Understand Artificial Intelligence Basics

0:00Hey friends, welcome back. We have another skill for you and in this one we're going to get to the fun stuff.

0:05The stuff that everyone's all hyped about. We're going to be talking about AI.

0:09We're going to talk about some of the benefits but importantly we don't just have to talk about

0:13the cool exciting stuff, we have to talk about how we can use AI in a responsible manner.

0:18So with that said, let's just kick things off and talk about how we can use AI to help us with our

0:23code development and our automation development writing scripts, that type of thing.

0:28So it won't be any surprise to you to know that we can use these AI systems to write code,

0:32to write Ansible playbooks, to write NetMeco scripts, that type of thing, because we have

0:36seen it already within this course. Now in our case we've seen things like using ChatGPT to

0:42write Python scripts, we've also used Clod, but quite honestly there is a whole ecosystem

0:47of these types of tools. We've got things like GitHub Copilot for example and quite honestly

0:52there are a lot more tools than this let me tell you. Now regardless of the tool that you select,

0:57ultimately they can all kind of do similar things. They can help us write codes, so that could be

1:03just a brand new creation of a script to change the OSPF configurations across your network.

1:09The ChatGPT, the Clod instance, whatever you're using, that can help you do that. But what it can

1:13also do, it can also debug that particular piece of code so if something is wrong and you're not

1:18quite sure how to solve it, the AI systems can help you out there. And in fact, to be honest

1:24with you, at least in my opinion, the debugging feature is one of the most useful features of AI.

1:29Gotta say, it really does save you a lot of time if you know how to use it correctly. So that could

1:34be debugging code that it originally wrote for you, which is kind of comical, or it can debug

1:40code that you have written yourself, whatever it may be, the point is we can fix these problems.

1:45And by the way, in a similar sense, we don't just have to fix particular problems, maybe it just

1:50might be that we can improve our code, for example. You know, it still might work in its current form,

1:55but it can be improved. Once again, having this AI and its ability to review what you're doing,

2:02kinda is super, super helpful. So there's so many wins here, okay? Alright, so we all know the old

2:07way in which networks were managed. What we would do is we would use like something like SSH to

2:13manually log in to a device, we can make configuration changes, log out, and then manually log

2:18into the next one, type in the configs for that, so on, so forth. And then came the era of automation,

2:24made things so much easier, whereby we could ultimately use these Python scripts, use things

2:29like Ansible to connect to all of our devices at the same time, and deploy these changes. That said

2:34though, we still had the bottleneck of actually writing these scripts, and this could take quite

2:39a bit of time, depending on the type of automation that you're doing. For example, if you are using

2:44Yang-based network automation, it's super powerful as we have already seen, but it can be a little

2:50bit difficult to work with. It can take a lot of time in order to get scripts working exactly

2:55as you expect. And quite honestly, it could transpire that writing code might take you days,

3:01weeks, and maybe even longer, depending on what it is you're doing. So this is where we come to

3:06one of the main benefits of using AI systems and network automation. Just like the way that network

3:12automation sped everything up, the use of AI ultimately speeds things up even further. It allows us

3:18to work at a quicker pace. For example, AI can tell you the type of libraries you should be using.

3:23Maybe the code that you're writing is going to have a lot of common boilerplate code that is

3:28not really inevitable, but it does take a lot of time to type out and, well, to get onto the page.

3:34Not the case when you're using AI. AI is so used to this type of stuff, it's trained on it, so it

3:39really is not much of a problem for AI to spit out this code at a rate of knots, okay? It really can

3:45speed things up. All right, so the speed factor, that's probably maybe the most obvious benefit of

3:50using AI. You can almost feel it in your bones, can't you? When everything is working in your favor,

3:55it almost feels effortless. Now that said, that doesn't mean it's the only benefit to be had with

4:00AI. For example, we also have the benefit of accessibility. Because here's the thing, when you

4:06happen to be introducing concepts of network automation into your team, it might not be the

4:11case that every engineer on your team has experience with Python and Ansible and Yang-based

4:17data modeling, for example. Ultimately, with AI, you kind of have this crutch. Maybe that's

4:23a negative connotation, but it's not meant to be. It gives you this support system. Like a good friend

4:28you can ask for advice to help you through a problem, this is what AI gives to everyone on the team.

4:33So maybe they don't have the most extensive Python experience, but they can still ultimately

4:39write a Python script, or at least evaluate a Python script, just by having a fairly decent

4:44understanding of the tool or task and using their natural language to communicate with an AI system.

4:50So at least in my opinion, I think that's really reducing the fear factor behind automation, whereby

4:56when network automation first came on the scene, there was a small subsection of network engineers

5:01who took that concept and just ran with it, whereas many other people just didn't really want

5:06to touch it. They were kind of terrified of it, to be honest with you. They did not want to use

5:11automation to make configuration changes, because if it broke the network, well, it kind of spelled

5:17disaster and they did not have the confidence to use these particular tools. Not so much anymore,

5:22in fact, the next benefit kind of leans into the previous one. When we have AI systems, it makes

5:28it a lot easier to learn a new technology or a new tool. I mean, I'm sure you've used it yourself

5:33when you're trying to understand a particular technology, the AI system can give you an answer

5:39about the very specific query that you have. Maybe the question you have in your mind is not

5:45thoroughly documented, but it doesn't make sense to you, because you can talk to the large language

5:50model, the AI system, give it all your context and then highlight what your problem is. Ultimately,

5:56that can help you solve that problem and will learn that tool, be comfortable with that tool,

6:00which again, like I say, leans in to the second benefit. It gives people that accessibility.

6:06Now, ha, it's not all roses here, like I said, when it comes to AI, there's a few pitfalls we have to

6:13be aware of and we're going to be looking at these pitfalls in more detail throughout this skill, but

6:18for the moment, let's just park that thought. Without going into too much detail, let me just

6:22say at a high level overview, you may have issues with things such as basic security when it comes

6:28to your AI automation. You also might have issues with data privacy, which is a very serious topic.

6:34In a similar sense, if you are using AI systems to develop your own proprietary software, you may

6:41actually have an issue with the intellectual property ownership of that particular tool.

6:45And again, we will expand upon each of these subjects later on within this very skill. Now,

6:50before we get to that, let me highlight one more thing. Within this particular course,

6:54we see a lot of focus on things such as CI-CD pipelines. Honestly, it's been a really heavy

7:00focus of ours because it's a really heavy focus in the world of network automation. And I just

7:05kind of want to draw your attention to how AI systems actually fits into this type of architecture.

7:11One of the most important components of building a robust CI-CD system is actually choosing the

7:16correct tooling. Just as we mentioned before, which type of tools are we going to be using to

7:21automate our network? Which type of protocols are we going to be using? Is it going to be RESTConf?

7:26Is it going to be the CLI? Is it going to be gRPC? When it comes to building out a responsible and

7:32robust automation system, the whole design of the process, the whole architecture, really requires

7:38great consideration and AI can really help you think through this particular problem.

7:42And in a similar sense, when you have selected the tools and you have selected the configuration

7:47style that you want to be implementing, AI systems, as we have already seen, can help you

7:51build these particular scripts and build these particular processes. And then also, when we

7:56happen to be making these changes in our CI-CD systems, we can also validate these changes via

8:02the use of AI because AI can help you analyze log outputs. It can help you identify error messages.

8:08Basically, what I'm trying to highlight is that the use of AI tools is not just beneficial in this

8:12small, isolated manner. Ultimately, AI systems help improve and stabilize the entire life cycle

8:19of our automation operations. Now, like I said, it's not just all benefits and roses and happiness

8:25with the use of AI. There also are some serious risks and considerations to think about. So,

8:30friends, how about we highlight some of those risks in the very next video.

0:00Hey friends, welcome back.

0:01So in this video, what we're going to be doing is talking about some of the risks, the potential downsides of using AI.

0:08So I suppose the first thing we will talk about is data privacy.

0:12All right, so let's imagine this.

0:13Let's just have some network configurations and let's imagine this as a real live network, okay?

0:18Not my lab environment, okay?

0:20So let's just do a show run and get the running configuration.

0:23And all I'm going to do here is just highlight it all and copy it.

0:26Okay, so copy here.

0:28Now, let's just say I go to a public instance of an AI system and I say, here's my configuration on my Cisco switch.

0:34And what I'm going to do is just paste it all in, okay?

0:36And then I earnestly just ask it something like, do you see any configuration errors or problems, okay?

0:41So I'm getting it to study my configuration.

0:43All right, so it thinks for a bit and it says here nothing jumps out as a syntax error.

0:47It looks valid and should boot fine.

0:49And it gives some suggestions about potential things I could do.

0:51Now, this is not the issue.

0:53This is why someone would like to use an AI system to get this type of feedback.

0:57But what actually is the issue?

0:58Well, this is a public instance of ChatGPT and I've just dumped in all my network configurations.

1:04Now, like I said, this is a pretty bare configuration, so there's not much on it.

1:07But imagine it was a real live production environment.

1:11I may have just potentially told it some sense of information.

1:14Maybe we've got some SNMP community strings in here.

1:17Maybe if my security is really bad, I've got some unencrypted passwords that I've just sent to this AI system.

1:23Maybe I've divulged the internal IP addressing in the system,

1:27my VLAN configurations, my access control list, my security posture,

1:30my routing protocols, so on and so forth.

1:32Now, if this here was just a conversation between me and a private system, a private AI system,

1:38whereby the data here was just, well, just private to me and not being shared to anyone at all ever,

1:45well, that would be fine.

1:46But that's not how these tools work, unfortunately.

1:49Very often, by default, the conversations that I'm typing here is used to, well, improve the actual training data.

1:57Because when you think about it, especially on the free tier version of these particular utilities,

2:02these tools are so expensive to run.

2:05I mean, like, seriously, the amount of energy they use is astronomical.

2:09And if they're just giving it to you for free or for, like, $20 a month or something,

2:14you can be pretty sure that there is an upside for them.

2:17The information that you are giving these particular tools and these particular companies,

2:22they're using it to train their models.

2:25So what that might mean, therefore, is that the configurations that you have on your live network

2:30may actually be the input today, but maybe in the future it is the output for someone else's request.

2:37Now, what I will say is that many of these tools happen to have an enterprise-style offering,

2:42whereby you pay more money, you have a type of business account,

2:46and because of that, the information, the way it's processed, is a lot more restrictive

2:50and it's a lot more secure.

2:51But the point is, is that when you happen to be using these particular tools,

2:55you want to make sure, make sure absolutely,

2:58that the information that you're sending is not being sent somewhere else

3:01without you knowing about it and without you actually understanding that.

3:05Because, you know, maybe some things that you send are not sensitive at all, and that's fine.

3:09But if you are completely unwitting about this fact,

3:12you might be divulging really sensitive information.

3:15So really the main thing to highlight here, especially, especially in a commercial setting,

3:20whereby this is you working in the purview of your job as an employee,

3:24if you really want to be understanding your company policy,

3:27if they tell you, hey, by the way, never, ever, ever put our configurations into a public AI system

3:33like ChatGPT, then definitely don't do it.

3:36You're opening yourself up for a lot of problems and you can be compromising your security.

3:40This might be what some people comically refer to as a resume generating event.

3:45Which means you'd be looking for a new job.

3:47So just don't do it, okay?

3:49So be aware of your policy.

3:50The next thing you can also do if you're going to be interacting with these systems

3:54is to actually sanitize the output.

3:56So say, for example, as opposed to just blindly pasting all the configurations,

4:00look at the unique data like here, like the IP addressing.

4:02How about if it's not necessary for the actual problem

4:06to actually divulge the IP addressing, for example,

4:09why not just sanitize it before you just paste it in blindly?

4:12Why not just go and, you know, change this information

4:14or deliberately truncate it and leave it blank?

4:16The AI system, if it doesn't need to know about it,

4:19it's not going to somehow impair its ability to solve a completely unrelated problem.

4:23So sanitize it.

4:24So again, pasting this in.

4:25Okay, let's just maybe say version.

4:27Let's just leave that as version redacted.

4:30Down here, hostname.

4:32Change it to my hostname.

4:33Username, John.

4:34Privilege 15.

4:35Secret.

4:35And it's got the actual hash of the password here.

4:37Change that to redacted.

4:39Same again here, redacted.

4:40And you can just keep doing this on all the sensitive data.

4:43Now, again, this is maybe not the quickest way to interact with AI,

4:47but it certainly is, it's a more secure way

4:49as opposed to just giving away the keys to the kingdom

4:52with respect to your sensitive configurations.

4:54So really know what the company policy is.

4:56If you have the ability to use the more enterprise version of these tools,

5:00whereby we have stronger privacy guarantees,

5:02definitely do that.

5:03And where you can, be sensible about replacing the sensitive data

5:07that you happen to be inputting into these systems.

5:10All right, so that's just some basics on data privacy.

5:12Let's now talk about something known as IP ownership.

5:15So let me go and give you a really, really simple example, okay?

5:18Let's just say,

5:19write me a network automation utility that tests OSPF.

5:22Let me hit enter.

5:23Now, as it stands right now,

5:24this is going to go and spit out a fairly simple and generic script.

5:28Now, that is not really going to be such a concern for you, okay?

5:31The problem, though, is if you try to develop this

5:33into some type of proprietary software,

5:36whereby you claim ownership of it,

5:38and even worse, if you try to monetize it and sell it,

5:42if in the event that you've just had an AI system

5:44generate this particular tool or large blocks of this particular tool,

5:48it may actually be the case that you have no valid claim to it at all,

5:52because all you did was just ask an AI system

5:55to go and create or mostly create it.

5:58And that, my friends, is not the same thing

6:00as actually creating something on your own.

6:03Now, what I will say is straight off the bat,

6:05I'm not a lawyer,

6:06and this very topic, it's quite contentious at this time of recording,

6:10so there is no clear path as to how to solve this particular problem.

6:14So I'm not going to go and pretend I know the exact legal bar

6:17whereby something becomes your ownership

6:20versus something that you cannot actually copyright.

6:22What I'm trying to highlight more generally

6:24is just be aware of this particular stuff.

6:26So if, for example, you just blindly copy this AI code,

6:30don't be surprised that the new tool that you have supposedly

6:33quote-unquote created is not actually yours

6:36and can actually even be used by a competitor,

6:38and there is nothing you can do about it.

6:40You'd have no claim to say,

6:41this is mine, you can't use it.

6:43If you rely just on AI

6:45and you're not making sufficient modifications,

6:47whatever that legal bar happens to be,

6:50just be aware that this is a potential pitfall.

6:52Now, that said, if you happen to be the main designer of the architecture,

6:56the main person writing the code,

6:58AI is just maybe assisting you here and there,

7:00modifying things a little bit.

7:02Well, you can imagine you would have a much stronger claim

7:05to ownership of that material.

7:07And, by the way, here's another one.

7:08See the type of code that is generated here?

7:11The way that the tool is so good at writing this particular code

7:14is that, well, many of it has been trained on open source material.

7:18Now, much of that original code that it was trained off of

7:21could actually have been written under a particular public use license.

7:25So the actual code that you're using

7:26that has been derived from original code

7:29that had a public use license,

7:31it might still carry that same licensing.

7:33Again, this is still not quite clear

7:35as to how all of this is going to play out.

7:37So, really, what I'm saying here is my advice to you

7:39is to be very, very careful

7:42when using these particular tools

7:44to generate and create what you think is custom and original solutions.

7:48So, friends, as we know, AI, absolutely awesome.

7:51Two thumbs up. We like it a lot.

7:53But as we can see,

7:54there are some potential pitfalls that we have to understand.

7:58One of the big ones, data privacy.

7:59Again, this is a very, very easy one to fall foul of.

8:02People make this mistake all the time.

8:04Do not let it catch you out.

8:05And, again, especially in the enterprise environment

8:08whereby you might be creating your own tools,

8:11including, by the way, what might end up being commercial products,

8:14remember to keep in mind the potential issues of IP ownership.

8:18In fact, you know what I should do?

8:20I should probably clarify what IP actually means.

8:22I'm just talking about intellectual property.

8:25This thing that we created, this thing actually belongs to us.

8:28And if anyone tried to use it without our authority,

8:32well, ultimately, we have some legal standing

8:35to challenge that and prevent those actions.

8:38So, friends, those are just some of the potential issues of using AI.

8:41How about we talk about some more in the very next video?

0:00Hey friends we're back we are back with a vengeance and in this particular video

0:04what we're gonna do is talk about one of my favorite not definitely not one of my

0:09favorite maybe one of my what you say pet peeves one of the things that

0:13annoys me the most about AI and this is the subject of hallucinations. Now if you

0:17happen to have experience with AI no doubt this has probably driven you

0:21pretty crazy at least well now that I'm used to it it's not so bad I've kind of

0:26half expect it but when I first got into the whole AI thing to be quite honest

0:31with you it was a little bit too trusting and that ultimately leads to a

0:34lot of frustration and confusion so we don't want that so what we're gonna do

0:38in this particular video is to talk about what hallucination is and of the

0:43importance of validating what the AI system actually generates for us because

0:49here is the thing if you happen to have used AI quite a bit no doubt you have

0:53experienced that AI can be super super confident you know for example you ask

0:58it to do something for you and maybe it gets something that you can clearly tell

1:03is wrong then you point that out and it says you know what you're totally right

1:07I'm sorry this is what I did wrong here is a solution that will definitely work

1:11this time and the first time you hear this it's quite reassuring you think wow

1:15this is really really good it must have solved a problem this time and must have

1:19thought very deeply about it this thing here this new corrected version is

1:24definitely going to work and then it doesn't work and you're left annoyed

1:28this is just the nature of many of these AI systems they are designed to sound

1:33confident but ultimately what they're doing is probabilistic reasoning and

1:37that is quite hard to say sometimes probabilistic so what that means

1:42therefore is that they are trying to get the what would you say the thing that's

1:47most probable from a statistical point of view so by its very nature it should

1:52be a little bit uncertain about the nature of the response that it gives but

1:57that ain't the case very often they'll just tell you this answer is absolutely

2:01certain and you should just trust it apparently but definitely don't do that

2:05because here is the thing in the world of networking and particularly network

2:10automation whereby you can use these particular generations to automate out

2:15across a large amount of devices making many configuration changes you can

2:21ultimately black hole your network if you're not actually certain about what

2:25the outcome is going to be from this code that was generated so that could be

2:29things like misconfiguring access control lists it could be misconfiguring

2:32routing protocols there's tons and tons of ways in which these systems can get

2:37things wrong let me give you an example okay one thing you might notice with

2:40respect to AI systems is that it might just get the syntax wrong maybe the

2:44training data was based off of older versions of the tool that you're trying

2:49to be using so maybe the training data uses an older version of Cisco iOS XE

2:53and the newer version has a slightly different command or a slightly

2:58different implementation if the training data is all on the old version then the

3:03AI system is going to statistically generate an outcome based on that

3:07version and then present it to you as this type of perfect solution almost you

3:12would be wise to always have a good healthy dose of skepticism when

3:17interacting with artificial intelligence next thing and this is the topic that we

3:21kind of kicked off this video about and this is the concept of hallucinations

3:25now if you're not certain what a hallucination actually is what we're

3:29talking about here is when the AI system happens to just make something up and

3:33pull it out of thin air something that doesn't actually exist so how could this

3:37present itself in the world of network automation well what it might do is it

3:41might invent a non-existent function inside of a Python library the way you

3:46actually talk to the AI system can kind of have an effect on this well I mean

3:51when I say that well if you happen to talk to the AI system in this way that

3:55almost expects a particular response it's almost like it freaks out it can't

4:00actually find an authoritative source on what you're expecting it to do but it

4:04can tell by your tone that you are absolutely expecting this particular

4:08thing to be there so it's just going to go and invent it out of thin air let me

4:12give you an example okay all right check it out what I've written here is a lot

4:14of rubbish see this particular module doesn't actually exist totally made it

4:19up but I sound pretty confident okay because I've specified a very particular

4:24version I've asked something that sounds really quite technical I mean LFIB

4:28does relate to MPLS and resconf is something you would use with an actual

4:32yang model so it kind of sounds like it should actually have an answer but it's

4:37just gibberish let's see though what the AI system happens to say because it

4:42sounds like I'm expecting a particular response let's hit enter so it's

4:45searching the web give it a moment and look at this right here now it says this

4:51is one of those areas where the iOS XE yang gets very under documented well I'm

4:57pretty sure why it's under documented this doesn't exist this model so it

5:01shouldn't actually be documented at all key point first so you don't chase

5:05ghosts okay so it's trying to help me out here it says here and this

5:09particular XE version this model does not expose LFIB adjacency statistics

5:14directly as a flat stats container instead it follows the MPLS model

5:20hierarchy now it doesn't do this because this here doesn't even exist so then we

5:24scroll on down we get a nice big green tick about the actual path to be using

5:28with this non-existent model and it just goes on and on and on and on so again

5:32this is obviously really annoying but notice the trend I sounded really

5:37confident the training data was obviously quite limited because this is

5:41quite a technical subject but if we don't have some type of expertise here

5:45to actually recognize that this here is all completely wrong you may be

5:49possibly putting this stuff into your network automation solutions and this is

5:52definitely not a good thing at all now let's be a little bit mischievous but

5:56we're doing it for educational purposes I'm gonna flat out confess that this

6:00model doesn't exist and I'll ask it why it told me these paths and then it's

6:03gonna say you're right to call that out that was incorrect there is no model of

6:07that type so now it's giving me a little bit better data but if I happen to change

6:10this and say the model should be called and all I'll do is paste in the model once

6:15again and let's add on like LFIB at the end something that sounds quite

6:18technical and roughly plausible okay so let's hit enter see what it does and

6:22look at this you're absolutely right the correct model is this I just made this

6:27up right there and now it's saying that's the one that actually backs the

6:31show MPLS forwarding table does it really not at all and then off it goes

6:35generating all this nonsense and what I'm trying to really stress to you is

6:39that you can't just be relying on this stuff this is why expertise is still

6:43very very important because if you don't actually truly understand this stuff and

6:48you just try to spam these solutions you might be really confused as to why your

6:53device is just giving you client errors all the time now one thing we can do to

6:57actually improve this we can use something called retrieval augmented

7:01generation and what this can do it can kind of constrain what your queries are

7:06actually going to be in essence what we're talking about here is grounding

7:10the actual answers with particular context so once again let me give you a

7:14quick example what I'm going to do here is open up a new chat and what I'm going

7:17to do is to ask it something quite vague so I'm saying which HTTP method and URL

7:22should I use to associate a site with a network profile now here is the thing

7:26see the terminology I'm using here this is quite specific to Cisco catalyst

7:31center so I should expect the AI system to perhaps generate some information

7:36about this but it might not get it correct so let's go and enter here and

7:39see what it says so it has recognized that this is catalyst center terminology

7:43and what it's done here is that it's giving me the actual URL or at least it

7:47purports to tell me that this is the correct URL now what I'm going to do

7:51here is I'm going to go and check that and verify this to make sure that this

7:54actually is correct so all I've done is search for the actual documentation

7:58right here go and click this and I'll get down to the API reference and let's

8:03click on API right here okay now the section I'm looking for is in site

8:07design so let's click this right here and if I scroll on down I can see this

8:11option for associates and as we can see right here this is going to allow me to

8:15associate a site to a network profile now look at the exact wording I had

8:20right here okay this is basically why I asked the system right here associate a

8:24site with a network profile but look at the endpoint they're saying it's going

8:28to be site site ID network profile then network profile ID where in reality it's

8:34actually going to be network profile network profile ID site site ID that is

8:39the endpoint we need to use so once again it was quite plausible what the

8:42answer was but it got it completely wrong what I want to do though is I want

8:46to give it particular context by just having it reference this documentation

8:51so check this out what I'm gonna do here is go back and when I get to this I'm

8:56going to choose to download the open API document okay so it's just gonna go and

8:59download like so now I'm gonna do here is choose to save a copy and I'll save

9:03this so what I'm gonna do here is choose to create a new project and let's just

9:07call us like catalyst helper and create the projects now what I can do here is

9:11specify particular sources that I want to include and to give this chat a lot

9:15more context so let's go and add the documentation alright so now the

9:19documentation has been uploaded as a source and all I'm gonna do is to paste

9:23in the exact same question and see if we get any better responses so let's hit

9:27enter and now you can see it's beginning to search the project context and look

9:31at this right here this looks a lot better see this right here it's now

9:34specifying we're gonna have to use the post request and the endpoints this

9:38actually looks like the correct endpoint this time and in fact if I go to the

9:41documentation and compare this it should be DNA intent API v1 network

9:47profile then you put in the profile ID then forward slash site and then the

9:52site ID and it also tells us any required parameters so again if we

9:56consult the documentation if we look right here it shows us the required

10:00parameters as the network profile ID as well as a site ID and those are the two

10:04that are required so friends be aware that whilst AI is awesome it can do this

10:09really annoying thing known as hallucination and we can ultimately

10:13mitigate hallucinations by giving our AI systems a lot more constraints and

10:17context as to what we're trying to do

0:00Alright, so we know about many of the benefits of AI as well as some of the potential pitfalls and

0:06in this particular video what I just want to do is to spend a little bit more time on

0:10kind of talking about how we can use AI in a responsible and secure manner.

0:15So the first thing I just want to highlight is just some kind of questions you should have when

0:21interacting with AI especially in the setting of network automation.

0:25So the first thing which is probably at the top of the list and that is always to verify

0:30what it is you're getting back from the system and we kind of hinted at this in a previous video

0:34but what does it really mean? We're not just talking about just being skeptical and you know

0:40executing the code with a little bit of trepidation.

0:42We're talking about actual steps that we can do to help mitigate the risk of

0:46something going wrong if we choose to execute the code that was written via the AI system.

0:51So say for example, the AI system is writing some code to

0:55templatize and generate some CLI configurations. What would you want to do?

1:00Well, first thing you would want to ensure that the CLI commands are indeed the correct ones and in a similar sense if

1:07let's maybe say it was generating some type of YANG payloads and it's going to be deploying it via NetConf or RESTConf.

1:15Once again, you want to be verifying this.

1:17So in the case of the CLI you could go to the actual CLI

1:21documentation of that particular model of the device you're using and before you attempt to deploy this on a real-life network

1:27go and check that what has been rendered is actually matching up with what you would expect.

1:31Same thing is true for the YANG based automation. Ensure that the actual modules are indeed correct.

1:36Use the tools that we've talked about already throughout this course. Things like PyYang or say for example

1:42YANG Explorer, that type of thing.

1:44These are the types of tools that you can use to help ensure that what has been suggested is actually real and it's not just

1:51some type of hallucination. Now again, maybe for you this seems really really obvious and you know

1:56it's just something you clearly would do anyway, but to be honest with you many people are not doing this at all.

2:01They're just kind of blindly trusting the god of AI and

2:04assuming that it must know better than them and just going in kind of gung-ho and

2:08well, I would say taking a reckless approach to network automation via using AI.

2:13So ultimately whether it's rendering CLI commands, whether it's you know

2:17using YANG based automation or whether you happen to be interacting with a particular controller

2:22the main thing to do is to trust the authoritative resource. Go to the actual documentation and make sure

2:28what you're seeing in the AI system is actually matching up with reality. All right, next point.

2:33So let's say you have done your verification and you are highly confident.

2:36Still not a good idea just to go and deploy this directly onto the production network.

2:40Instead what you would want to do is to test this in a nice isolated environment.

2:45So that could be things like using EVE-NG for network simulation or as we talked about earlier on throughout this course

2:51you could use Cisco's modeling labs to model your topology and

2:55just go and inspect what the AI system has actually done when you deploy this code or when you run this playbook that type of

3:02thing. And another tool you could use is one we have seen already throughout this course once again is to also use the

3:08Cisco DevNet Sandboxes very very useful. Once again

3:12they give you a nice isolated and secure environment

3:15whereby you can kind of go crazy. And the worst case is as if you happen to break something

3:19you never really broke anything in reality. It was just a simulation. Now another component

3:24which is very very valuable and this is the kind of thing that's going to keep you in a job is to use your own

3:29expertise. Really spend time learning these technologies, become comfortable in them, become an expert in them because even though AI can't allow you to work

3:37quicker more efficiently

3:39ultimately you should see it as a tool to augment your own work. And if say for example

3:43the AI is going off the rails because you have expertise

3:47you are more able to easily detect that. Because here's the thing when it comes to making changes and networking

3:54particularly you may push out a change and it might seem like it has worked

3:57but you also maybe have to think about the downstream effects of the change that you just implemented.

4:03Maybe it transpires that the BGP configuration that the AI system actually told you it's using the correct format

4:10it's using all valid values, the configuration does apply and it can actually work.

4:14But maybe the AI systems are not really thinking about how that might affect your MPLS network.

4:19This is once again whereby you can shine when you understand the holistic picture and you have expertise and confidence in the technology.

4:26Then maybe when the AI system gets a little bit gung-ho and recommends changes to a particular technology

4:31you have a good enough

4:33understanding to know all the potential side effects that that may cause and you can go and check them and verify them in your

4:39isolated environment before deploying it to production.

4:42Now again, just another point to talk about with respect to actual security when using AI.

4:47Now previously we talked about these public AI utilities such as

4:51ChatGPT or Clod that type of thing and we highlighted some of the potential pitfalls that you can fall into when using these particular tools.

4:59For example, you happen to paste in your API token directly into an AI system.

5:04Maybe you divulge your inner security posture of your network configurations.

5:08As we know the real vulnerability of this as the data that you're sharing

5:12well, it can be shared and it can be used elsewhere.

5:15So one thing I just want to highlight to you is that whilst we can get these enterprise grade

5:20implementations of these tools, one thing you can also do is run your own local and private AI system.

5:26So the benefit here is that while it's offline

5:29it's just between you and the AI tool. Now the downside to this is in order to be able to run these

5:35well, these more modern models, quite honestly, you need an absolutely massive amount of compute and power.

5:41So to be honest with you, to use the latest greatest best models in the world

5:45you're kind of gonna have to go the public route whereby you have access to Clod whereby you have access to ChatGPT and their

5:51latest models and their infrastructure to run those models. That said though not every solution requires the most advanced AI.

5:59So what I want to do is I want to just briefly show you how we can install your own

6:02local AI system and we can see how that can benefit us in the world of automation.

6:07So check it out. What I'm gonna do here is to go on to my Ubuntu machine.

6:10Now what I'm gonna do here is to install a particular utility known as Olama.

6:14So let's go and open up the browser.

6:16Now all I've done here is search for download Olama and let's just click this right here.

6:20And what you're gonna see here is we have all the options to download this if it's on Windows, if you're on Linux or if

6:26you're on Mac OS. Now in my case here, I have a Linux machine.

6:29So let's just click this and we can see here the command is this one.

6:32So if I just go and paste this in and to enter, what it's gonna do is to go and begin

6:37downloading Olama onto my machine. Now in the meantime, if we just go back to the browser here,

6:41what you're gonna see here is we have all of these particular models that can be run on Olama.

6:46So if I click this right here, you can see here

6:49these are all the things that you can ultimately use. You can see here

6:52we have tons and tons and tons of models. Now one that happens to be particularly powerful for coding is this one right here,

6:59the Quen 3 coder. But one thing to note here in order to run this at its most powerful,

7:04this has 480 billion parameters, which requires an enormous amount of power.

7:10So what I'm saying here is that when you happen to choose these models,

7:13you really have to consider your environment and what you can actually power. Now a good generic use model

7:19that is not so intensive that can be run on a regular laptop is this one right here, Mistral.

7:24So what we're gonna do is once Olama is ready,

7:26we're gonna go and pull this particular model and use it. And even though it's not the most powerful,

7:32it should be enough to kind of get us going in our network automation scripts. Now straight away here,

7:37we get this warning and this is because I have downloaded this in a virtual machine on my system.

7:42So we're not getting access to the GPU,

7:45we're only going to be running it via the CPU, which is going to make things run really quite slowly.

7:50So with that said, all I want to do now is to show you how we can pull in the actual Mistral model and use

7:56it. And then what I'm going to do is to slide on over to my MacBook, my native computer, and then do the AI

8:01there, whereby I have access to my GPU.

8:03So the way I can get the model is really quite similar to Docker.

8:06All I'm gonna do is say Olama pull and then Mistral and say I want to get the latest version.

8:11So if I hit enter, what's gonna happen is we're gonna begin pulling this down.

8:14And if I want to run this, I can just say Olama run Mistral, okay?

8:18So if I hit enter here and

8:20what it's given me is an error straight away saying that we need more memory than is available inside this particular virtual machine.

8:27So pause, let's hold that thought. Let's now take the same installation over onto my MacBook, okay?

8:32All righty-o, so now I'm over on my MacBook and if I just say Olama list, we're gonna see here

8:37I have already installed Mistral as well as another model as well.

8:41So if I want to be interacting with the Mistral model, I can just say Olama run Mistral and to enter.

8:47And now we drop into this command-line interface and like I say, this is completely offline.

8:51So let's just ask you something quite simple.

8:53What is the capital of England? And it tells you straight away that the capital of England is London.

8:57So it seems to be working. Now here is the thing.

8:59What I can also do is use the graphical user interface.

9:02If I just go and open this up to click on this logo, as we can see right here,

9:06I have this little chat interface and down here what I can do is select the model that I want to use.

9:11So let's choose Mistral.

9:12I'll just say, can you write me an Amico script that will connect to Cisco iOS device?

9:17The IP address is 192.168.1.201.

9:21Username is John and the password is Cisco123.

9:26Just say send the show version commands and print the result.

9:30So let's hit enter, give it a moment and now it's beginning to write this out.

9:34Okay, so it's given me a response,

9:35but what I've noticed is that it's actually using the getpass module to prompt the password.

9:40So I'll just correct it and say, obviously this is good,

9:42but hard code the password of Cisco123 instead of using the getpass module.

9:47So let's hit enter, see what it does.

9:49So now it's beginning to correct things and now we have our code.

9:52Let's go and just copy this right here.

9:54And what I'll do here is create a little script called olamascript.py

9:57and I'll just paste this in like so.

9:59Okay, so let's just save this.

10:00And if I just run the script and to enter, let's see what we get back.

10:04And bam, we get our response back.

10:06And you know what, just to highlight this, if say for example, I disconnect my internet.

10:11Okay, so if I try to ping the internet, we're going to see here I don't actually have any access,

10:15but the cool thing is, is I can still use Olama.

10:17So check this out.

10:18Let's imagine this is now the zombie apocalypse

10:21and, you know, I don't have access to the internet.

10:24I don't know what I can do.

10:25But the good thing is, is I still have access to the intelligence of AI.

10:29So whilst I'm disconnected, I could say something,

10:31let's say write an Ansible playbook to do the same kind of thing.

10:34And if I hit enter, what is going to happen is even though I'm disconnected from the internet,

10:39I'm still now getting back my responses from Olama.

10:42So see this right here, all this is coming out.

10:44And as we can see here, it's given me all the instructions to do things like make sure Ansible is installed on the system,

10:50make sure I've got a host file, that type of thing.

0:00Alright friends so let's now validate what we've learned within this skill. So the first thing

0:03we're going to do here is to launch the lab below. So let's go and click this and then once it loads

0:08we're going to go and open up the Linux Mint virtual machine, run the command olamalist to see

0:13all the available models installed on the system. Then we're going to run the Mistro model and use

0:18it to generate a Netmego script that's going to connect to the Cisco iOS device in the lab with

0:23the IP address of 10.0.0.1. The username is going to be John and the password is cisco123 and the

0:28script should run the command show ip interface brief and print the results. Now once you get

0:33your result the last thing I want you to do is to test that script against the lab device. So

0:38friends go build up your lab give it a go and I will see you in the very next video where we discuss

0:42the solution.

0:00All right friends, the first thing we're going to do is to run the olama list command.

0:03So let's hit enter and as we can see here we have the Mistral model available on the system.

0:08So what I'm now going to do is I'm going to go and use Mistral to generate this netmicro script

0:13to talk to the device of 10.0.0.1. So all I'll do here is say olama run mistral and hit enter.

0:20Now what I will say is because this is in a virtual machine things are going to be really

0:24quite slow compared to how it would work if you had a GPU available. But nevertheless we're just

0:28going to have to be patient and allow it to run. All right also let's write this I'll say

0:33write me a netmicro script that will connect to a cisco ios device just say ip is 10.0.0.1 username

0:41is john and password is cisco123. I'll just say just hard code the credentials into the script

0:49and I'll say run the show ip interface brief command and print the results. So let's hit

0:55enter and see what we get back. All right so it's beginning to respond and as you can see straight

0:59away it's a little bit slower in typing out the response but again like I say it's a virtual

1:04machine so we're just going to be patient. Now straight away I can see here it's wanting to use

1:08the getpass module once again as you can see here these particular models are not quite as smart as

1:13the higher end ones but you know it's still going to do the job. In fact what it's doing is actually

1:18hard coding the credentials as I said but for some reason it's importing getpass even though it's not

1:23going to use it so that's a little bit bizarre but nevertheless. So let's try to correct it and I'll

1:26say simplify the script remove getpass since it's not being used and I'll say don't set the terminal

1:32length to zero and remove the fine prompt. So let's hit enter and see what we get back this time.

1:39Now one thing I will say with respect to the Mistral model it's just a general use model and

1:43it's really quite simple. It's not designed for coding, it's not designed for network automation

1:48so you can expect some mistakes. Now if you happen to have a more powerful machine at home then by

1:53all means go and run a more powerful model that is optimized for coding. Alright so it's now

1:58beginning to respond back with the actual results. Alright so it now has completed let me just

2:03fullscreen this for a moment and all I'm going to do is to go and copy all of this like so okay so

2:08let's just go and grab you and I'll just go and copy this and let's exit olama by saying forward

2:14slash by and let's just create this little script and let's just call it again like olamascript.py

2:19and I'll just paste all this in like so and let's just save it and then finally we'll go

2:24and run this script like so and as we can see the script does appear to work. So friends that is us

2:29for our validation challenge on artificial intelligence. I hope this has been informative

2:33for you. I'd like to thank you for viewing.