Who's Afraid of ChatGPT?

Machine-generated from the episode audio and not hand-corrected, so names and technical terms may be imperfect. The audio is authoritative.

All right, welcome back to another episode of the CTO Advisor podcast. com, all as cold stuff. If you think in policy as cold, infrastructure as cold, you know what, we're going to get into maybe one day content as cold. That would be kind of cool, right? Ned, welcome back to the podcast. Thank you. Great to be here. Thank you for having me. And I'm excited to talk about this conversation because it's a little inward facing, but it's also very topical.

It is very topical. You know, as of this production, as this airing the podcast, who knows, me and you might have been replaced by AI. Our jobs are done, man. ai, other large language models and chatGTP has made the need for everything from developers to network engineers, to content creators. Our work is done, man. We can retire. Sweet. I thought I had another 20 years, but it's great. I'm just done. Oh, I hope that universal basic income comes in pretty soon.

I mean, so the reason why I'm having you on talking about AI is because you had a really great tweet about how you've integrated AI into your workflow. So without getting into kind of the hyperbole of AI replacing content and knowledge workers or content creators and knowledge workers, I would love to break down how you're using AI and how you believe AI will enhance the productivity of basically not just technology folks, but knowledge workers in general. I think it's a good, broader conversation, especially for the CTO audience.

And as we're thinking about enabling entire organizations and industries. Yeah. And just to provide a little background, I'm sort of a one person operation. So net in the cloud has one full-time employee and that's me. And my goal is to create educational technical content for technical professionals. That's my goal. That's what I try to achieve. And in order to do that effectively, leveraging other people and other technologies is part of what I do. And what I try to do is farm out the stuff that I either don't enjoy doing or that I'm not particularly good at.

So like graphical design, I am not good at graphical design. I own that. That is fine. I'm happy to farm that out. Writing social media posts based off of stuff I've written, whether it's a blog post or made a video or whatever, I don't enjoy doing that. So I actually have someone that I pay part time to comb through the things that I produce in a given week and write some posts for me, whether it's on LinkedIn or Twitter. And then there's the parts that I really like, which is reading and learning about technologies, tinkering with things, figuring out how it actually works, and then writing a script or creating a post that explains it in a way that I think people will find accessible.

So that's sort of my goal and what I'm trying to do as a one person operation. The reason for my Twitter thread is I was trying to think about how AI is going to impact all kinds of different jobs. At this point, I've been out of the direct industry for four years. I was a consultant for eight years or something, and I was deep in the tech, and then I went independent. I've been doing this for about four years. So I didn't want to pontificate on what's happening with them without first thinking about what's happening with me.

And the more I thought about it, I realized I've already started using AI in a whole bunch of ways in my production process. Just without realizing it, or without really thinking hard about it. So I just kind of walked through some of the ways I'm using it today. Probably the biggest example is GitHub Copilot. I know, you gave me a look like, what do you mean you're using Copilot? Yeah. So, you know, this is really, I've heard of Copilot for like application developers, etc.

But you're more of a policy as code, infrastructure as code, you do a lot of stuff in Terraform. And I'm having trouble mentally connecting, you know, for this kind of AI-assisted for application development, for infrastructure as code stuff. I guess, you know, infrastructure as code is just code, so. It is. And Copilot is really there. It's like autocomplete on steroids, basically. So if you've used an IDE in the past, you would start writing a function by doing, and I'm going from the Java days, like you start writing public static, and then it would go, oh, you're writing a function here.

Here's a suggestion for how you would finish the form of this function, right? Copilot takes a step further. And maybe you wrote a comment about the function before you started writing the function. I can just put a comment in that says, this resource block creates an Azure resource group. And then start typing resource and Copilot will go, oh, well, this is how you do that. So it'll just fill in everything for me with the best guess. And I can accept that best guess and tweak it.

So it's kind of like having a pair programmer that's, they're not great, but they kind of know what you're trying to do. And they can do the drudgery of the actual coding for you, you know, filling in all the little fields or at least getting the formatting right. So you can tweak it to your actual intention later. So I'm a good enough application developer that I can write some really bad code, like I can, some tragically bad code, but to my credit, I actually have really great fundamentals around logic and understanding structure, et cetera.

I just hate the minutiae, like the, you know, going back to my eighth grade and seventh grade developing skill and trying to find why the syntax, why there's a syntax error in my, the code that I was writing was probably 90% of my troubleshooting, what I, what I needed to troubleshoot. I wasn't troubleshooting the logic of the application so much, as much as the syntax. So this sounds like for developers like me, it sounds like a godsend. Right. It can help you take your intention and kind of produce valid syntax for you.

And that in turn will help you write code faster. It's definitely helped me accelerate code. And so a stat that Microsoft published when they were talking about Copilot so far was something like 45% of code is being written by Copilot. And people were up in arms about that. And I was like, no, that actually sounds about right. If I'm just talking about literally what Copilot suggests to me, that percentage of code it suggests is probably 45 to 55% of the code that I write just by lines and characters.

It doesn't mean that Copilot is actually writing all that code for me. It's just taking out the drudgery of writing. So I can't create a statement or comment that says, hey, write a application that will scrape eBay for this type of auction and then reply and then auto bid for me. It's not a pilot. It is a Copilot. So how would I go about writing that type of app? Well, what's interesting is in its current form, you're correct. GitHub Copilot will not write that application for you.

It can help you write functions and basic structure and produce variables, all that kind of stuff. But Microsoft announced GitHub Copilot X. And what that adds to the whole feature and functionality is a chat GPT like interface where you can have a conversation with Copilot and tell it what you're trying to produce. And it has that context and history in the chat to help you produce and refine. So today I can go into chat GPT and say, I'd like you to create me a Terraform configuration that includes these resources go and it'll produce it.

And then I can say, that's not quite what I meant here. Refine it in these ways. That's what we're moving towards in terms of code generation is just refining that so you can start with an intention and refine the intention until it produces the code that you're looking for. com and they were talking about this in the sense of DALI for prompting for images, etc. But I was really interested in this conversation for content, written content in code. Is this now the next major skill, the ability to prompt these AI interfaces for the correct fit and form?

To a certain degree. So there's this idea of a prompt engineer, which is someone who's just learned the ins and outs of how to talk to the conversational AIs that we're working with today to produce the desired output. And just like any other person who works intimately with the system, you get to know its foibles, its expected syntax, and how to massage it to produce the best results. And so if you look at some of the prompts that people have used to produce images with DALI or MidJourney or any of these image generation platforms, the prompts end up being whole paragraphs where they've just iterated until they found it will produce the thing that they're looking for.

What's interesting for me is the way that I started using Copilot for content creation by accident was I had Copilot enabled. And I also write most of my documentation in Markdown. So when I'm creating a demo in Terraform or something like that, I write a readme so people know how to use the demo if they want to try it out. And Copilot was enabled for Markdown too. So I would start typing a sentence and it would figure, hey, I think I know where the sentence is going.

And it would suggest text to me to complete the sentence or the whole paragraph. You know, how about half the time it was close? So I could just tab complete the sentence and then go back and fix the part that doesn't sound quite right or doesn't sound like my voice, right? I also write scripts for various videos or, you know, whatever it's for. And as I'm writing scripts, I also write those in Markdown. And hey, here's Copilot coming along again, suggesting either sentence completion or whole paragraphs.

And it has actually accelerated the amount of scripting I can do for a given video by helping complete the boring parts, the mundane parts, you know? And so I can get the meat, the gist of what I'm trying to say across. An interesting thing somebody talked about was what they were envisioning is I would like to give ChatGPT or a similar thing five bullet points, well-written bullet points, and have it produce the blog post that someone would summarize into those five bullet points for those people who like to consume blog posts.

And I was like, oh, that's me. I don't want to read the whole post. Just give me the bullet points. Right. I've actually just retweeted something to that effect about an email. com quite a bit. And what I'm finding is that you're absolutely correct. For the mundane stuff, it's really great. It helps me to remember that there's probably five elements of a topic that I want to hit. The problem that I run into is that it then quickly becomes the PowerPoint presentation template.

Like when you first use like, man, this is really cool. Then after a while, you're like, you know what? It's kind of limited in imagination. But with the chatbots is even worse because sometimes it can just be flat out inaccurate. And like the other day, I said, hey, write a resume for Keith Townsend, the CTO advisor. And it started out brilliantly. It summarized my career experience really well. Then when it got into the details, like my career history had me working at VMware for seven years, EDS for 10.

And I'm like, I worked at VMware for like six months. And I worked at and I've never worked at EDS. And it had all these details of all of my accomplishments at EDS. And it was just flat wrong. But that's not the point. The point is that I have a really great structure that's gotten me 45, I think 45% of the way is a fair estimate. Like even if you look at the highlights, the highlights themselves were right. They were just in the wrong company.

So it was a great idea. It was a great idea starter, not a complete thing. Do you see with great prompting the or improved prompting skill? If I were to invest time in leveraging AI today, I'm a mid-level engineer or content creator, is it time better spent refining my ability to code? Or is it time better spent refining my ability to prompt? I think where we're at today, the content that's produced by the AI still needs a lot of human touch for it to be correct, whether we're talking about code or content, reviewing what it's written, you'll find lots of factual inaccuracies and reading the code that it writes, you'll find a lot of inefficiencies in the code.

So I think the two are comparable because chat GPT doesn't really differentiate between the two. It's all language, right? What it's not able to do today is fact check itself. It's not able to do that next level of abstraction, that next level of value add that a human needs to provide today. Now, that being said, we just crossed a threshold in terms of AI being accessible to the average user and being good enough to produce that first level content. The next level is not that far away where it can now fact check itself and use input from humans to retool its model to be better at producing factual and input that is almost imperceptibly the same as what a human would produce.

I'd say we're only a few years away from that. So where would I focus my efforts? I think platform engineering to a certain point or not prompt engineering to a certain point, understanding how these tools work, how they're evolving and how it can make you as a human being more efficient. As the single proprietor of Ned in the Cloud, I am unlikely to fire myself. Like I am not going to make myself redundant. What I can do is make myself more efficient and I can produce things that are better and higher quality by focusing on what differentiates what I make versus the drudgery of just, you know, hocking through this or that.

So a good example, I had to write an SOW recently and it was an SOW for something I hadn't produced before. I asked Jack GPT to write the SOW for me, gave it some basic parameters. It produced something that was 75% of the way there. And then I spent 10 minutes fixing it up to be the correct language and in the right format. But it did so much of the work that would have been awful for me and maybe save me 30 minutes and hours worth of work.

That was absolutely worth it. And I only expect it to get better over time. So I think of AI as the biggest threat right now that AI presents to is to the other freelancers that I work with because it can displace what they do for me because I just need it to take care of the drudgery. And that's what I pay freelancers for is like, you're good at this. I'm not, or I don't want to spend my time doing this. I'm willing to pay you.

Oh no, here comes AI displacing you. Yeah, I think the opportunity, I want to get in one last question before we end, but I think the opportunity even for those freelancers is basically to up level and do more. So offer you additional value on top of what AI can, on you doing the AI yourself. I think there's, again, opportunity for everyone. And I think a great example of this is to talk about, about AI and AI ops and chat GPT and AI ops, uh, industry that I love to pick on is one of, uh, is one that I grew up in, which is the network engineering field.

You know, network engineers are loathe to use automation because they believe automation will displace their work, their, uh, their jobs. Simply put me being able to say, Hey, chat, GTP, right. A, uh, uh, VPN policy between two nodes is extremely low value today. But what I am excited about is the ability for chat enabled AI to help with troubleshooting industry specific problems, et cetera. How should folks who have specialized skill, whether it's infrastructure's cold network engineering and these, uh, general chat entering these fields and being really good at basics like ally your freelancers, how should they be thinking about these tools?

Well, I mean, it would start with the fact that leveraging AI lets you do more with less, you know, you are offloading a portion of your responsibilities onto the AI, which means that you can be more effective as an individual contributor, but that means you have to be more effective at the things that differentiate you from the AI. So if that's troubleshooting skills, the AI is going to present information that may or may not be useful. You as the operations engineer have the contextual awareness of the architecture of the entire system and what normal looks like and what problematic looks like.

And so you can use AI to gather the data, to present it in a useful format to you, maybe even look for some insights, but you're driving the conversation at the moment. And so you need to know how to steer that conversation to get to the end result of here's what's wrong. Here's how you fix it. Hopefully, you know, people are focused more on those troubleshooting skills and the, the logic and a larger context of their environment and not so focused on how to write something at a command line.

Cause that's kind of what we were taking network folks to task for is like, stop focusing on command line arguments and focus on the bigger picture and maybe represent things with code. Well, now we're saying maybe representing everything with code and learning the minutia of code. Isn't the big thing. Again, it's the bigger picture. It's understanding how all these pieces connect together and the larger context of the business in which they connect. So now we plugged your website a couple of weeks ago, this go around.

I want you to plug what's coming up. What, what's what AI assisted piece of content or educational work are you working on now that we can, we can look forward to the next month or two or, or in the near future. Oh, sure. So I am just finishing up the last course in a learning path for Azure virtual desktop. So if you are interested in deploying and managing Azure virtual desktop in Microsoft Azure, I just finished up a six part learning path on Pluralsight that is focused entirely on that.

So if you're thinking about learning it, or you want to sit the exam that's associated with it, maybe check that out. So hopefully you snuck in some automation, some as cold stuff in that, that'd be really, really cool. Cool. Carl's over for those who want to follow Ned on Twitter. He is a net. What's your Twitter handle? Ned, 13, 13, Ned, 13, 13, really easy to follow. I am also still on the bird site. I don't know.

They're doing something special as of this recording with verified accounts. So I don't know if you'll see my tweets or not anymore. Who knows with our friend, Mr. Musk, but I am on Twitter at CTO advisor is the handle DMS are open. You want to learn more about the CTO advisor. com. Make sure to rate the podcast. I went into, I've been neglecting to actually do that. I went into iTunes a couple of weeks ago. I realized the last rating was something that I gave myself.

So let's, let's get some more recommendations in the podcast. Let's talk to you next episode of the CTO advisor.