Just what is AI anyway – Cloud Field Day 22 Delegate Roundtable

I'm Alistair Cook. I'm the host of, uh, the, uh, lead Forward Cloud Field Day. So we're gonna pick off this particular round table with a discussion of what is AI anyway.

And so what is ai? Because it used to be that AI ML was the acronym that we all knew about. This was the thing that was gonna save us because it would be able to look at all of our data and give us insightful guidance about what might be happening in the future.

And then this behemoth rolled outta nowhere, released into the wild before it was really ready. And generative AI turned up. And all of a sudden, AI ML is not what AI is, and AI now has to be Gen ai.

And we had all of these people at, at, uh, keynote conferences telling us all about Geni. Um, which probably sounds better if you don't have a New Zealand accent. So panelists, delegates, what the heck is ai Anyway, Weller.

It's funny you should mention that. I think we saw this with selector ai. In fact, it's interesting if you, if you Google 'em, they're selector.

But in the, uh, presentation, they were selector ai. And in certain places they, they've added that AI goodness on the end. And I think that has to do with what you're talking about, which is the rise of gen AI and the, and the collective buzz around it.

But as we found in their presentation, it got very muddled because they've been working with ML before machine learning, which is a form of ai. And then now they've got AI in the front end with LLMs. And when I say AI in the front end, I mean generative ai, and they just kind of mix their messages.

And it was, it was very confusing. So I think in the, if you're going to be talking about both, you need to be clear on which one you're talking about. Because I think ML has been around in some form, another form Orano form or another for a long time.

And it's much more baked than Gen AI is. Which is much flashier, which is much accessible. More accessible conceptually to the, to the public, myself included.

It's just, you go in, you put something in and chat. GPT fills, uh, uh, responds and fills something out within 10 seconds, less than 10 seconds. So I think we have a problem because the academics broke our brains.

'cause they came out like it was, it was neural networks in the seventies. That's what the, a lot of them went to school for. And then it was machine learning.

Remember, machine learning is like calculus. But what you do with calculus is ai. Like the thing, AI is the thing you do with machine learning.

AI is like, uh, you know, the product of machine learning, machine learning is the science like, but AI is something that goes beyond the, the math behind it. I think that's what blew it up. And even there's a company that I knew they were called Octo ml, and after 18 months of getting beaten up trying to explain what they do, they changed the name to Octo ai.

And no one asked what they do exactly the same company. They, but because AI was now associated to do a lot of other things versus machine learning was just, it just didn't. Yeah.

And it's, you know, I don't know why it was probably gen AI that broke all of us that that became the thing we would call it, but uh, broke us. Meaning like our spirits or when you broke what? Oh, everything.

The bang, it, it broke our understanding of what real words mean. Okay. Which is a tough part too.

We also have marketing versus reality and, uh, oh boy. I could go on for an hour on that one. You could Stop.

You didn't, didn't, didn't we hear chat bots or ai Yeah. Chat bots have been around for years. I mean, isn't that a AI ops, let me throw that buzz word Out.

Well, I think ML ops, for the longest time AI was the future that wasn't here yet. And once it arrived, we gave it a new new name. So chatbots were ai, but then they became chatbots and we stopped calling them ai.

'cause it's just a thing that I have. And image recognition. Well, we got used to image recognition.

So that's not AI anymore. That's just image recognition. And then generative AI came along and somehow it's managed to retain the AI label and in a way, muddy the waters for everything else.

'cause when anybody says ai, now it's safe to assume they're talking about an LLM. Right? And so we needed that with some of the presentations this in over the last two days is someone would mention ai.

And we had to get them to clarify what they meant by that. Because in some cases they were talking about multiple different types of AI within their product. And there were parts where it's like, okay, it's great that you're using AI to make the response more friendly or make the interim interface more friendly, but you're using AI in this other thing where I'm very scared you're gonna give me bad information or hallucinate.

And they had to roll back and go, oh no, that's just regular old machine learning over here. I was like, well just say that. I know you want to say AI everywhere, but we've poisoned that term.

We could say algorithm instead, but we, it is no, no longer fun to say that one. So I think you have a good point that people just get, if you say ai, they get sucked into LLM, um, or generative ai. But I think a really base level what AI is and should mean is, um, you know, computing, doing something intelligent that it's not explicitly told to do.

So that's really where, you know, machine learning becomes part of it because you're just giving it data and then it figures out what, what the result is versus just, you know, hard coding it. Um, so, you know, saying, oh, it's, it's ai it's not machine learning. Well, that's kind of a silly statement because, you know, all these LLMs are machine learning.

That's what they are. There's, you know, obviously all sorts of different machine learning algorithm. But, um, I think you, the conversation needs to go back to, well, what's the point of, you know, artificial intelligence in the first place?

I think that we need classification. You know, there's image recognition, there's natural language processing, there's recommendation engines, and now there are generative AI tools that offer something different. Um, in the case of cloud.

'cause that's why we're here. The question is what AI is required to oversee a cloud and where is the value in all of those categories and the application of what we're talking about? And so we see in all of the major clouds, there is a little chat bot interface that's using a large language model.

Mm-hmm. So although it's, yes, we've had chat bots for a long time. They often were very script driven and narrow in what they were produce.

And so they were, as, as Mitch says, they were pretty much direct it, this, this is the path you go. Whereas the, the generative AI models, and I definitely think we need this distinction between generative AI and what's I normally refer to as predictive ai, which is the machine learning part. You know, this, this is saying based on the patterns of the past, this is what's gonna happen in the future.

And that's exactly how large language models work. It's just that their vision of the past is not nearly as specific and still large language models and neural networks. I'm, I'm not certain that that's the right framing to think about.

AI is how does it apply to the cloud? It's a little bit like asking how does TCP help with the cloud? Oh, I was just talking about our conversation this week.

'cause we're at cloud fields. Yeah, Yeah, yeah. No, no, I totally like, I get that.

But I think AI as itself is such a, a foundational, large spanning technology that it's more akin to something like TCP, so not as standardized that it's, it's going to be useful in uh, uh, so many different ways. And with so many different applications that it's hard to just zoom in and go, well what is TCP good for? Well it's, it's good for, you know, transporting data from one place to another and doing a lot of handshakes.

It's like a politician, but, um, not, Not DNS Ned. Well, you can use a T CCP for DNS, it just sucks. And, and there's a lot more handshaking, but you can use it for a lot of more data transfer and, you know Right.

Use it Data AI's really good at taking really large volumes of data and finding patterns and munging it into something that humans can use. And that is so broadly applicable that I'm, we haven't even started to distill the number of use cases now that we have the hardware and the tool set to start using it effectively. I Think that's a big Problem use cases, right?

So we can go on and we can use chat GPT to say what, uh, what hotels are available this time of year in, in Madrid. 'cause my family and I are thinking of going, but I think in, in the business areas, when you start talking about gen ai, there's just not a, a, a, a real good understanding of how it can, um, how it can, uh, help the businesses. And I think machine learning much more, more practice, much more, uh, understood, but a lot less sexy, more on the backend.

I mean, you've got, you've got, uh, AI that you're helping to run your own business, and then you've got AI that, uh, customers might be using. And there's, there's different vantage points and frameworks. Um, so those are, those are, I think Allison, you snuck one in that I want to jump on further is that I think you hit the question that answers itself is that ai, you know, the cloud enabled us to experiment quickly with ai.

And one of the first things we began to do was use that AI to then try to manage the cloud to optimize how we produce ai. We actually have the first self, you know, managing ecosystem there where we're putting a lot of compute and ideas towards how do we do ML and AI effectively cheaply properly, and we're using it to help us build it. Mm-hmm.

This is the first time where the technology is helping us to make the technology better, which is probably why the machines are gonna eventually take over. And that's why I say thank you every time and, and chat gt so it won't kill me first. You're, So There's also a piece I wanted to bring in from ned, which was that a lot of the machine learning that the traditional AI, or what I often refer to as production AI is now just a feature of something.

Mm-hmm. We've, we've all seen the, the big enterprise management tools that have taken a whole host of data and work out what the trends are. Uh, that's no longer being sent to us as, you know, shown to us as being an AI ml, you know, an ML tool.

Um, it's just, this is the way it works. It's just a feature. I don't think we're seeing a lot of, it's just a feature with generative ai is that to do with maturity and the timing.

I think it Is just a feature. Every single company that has presented an included generative ai, it's just a feature of their much larger products that Usually to chat with their Dogs. Right.

But I, but I think part of the problem with where we're at is so much hype around it that people bring the AI part right to the front mm-hmm. And they slap AI on the end of their name. Um, but they need to get back to the business use case and that, you know, they're doing intelligent monitoring or they're doing intelligent, you know, travel booking or, or whatever.

But that's a feature of, you know, whatever their actual business use case Is. I don't want to bring it up, But Metaverse, I just wanna Sure. I understood what angle you're coming at It from.

So my, my primary angle is that AI is not a product. Mm-hmm. AI is in a particularly generative AI is a feature to be used in a product.

Right. Um, coming back to cloud field day 20, Bobby, Bobby Allen, right. AI is not the thing.

AI is the thing that makes the thing better. Mm-hmm. It's a tool.

It's, it's a tool. It's a feature. It's a, something that we're going to use.

And that's why traditional machine learning has, has disappeared. John, are you seeing lots of products that actually have generative AI in them as, as a feature are, we're not there yet. No, not at all.

I, I, I feel that no matter where you go, you walk down the conference, you see the latest website or something, and there's an AI word in it, right? And they're just, we're doing ai. No, you've been doing automation behind the scenes.

You've been doing a feature of it. Uh, but you're not doing ai. Everybody takes in and thinks about AI and they think generate AI right away that we, we, like Ned said, we've muddied the water so bad that we're all assuming it's Generat ai, we're going to be doing something.

We're gonna be asking questions and getting the answers as pulling it from the, you know, uh, the LLMs and just pulling all the information. But in actuality, they have been doing a form of AI for a while. They just threw it on there for their marketing.

And I, I think we need to kind of get away from just calling everything ai and really what are we doing with that? Uh, Eric mentioned that intelligence, we're doing this thing that's ultimately helping you better. And it's not ai.

We're just doing some automation and intelligence behind the scenes. Are we ever gonna, not everybody needs ai. No, not everybody needs generative AI in their product, Especially generative ai.

I feel like a lot of it is just getting, is this getting rushed out because they want to have it on their product? Like Yeah. Your point about that all the cloud providers have the chat bots.

Does anyone ever actually get any use out of those things? I feel like they're generally useless. Right?

It takes so long to, to type in that the, and you're talking to it and it's still the AI thing that they had originally. It's not generative, right? Oh, please press one for this prompt list.

Like, okay, did you please select from these options? The One that, or even the ones that teach you like, okay, how can I do this in, in, in Cloud X? Like, how do I build, how do I build this in a WSI don't find those things to be particularly useful compared to just reading some good documentation.

Mm-hmm. Now this is the interesting challenge. We are, we actually, the real audience anymore, like the people that are gonna be most profoundly affected by AI are not the people in this room.

Like, we're gonna ride this beautiful wave to whatever's next, just like we did with cloud and OpenStack and Well, I did anyways, and, uh, sorry. I was the only one that was team OpenStack. I tried, uh, and hardware and whatever.

I mean, you come out of fab and, and chip world here we are talking about cloud and ai. We'll survive it, but 7 billion people will be affected by it. And far more impactfully than we will in this room.

And drop the mic. We're out. I do think that we need to, I wasn't kidding, Ned.

I do think we need to think about, we're talking about how to manage clouds, right? I think that there are a ton of use cases for generative AI across industries, especially where people use a lot of words. Um, you know, customer service.

These are not, these are not these use cases that would apply typically to IT environments. Um, we know that machine learning has been integrated into a lot of cloud control, uh, especially within the largest of clouds. Um, I do think that there is pressure on companies, especially companies who are seeking valuations to associate themselves with AI because, uh, VCs are providing better valuations for AI related companies.

It's just the truth. So we should expect to continue to see this. Alison, I love that you touched on customer service and that use a lot of words because a lot of the rept, the companies you talk to, oh, we have AI in our customer service now.

No, I still have to say representative a hundred times and I don't get representative. I go back to the beginning thing. I literally want to tell you, you know, today I would like to actually talk to somebody real Sure.

We'll get you that. No, please. So that's exactly right.

Use a lot of words that say they're using ai, but they're not really using ai. They're just using their automated system to pull out keywords, Bring back movie phone. But because you pick up the phone, it already knows what you really wanna see and it just tells you go see Batman.

It tells, even waits like it's goes see this one, it's playing at seven. I do think that content creators, of which many of us are, um, are one of the earliest markets for, uh, adoption of generative AI at my company. We use it every single day, all day long.

Um, creating, um, images, you know, trying to create efficiencies in terms of content generation. Um, I think that that's a very important aspect to, to think about how this, this is, this technology applies. It's just like, um, vi visual recognition would only apply to the industries where that is valuable.

It's not that we can just say that every single industry is going to change because of that particular technology. So maybe the question that we should be asking is how do IT departments change, if at all, based on gen ai? What, what, what's the answer?

Yeah. And, and which parts of the drudgery? The repetitive tasks, the things that don't bring joy to most workers.

Mm-hmm. Which of those things can we take away with the ai? One of the most interesting Developing portions of AI is, and I don't like this term, but it's the one we picked, AG agentic AI was, oh, you had to go there.

You're so sorry. We're gonna need another half hour tack on the back of this one. I was having a very interesting conversation, uh, with one of the people who wasn't their presenter on John Capa Bianco.

He was with Selector, but he didn't get to present. Um, we were talking about how he has been testing out AG agent AI on his local machine. And he told me a story, I'm sorry I'm stealing this from you, John, but he ha asked the AI to copy a file from its local system to a remote system and didn't really give it much in the way of instructions on how to do it.

So it attempted to use SCP to send the file over, but then quickly realized that when you use SCP, it prompts you for a password. And the agentic AI couldn't type in the password. So instead it downloaded a utility that will inject the password into the process and then use that utility to inject the password into the process and got the file transferred.

That is wild. And if you think about the drudgery of what it workers need to do, if there's an agent that can accomplish a task for them and they don't have to be so procedural about it, that is going to make their lives a lot easier. Of course, who gets the blame when it does it wrong?

When it breaks production? You can't point at the ai, you point at the person who kicked off the agent. Yeah.

And then we are heading towards that, uh, singularity where the machines take over the rule. Thank you very much for downloading a replacement for your owner. And, uh, here comes also spreading that, that AI from the, the source machine where you are running it to the destination one.

So of course the, the AI has to be on the destination to be able to confirm the transfer is complete. It's, it's only logical. Yeah.

It is only logical. I never liked the word reason or think like reasoning is such a, it's a very tricky word in what we talk about, what AI is doing and, and agents agentic AI is getting closer to that work and least sense of like the realm of capability of another agent. And it will try, ask it to do things and it does, you know, work around some stuff, which is pretty interesting.

And then the irony is that they become very good at reasoning, but the people that built it, like probably wouldn't pass a touring test. There's incredible amount of mathematics behind the stuff to then try to stimulate, you know, reasoning. But then as soon as you get over to mathematical reasoning, it falls apart, which is interesting.

You know, I start to think like, what bloody meat, where are we that we can just be replaced easily with for the task-based stuff yet, you know, the really complex stuff is there. So then it becomes where do we use AI most effectively? There's the, the classic example of the way you trip up a generative AI to do math.

And you say, you know, I, I picked seven apples from one tree and six apples from from another tree. Three apples were red. How many apples did I have?

And the AI says, well, 7 6 3, clearly you had 16 apples, but we all know I only had 13. Sounds like there were the us uh, the British Postal system treated this postmasters. But that's a whole other count of worms that probably isn't relevant to the US audience.

I did want to circle back again to Allison's starting point of AI actually changing the way clouds are managed because we know we see a huge amount of complexity in the environments that are being built on clouds. And that seems like a place where replacing humans with something that understands large groups of very structured data might be a good thing. Has anybody been using any, well that'd be ML tools from managing cloud.

Well, I know Mike is big into the finops portion of things. Have you seen useful implementations of AI and ML for the finops movement? I can't really say that I've seen that so far.

I don't know what John's opinion is on that, but I don't, I mean, I feel like that's where everyone wants to go, right? And it's kind of, there's, they always talk about the two sides of it. There's the, um, AI for finops and then there's finops for ai, right?

Which is like, how am I gonna optimize, how am I gonna optimize the AI spend versus, you know, using finops processing on the, on the cloud spending data. So, right. I haven't, I don't know that I've seen really groundbreaking developments in that area so far.

There hasn't been anything that's really been implemented company-wide or world enterprise wide, while other companies are investing in it and looking into it in order to say, what are my top spends in this region for the environment? And real generative AI pulling and, well, not real general ai because it's not, it's AI and real time data, but like Mike said, the biggest discussion in the last year has been, can we use AI for finops? And what about finops monitoring our AI usage?

Yeah. Because is it worth it? Is it the thing, so there's two sides of the coin for it.

And is the, is it cost effective to use it for finops? Because ops is all around, Let's ask ai, It's like 80 plus 80 AI says keep it because else you'd kill me. So I need to, I need to stay living.

Do not downsize me. Yes. Literally, they're the only ones that are exclusion.

The multiplying factor with using AI for anything is huge. Like how did we get to where we are? I think like as humans, we all came together as a group because we were, well, a lot of us came from the VMware ecosystem and how did we get better at it?

We've met with a bunch of other people in a room who had done similar things that we learned from, oh, almost like a an LLM where we continued to add knowledge to the room and we could pick and pull from it. But that, that was a very human hand-to-hand type of transfer of knowledge. Now I can go in and all of that knowledge about how thousands and millions of ways you can build VMware infrastructure.

It's inside all sorts of LLMs. So I can ask it a question in fast path to the next thing, but there's also an opportunity to go and say, what are the learning steps I should go through As I look at, I wanna learn about how to use Lambda versus how to use EC2, I can go to this tool and have it really guide me towards, and it's based on how other people have done things in reality. Mm-hmm.

And I would never have access to that level of learning. And it sure it's wrong is you're gonna give us any bond for a bunch of different roads, but so will the person at the, at the AWS user group, 'cause how they did it may not be the right way. It's gonna be a way.

So I think that's a huge multifactor, Oh, sorry. K go ahead. Hey.

No, I was just gonna say, um, yeah, one of the ways I tend to use, um, ai, I don't know, I have like this battle because a lot of times I feel like when people say ai, a lot of people I know who are not, you know, in the tech space, they instantly think of chat GPT. I think that's, that's something as well, um, where there's just a lot of training. But, you know, I like to use things like chat GBT and um, Gemini and other stuff like that and thought and whatnot, especially chat GBT.

Um, I use it to create my own form of learning. You know, to to your point where I, I sprayed up, just create my own learning path. 'cause I have the way that I like to learn.

So one of the things I do with stuff like, you know, um, code theft auto. So I teach people how to terraform, which is really just teaching myself terraform by incorporating a grand theft auto theme to it. And then I turn it into a tabletop game because in college I played Dungeons and Dragons and so it's just like this whole thing.

But, um, that's how I like to use it. But I do want AI in cloud to do things like sometimes I'm running into a networking issue and I would love to be able to type in and say, Hey, let's take AWS with, uh, Amazon, uh, developer with Q and I'm just like, okay, you see I'm getting this error. You see what the error is?

Okay, well just go fix it for me and then I can go do these other things. That would be nice. Q is nice because it has the context of your live environment within it.

So there's an advantage that they have because they own access to your tenant data. So you can actually get really good rag level of intelligence because it knows what you're actually running in your environment. So there's added context.

What we then don't get though is now this is the competitive play. How do I do this in Azure versus AWS versus Google? And what's the AI that knows that answer, right?

So they're all different levels of lives between the, the, the bunch of them. But I think that's each individual one. If you stay within an ecosystem, it's gonna be very cool.

But I do like the idea that there should be a broader, almost a democratized shared place where we know there's less chance of influence towards a bias. Mm-hmm. So in the, the coding assistant and beyond that software development AI assistant Microsoft copilot is, is the, the big one in market that has interesting reputation.

Uh, but then all of these tools have an interesting reputation. But to satisfy, Eric, I am blanking on the name of a company that briefed me a little while ago that has a, um, cloud and vendor and neutral software development AI for the entire software life cycle, including the, the sort of, uh, more expert system of when I start a new project, just lay that out for me or lay out how I'm actually going to approach building my entire application. The pipeline beyond it, the authorized it, it isn't just, uh, I type, uh, a prompt in, in comments in my code and it tells me how to, to uh, connect to an S3 bucket for the example that I literally taught in, uh, AWS training courses.

Uh, so yeah, they can be much higher and much more diverse places We used to make fun of it 'cause we would say like, don't use chat GPT because it was just trained on Stack Overflow and Reddit and it doesn't really act like a human. I said, you just described my lead developer. Like that's, there's nothing wrong with that.

It's just that it did it at scale. Like we, we can ask at things and it can be wrong, but as long as it's right enough that it leads us towards something, I just don't take it at nothing. You should read out of a lot of these inputs when generative AI is in place, is defacto truth.

There is no objective final answer. Well, it's sort of like we used to tell people not to use Wikipedia. Right?

But now, I mean, that's the first thing that I go and look at when I'm looking for more information. That's right. And then I'll follow the links to the additional references.

But that's my like first stop. And it tends to be far more accurate than if I ask somebody sitting next to me about it. You know?

So I, I think coding assistance, are they always gonna get it right? No. But neither is a human.

And if it can increase how quickly I can develop software that helps me manage the cloud, hey, brought it back then that for me is a net plus. The next step beyond that is to share that information with others. So having a virtuous cycle where, oh, I just figured out a much more efficient way to manage this portion of Azure, now I need a way to share that with the model and with other people who are facing similar tasks.

And maybe that's where things tie into open source and, you know, publishing your work and then having an AI being able to legitimately ingest that as part of a, a feedback loop. And I don't know if we've really figured out that portion of it. And I feel like that's a, that's like another hour long discussion.

And on that note, I am going to follow the Grand Tick field day tradition of saying we're out of time Fox, we could go on for hours. We're gonna stop right now. 30 more.

Thank you for joining us on this, uh, delegate round table here at Cloud Field Day 22.