it wasn't just Salesforce in terms of using multiple LLMs on the backend. Uh, any number of the companies we met with at reinvent, you know, and there were 60,000 people at reinvent.

Yeah. But another interesting thing that we didn't mention about that trust layer was that it also prevents your data, whether it's data you're putting into the prompt or data, you're, you're getting out of it from being, you know, Borg Borg into the LLM from being, you know, assimilated into the LLM. And so I think that's another big issue again, as we take these baby steps into this brave new world of how do I protect my IP with these co-pilots?

If, if, do we have a multi-tenant, co-pilot? Like we're seeing, let's say in the Microsoft case, right? How do I make sure, now the, one of the companies we spoke to and I, this, I don't know if we can mention, but one of the, the company's name, but one of the companies we spoke to contractually, contractually contracted, that's kind of a double entendre, but they, they, they entered into a, a contract with OpenAI that OpenAI Chacha BT would not use the data put into the prompt or what's coming out.

Now we talk about regulations around ai. You know, when I went to law school a hundred years ago, I had a law school professor who said, the law always trails technology by three to 10 years. I don't know if we could wait 10 years for this, I don't even know if we could wait three years.

But we need certainty around IP ownership if this stuff is really going to u take off the way we want it to take off, right? I can't, I can't trust Salesforce to trust layer to make sure that no one, or maybe I can trust Salesforce. And maybe that's what's needed.

I need that trust that my IP is secure, sacro cit, it's not gonna be violated. And when we start talking about plugging into multiple LLMs and they, and chat bots and so forth, in the back end, I want to know who I'm plugging into and whether or not my IP is protected. And maybe that's, that's a great angle for Salesforce, right?

Where the trusted force, you don't have to, we don't have, you know, worry about who that LLM is in the back. I don't know. But I, you know, I think, I don't know if we could wait for the government to, to figure this out.

It seems like, it feels like we're at the beginning of the cloud era times a hundred where, you know, is kind of, is it secure to operate in the cloud, right? Eventually that question got answered. IP is one issue.

Data protection. I mean, just, there are dozens of unanswered questions For it. But for instance, we have the GovCloud, right?

You have GovCloud where security and, and access and where stuff was hosted was very well defined. If you wanted to have, you know, the government using your cloud platform, I think we may need something similar here. Maybe not.

I, maybe, maybe Salesforce provide, maybe that's the, the ticket, you know, that they'll all provide this trusted layer. And maybe at that trusted layer is where they all talk to each other. That'll make Mike very happy.

Course It's about making a mike happy. I'm glad we cleared that up. Trying to have, we do things here at tick text.

We always think what will make Mike happy? Oh, What would Mike do? What would Mike do?

Aren't You glad you signed up Amanda for this? Yeah, I'm sorry. No, I love it.

Okay. I love hearing you speak about it. Um, so to that note you mentioned, um, so we're speaking about the large language models and how many there are and, um, we have the private ones and the open ones.

And with all these large language models strewn all over the enterprise, how do business leaders best manage these? Who is going to manage these? And since the government's, um, not really quite there yet, it falls on the enterprise, I believe.

So what do y'all have to say? I think there's gonna be multiple classes of LLM. So there's the foundational ones that, you know, the big tech companies have built and there's open source ones and they're all over the place.

And I think people are extending those using things like vector databases that they stick in front of that so that it can see your data and kind of be extended. But ultimately, I'm not sure that I need or want a general purpose. LLMI think you're also gonna wanna customize those LLMs, build your own and kind of build some data into that.

Whether you want to build an entire LLM from the ground up is, you know, maybe, maybe not. That's a lot of work and maybe something that Morgan Stanley would consider, but there's a lot of nuances along that curve. And I think, you know, the answer might be all the above, You know, quick plug at the end of, uh, to, towards the end of today's, uh, AI in action virtual event.

We have about a 40, 45 minute video playing of a hackathon that we hosted here at Techstrong Headquarters, I think back, I guess in late August. And, and we had, we had some, some really smart people, way too smart for me. Um, some of the people put together the whole DevOps and DevSecOps movement and, you know, they explored this, right?

And the, you know, for those of you who are new to AI or just dipping your toe in the water and are not quite sure, the way they explained it to me is think of these giant LLMs, like an OpenAI or Bard or llama or these as, as your long-term memory, if you will, right? There's a huge amount of data there. It's a huge data lake.

And then the custom LLMs that you can create from a Vector database, you know, inject in, think of that as short term memory, right? So now you have your, your large LLM and let's say your custom LLM. And when you then query or, you know, chatbot put a prompt in it first looks into that short term memory, that custom LLM, and then what it doesn't pull from there, it pulls from the bigger LLMI don't that that secret's outta the hat, right?

That that's no longer gonna fool people in in Oz anymore. They know it's just the man behind the black curtain. So everyone's, and you mentioned, I think it was not, not, uh, what's the site developers go to?

And you could get code, um, Mike, you mentioned it. Open stack. No, no, no, not open stack.

Um, Are you talking about HuggingFace or No, you go, you go and ask questions and the community answers. Oh, um, Oh, oh, stack overflow stack over Stack Overflow. Th this is ex, sorry, Mitch Senior moment, but this is exactly what Stack Overflow did.

I, I interviewed their CEO, it's on our tech drunk TV if you want to take a look at it. They took almost their entire base of questions and answers and created a custom LLM that sits on top of another broader LLM. And, and you can query is, is that gonna be the model going forward?

Maybe with, you know, some sort of, uh, copilot that makes that easier to query that big thing for you and better gives you a better prompt? Maybe, maybe, you know, I've spoken to other people who say, nah, that's not the answer. That's like a, it's an ex it's a dead end ex.

It's like the Neanderthals, right? It was, it was, it was, it's on the branch of, of humanity, but it's kind of a dead end Knuckle dragers. Okay.

I mean, the, the more, the more data that you throw into the L lmm that's not vetted or specific to your task, the more likely you are to get a, you know, hallucination Or poisoning, right? Which is again, something about that trus layer. The Trus layer says, you know, they're gonna weeded out the hallucinations in poisoning toxicity or whatever you want to call it.

We're coming up with some great names for this stuff. I think that's just one use case, which is, you know, I get this big massive amount of data, Wikipedia, you know, whatever, whatever it is, right? Put in LM generative AI chat, natural language in front of it.

Um, I I, I seriously wonder like, is every organization gonna be going out there building LMS for themselves or training LMS for themselves? Seems to me this is ripe for sort of like the cloud hosting environment. I want you to do my training, my LLM for me.

Yes, it may be my data and you'll protect it and I want you to operate it for me. 'cause, uh, are you, are you gonna have, you know, 20 different groups around an enterprise going around creating your own LM LLMs with different versions and aging of the same data and you know, you it, it'll be a mess. I don't see how you can, uh, maybe we will figure it all out someday, but it seems pretty complex to me to say it's, you're gonna do this on, uh, on in many groups at large scale and have these things as sensical.

I think you're onto it. I think it's gonna be a mess because the LLMs are down into the size now of terabytes, right? I mean, it used to be these large language models where petabytes of stuff that you built and somebody had to go and curate all that stuff for you now.

But with each passing iteration, I can now use massive LLMs and I can use what are essentially by comparison tiny LLMs for a specific purpose. So I, every developer I know is gonna be like, yeah, I'll just use this. And away we go.

And, you know, somebody on the DevOps side, I'll figure out the best later. I I think eventually you automate, have a co-pilot that builds your custom LLM for you, puts it, you know, puts it in great gathers data, puts it in the vector database, and there you go. So that brings us to the next question.

Seems like a good question to ask now is, will there need to be a convergence of DevOps and MLOps? I think that happened way before generator of ai. Yeah, I agree.

I think it already has happened or already needs that. But if it Hasn't, in, in, in the, in the marketing lexicon, they called it AIOps. I don't Right?

But think that those are necessarily the same thing. I mean, AIOps is more about applying AI to the actual management of IT operations. MLOps is the, you know, the best practices that data scientists are using, the construct the model.

But the problem is, is once they construct it, they gotta throw it over the fence to a DevOps team to deploy it. And sometimes that's through an API and sometimes it's an actual artifact that will get embedded in the application. But I think there's work to be done in that space still.

I do too. I mean, think about it this way, is if you had multiple work streams that are happening in parallel that need to be quarter coordinated to, you know, produce some sort of, uh, releases into production and ai, LLM is very much like that. It's not just a database, right?

It's something's being trained and evolved and updated and how does that flow in integrate into the workflow along with, uh, you know, other software that's being created. So I think that's where the gap is, Mike, of figuring out how much, where, how do we, how do we kind of, uh, integrate the workflows to the degree necessary. So if we're not going down divergent paths, Right?

And there's no concept of version control in the land of LLMs, right? When you retrain the LLM, it's an entirely new thing. 54, that's version control at some level.

They're, They're all different in, in that sense that they're not like version 10 of my previous app. They're kind of fundamentally different software entities. Yeah.

And with different functionality. But you know, back to back to what we were saying about AI ops MLOps, I, I think unfortunately a lot of people meant MLOps when they said AI ops. And I think that's the way it came down to DevOps.

Um, but, but that being said, here's how I ha to me, the real question is how does that world of machine learning and AI ops, if we can call it that, interact with generative ai? 'cause I think to me this is akin that you need an atomic bomb to set off a hydrogen bomb, right? Because that, that you, you need, that's how hydrogen bombs work.

Now. I think what's gonna happen is we can create great insights and, and recognize unbelievable patterns and, and, you know, observe for observability sake, amazing things using ml, MLOps, you want to call it AI ops, whatever. But taking the findings, taking what you find in those patterns and then creating actionable steps, uh, real, you know, taking action is I think where generative AI somehow gets that feed as a prompt or whatever you want to call it, and then does something as a result, you know, using AI in, in its pure sense.

So I think there's a world where these things can converge and come together and, and you do have that nuclear age, right? Then throw some quantum computing in there just to make your head spin. But that's where we're headed.

I think you're onto something there in the sense that, you know, if I look back in AI models, there came in two flavors, right? They were all machine learning derivatives, but they were, one was predictive and the other was causal. And now we added this generative capability.

But I, you know, I don't think generative eliminates the need for the first two. I think as we go along, all three of these things are gonna start showing up in various applications and platforms and we're gonna need to weave that into something that feels like a fabric that works. Yeah, I think, I think maybe to the question about sort of, does, does the atom lead lead to hydrogen bomb?

You, I'm gonna grossly oversimplify, but you get an idea of this from the video about the hackathon later on when we say training in LLM. Uh, really what a lot of that training is, is training, uh, the how to resp, what kinds of prompts to expect and how to respond to those prompts. So you get meaningful information back out of the LLM, otherwise you'll get whatever and you know, it doesn't know what to expect.

And that's why you can get really weird ization a lot of the time. So I kind of believe over time that training process is something you can develop models for, right? There's certain ways of doing different kinds of training of different kinds of data.

Like APIs behave this way, chatbots behave this way, you know, pick, pick whatever the use cases are. Maybe that AI can then help us with the training. Maybe it's co-pilot for training of the LLM.

I think that's where AI kind of could fold back on itself. Alan, I'm not quite sure if it's as big as, you know, creating the next big, big bang or not. But wait, Wait, are you saying that the hiring of retirees to tell machines that this is a cat a thousand times is not gonna be a long-term job?

You mean I don't have job security, what you're saying, I'm not gonna go there. You know, keep up. That's me into this one.

Um, Amanda, back to you. Yeah, moving on. Um, so we know business leaders, they're all trying to implement this AI technology in some way.

And we know there's many positives to harnessing ai, but let's talk a little bit more about, um, some of the cons and the cybersecurity issues. What are those issues and how can we address these issues moving forward? Some of it is just classic phishing, right?

People are gonna fish their way into these AI models, grab the credentials, and then try to poison them. And you can get folks out there who will attempt to deliberately introduce a hallucination into an AI model if they're allowed to. So that's gonna be, you know, I think the most common cybersecurity issue up front, Mitch, I don't know.

Uh, I, I agree. I think there's an flip the coin the other way. And the fisher could be the, the, uh, LLM, the generative AI that's talking to you.

You think you're talking to a person through an email or chatbot, suddenly they're now in, in the middle attacking you and you think you're talking to who, who you think you are. And that's not, so that, that seems like one, I think another really big one for everybody is DA is data protection. Like how do I know that if you stole my model and all the training that's gone into it, what's the value of that ip?

It's gotta be the value of the data times, whatever, right? That would be talking about stealing ip. That, that seems to be a big threat is the data.

I know there's a gentleman in the back of the room that's a white shirt. I think his name is Allen. He has a question or you, you No, well, no, no.

I I I like to let Mike, usually you gotta give Mike three things to say. 'cause he always has three things. Mike's gotta be happy.

Think I, I think the other big issue is that, um, it's software, right? And it's built using open source components and the open source components have vulnerabilities in them and they're gonna be exploited just like regular software. Sure.

Yeah. And, and we don't have, Whether it's open source or not, there'll be vulnerabilities. I don't want to, I'm not gonna bash open source here.

No. All right. They'll have to do Hoover in animal house.

Not gonna, we, we're not gonna be part of this. We'll start playing some patriotic music and walk out. I don't care what Dean Wormer says Either.

Either way. I think you just called me Dean Wormer, but okay. I know a Dean Wormer, I'm a husband Dean wormer.

But anyone, um, the issue becomes, it's harder to remediate these AI models once they're in production because you gotta pull 'em all back and retrain 'em if you got an issue with the vulnerabilities. So there's no like, concept of patching an AI model. So I think this is gonna be a bigger issue than we think.

Yeah. So as a security person who's been an InfoSec, cyber, whatever you want to call it for a long time, does anybody really care? Does anybody really care?

Every, every innovation in my lifetime, the security people have stood up and said, wait a second. This, this is gonna be a security nightmare. How do you look in stripes?

Let me throw some FUD your way. We, we halt, halt immediately. Stop wall activity.

The trains left the station folks, right? It's up for, it, it for security and for those interested in it, we gotta hop on the moving train here, right? But let's not delude ourselves into thinking we're gonna pull the emergency brake and, and figure out a way to secure it.

It's happening, it's done. We gotta live with the consequences. So instead of trying to prevent security things from happening that are kind of, it's already, you know, the barn's door already opened, the cows have run out.

We need to start coming up with strategies for how to respond to these possibilities. And response is, is is where it's at again, we're at the beginning or at the beginning of the beginning. Security will evolve.

We'll adapt, we'll be there. But I, you know, I think AI's gonna have, make security better in some, in some ways, but it'll also pose some challenges that we'll, we'll have to figure out. But let's not dilute ourselves into thinking somehow we'll slow this down as a result.

Just The best, best way life don't work that way. Best way to stop a train is not by jumping in front of it. Is that what you're saying?

Yeah, exactly. As Dick dastardly would, would know, Tied to the train from Tying Penelope or whatever her name is. To your point, whenever there's a threat or, or some, you know, something disruptive, you can, it can be disruptive.

Or you can also say, well, how do I take advantage of that? How do we turn that to our advantage and secure? Yeah.

Right. That's what I think the energy should be. I think.

I think when you pull the emergency brake, the first thing everybody asks is, A, who pulled the brake B, Y and C Throw them off the train. Why are they Still the Train? But you've been riding the Long Island Railroad and Metro North too long, Mike.

And in other places they don't necessarily say that. Okay. Especially the throw 'em off the train piece.

Um, but it's okay. Evander, I think we probably have time for one more. Yeah, Yeah.

Um, so yeah, there are two sides to every coin, but, um, to your point, there will be some regulations down the road. I'm sure they'll, um, have some, some government say in, in ai. And so how do business leaders need to address the compliance, um, aspect that's gonna come down the line?

I think business, but They always do. They'll check the box, They'll fill the forms out. I think there's a little hesitancy though to kind of go in full bore into production.

I think everybody's experimenting, but they're looking for a little clarity. 'cause nobody wants to go build something and then have to roll it back. So there's kind of a lot of pressure, I think to, to at least put some regulation on the board that people can count on as being somewhat stable.

And hopefully there aren't 52 of them in various states. And I'm wondering at the end of the day, if, um, you know, these regulations come down the pike too, somebody's gonna try to start auditing this stuff and then they'll be, you know, explainability requirements and all kinds of funds, things will happen and they'll all come with a fine at the end of the day. So I think we should move sooner than later, regardless of what the regulations are.

And like, let's give people at least some sort of guardrail that count up. So I disagree. We're not gonna have 52 different regulations.

We'll have two or three states, maybe California and New York Arc, Florida and Texas ain't doing, and that's exactly it, Mitch. We'll look to the EU to come up with some regulations, and in about two, three years they probably will, But already are they already have it in in revision, right? Or review.

Well, Yeah, but it could take, it could take until they pass it. And then they just 'cause they pass it doesn't mean it starts being affected, effective day one. It'll be two years, three years.

But that being said, let's not again, let's, let's be realistic. There's a lot of frigging money being invested in this AI stuff. They aren't waiting, they're not waiting for compliance.

They'll comply later, right? The, the, again, the chain's left the station, there's billions of dollars being poured into it. Silicon Valley has taken a, a hard right turn into ai, as has the rest of the tech world, the VCs, that's all that, not all, but you know, that's where the money's pouring in.

Follow the money. So lesson I've learned, I, I do agree. I think there'll be a couple that most, you know, be the predominant whatever the reg regulations are.

One from Europe, one from US North America. Uh, it it, to your earlier point, you know, Alan, about the legal catching up with the technology. Same thing in regulation.

We, we do, we are moving faster than we have in the past, but I, I think they'll, my guess this is a total guess, they'll probably build off what they've already done around data privacy protection and they'll kind of address that part of it first and maybe take some attempt at like, you know, what's responsible AI and not defining it, but do you have processes in place to make sure you're using AI responsibly? And you have a definite, you know, they'll have the, you have to define it, but then you have to follow your process, those kind of regulations, because they don't, they don't know any more than, I mean, we know more by, by implementing it than they're trying to write regulation for it, right? So I, I think right now most of the talk around regulations revolves around data privacy and ip.

And that's probably a good place to start. But I think there are other regulations here that are going to pop up. I think you're gonna have something akin to like Asimov's Law of Robotics or something, right?

In terms of what we're gonna let ais actually do, right? Let's not, let's not give them the key to the car, the keys to the car and regulate ourselves to the backseat and, and, you know, 'cause who knows, right? Um, Isn't that a Tesla?

But you know what? They tried the driverless taxis in California and they pulled back, right? Um, I, I think, I think you're going to need some sort of laws of AI like, like robotics, where, you know, do no harm, do no evil, do whatever.

I don't know. Yeah. Elijah j Bailey, which I'll, I'll give a quick, a quick, uh, a quick, uh, plug.

RSAC this year, again, we're putting on our DevSecOps, DevOps Connect DevSecOps, but it's DevSecOps in the world of ai. And our keynote speaker is none other than David Bryn, multiple Hugo New Nebula Award-winning PhD from JPL and probably the one of the foremost futurist on AI in the world for the last 25 years, who also happened to have written the sequels or one of the sequels to the foundation commissioned by Asimov's family. So he's really familiar with the laws of robotics, and he'll probably talk about this stuff.

If you're going to RSAC, it's Monday, May 6th, we're there all day in the Moscone Center, and we even have, uh, somewhere on Textron, you'll, you'll have a free code for an expo pass, which will get you in to our event that Monday. So all those Asimov books, they end happily, right? Um, yeah, Yeah.

Something like that. Well, the next generation comes along. Well, you the next book when the other ones book, right?

You read the next book, The next book. My 2 cents, my 2 cents in the whole thing is we can't afford to be Luddites, right? We gotta move forward, we gotta kind of keep this thing going.

But I would say that a certain amount of caution is required because as one folk shared with me the other day, he said, you know, the thing about ai, it's one thing to be wrong. It's another thing to be wrong at scale. Yeah.

And it may happen, right? Right. That's what makes counts.

You know? So to that note, as we get to a point where we have just a few minutes left, let's wrap this conversation up with your best advice or your best key takeaway for our audience as it relates to AI in the future. Mike, would you like to go first?

Don't, don't quit your day job. You're not gonna be replaced anytime soon. So show up.

I agree. Go to work, work might be better, might be less toil might be less painful, but the whole notion that, um, you know, we're all about to be replaced by a bunch of bots is, uh, farfetched at best. Mitch, my my advice is go learn something about it.

Right? There are so many resources available to go try this or that. I mean, you can try chat GPT if you want to, but it's not that hard to go do a little, little kind of kicking around and trying these things out and taking a Udemy course or whatever.

The best way to understand it, I think is right now is to actually just apply some of it yourself in your own, you know, sandbox. I'm not saying go change the company, but that'll help you get beyond sort of the marketing talk and like, okay, I understand a little bit about what it means to train a model. I couldn't do one, but I, I've tinkered with it enough to know I, I've, uh, tried an electronics kit enough.

I can build a radio, but I'm not gonna build a real radio. You know, that kind of thing. So I would say this, look, the three of us, we're a little older than most of you out here watching this, right?

I wish I was your age right now with what's going on with ai because we're, we're set to boldly go, boldly go where no one has gone before. No humans have gone before. And embrace change, embrace progress.

Embrace these new technologies because as I've been saying all along, the train's left the station. Hop on that train, ride that train for everything. It's worth the, the, the potential is boundless.

There's so much opportunity there. Don't, don't be a stick in the mud and try to drag your heels. Do what Mitchell said.

Get some training. Don't be a Luddite. As Mike said, embrace this opportunity for what it is.

Every 20, 25 years or so or so, it seems like we throw the cards up in the air and we see where they land this time. This is one of those cards up in the air moments. Take advantage of it.

And that is a great note to end on. I wanna thank you all for coming and listening to our session today. If you miss some of the other sessions, they will be available.

And stay tuned for more great information on AI in action.