Brian chats with AI expert and WordPress developer Keanan Koppenhaver about how AI tools are transforming workflows. They explore AI’s strengths, limitations, and practical uses in coding, brainstorming, and streamlining creative projects, offering a framework for gradually adopting AI in tech workflows.
Brian (00:00)
Keanan welcome. How are you doing today?
Keanan (00:01)
I’m doing alright. It’s good to be here. Thanks for having me.
Brian (00:04)
Yeah, for people who don’t know you, maybe you could give like kind of a brief bio background, how you got into WordPress, that sort of stuff.
Keanan (00:10)
Absolutely, yeah. So my name is Keenan. got into WordPress straight out of college. Actually, I started working at a WordPress VIP agency in Chicago where I live and worked there for a few years, just kind of getting interested in WordPress, starting to use it. They use it for most every project. so anything from, you know, small local grocery stores all the way up to some of our VIP clients. After that, just bounced around the WordPress world for a bit, worked for a pretty big tech blog that was all built on WordPress and then started my own.
WordPress agency. And so, you know, had lots of friends in the community and wanted to give it a go for myself. So with my business partner, we started our own shop a few years ago. And yeah, that’s basically kind of brings us almost to today. I’m no longer involved with that agency on a day to day basis, but very thankful to the world of WordPress and everybody, including yourself, who, you know, I’ve gotten to meet through that community.
Brian (00:58)
And I think of you as like also partially like agency buddy who understands the agency life, but now it’s much more like my go-to AI expert, like the person I am like, all right, what would Keenan say about this when it comes to AI? And I’m thinking because of WordCamp Phoenix, think, I don’t know if it this year last year where you gave a presentation and it was the first one that I saw where it was like you very neatly explained like, here’s what generative AI is good at.
Here’s what WordPress is good at. Here’s why you should pay attention to this, kind of. And then was like, maybe one of those first moments where I was like, okay, I kind of see it now. So like, maybe how did you get into AI from WordPress? And maybe a little bit about that talk if you remember it.
Keanan (01:41)
Yeah,
it’s a good question. So, I think my first kind of aha moment with AI, which I think a lot of people had as well was when Chai GPT came out. So that was the end of 2022. and I hadn’t really been following any of the, the GPT models or any of the generative AI work before that, but, I used Chai GPT and my, I think my first prompt was something like, you know, write me a poem in this style. and it just like kind of drove home for me, even though the models weren’t anywhere close to what we have now, that these new tools were really great at a separate
kind of problem solving that I knew how to do with code or any of the other technical tools that I had. like you mentioned, WordPress was my frame. And at the time WordPress, like in that technical work was really all I was thinking about. And so it kind of became obvious for me that like this is a content based language based sort of technology and blogging and web app building. And these language models were going to have a huge impact in all the areas of that. So just started kind of experimenting and thinking about how that would work. And, you know, again,
There are no like instruction manuals for this technology. There’s no very clear like with programming languages you have this language is good at these kinds of things, but not good at these kinds of things with a lot of these generative AI technology. They have weird knowledge gaps and skill gaps that like you would expect them to be able to do something very well and they can’t. And so really just like dove in and tried to start experimenting with can this write me an outline for a blog post or can this take something I’ve already written and help me edit it or can this write me a WordPress plugin or.
Brian (02:44)
Mm-hmm.
Keanan (03:02)
You know analyze code errors that I’m getting so that I don’t have to constantly like go back and forth debugging if it’s something simple. And so yeah, my my talk in Phoenix. It was February of this year was really about just like helping people. Have a little bit of a better idea. Just generally what these language models were good at and what they are good at and the frustrating thing about that talk is it’s very quickly out of date, so I like if I had to go back and do that right now it would no longer be accurate. There were things that I mentioned that the models couldn’t do.
Brian (03:23)
Yeah, yeah.
Keanan (03:28)
that they now can, especially with things like O1 coming out and stuff like that. But generally, you know, I try to give people that mindset of, know, if you want perfect factual recall, if you’re really looking for a search engine, like that’s not what you should use these tools for. But they do unlock a whole separate class of solutions that you there were no good ways to solve before.
Brian (03:32)
Mm-hmm.
Yeah, and it’s, I’ve been using it a bit for like, you know, a bit of a search engine where I’m like, is this something where I know there’s tons of good information already out there on the internet? Then it’s like, okay, it can kind of synthesize that or something like that. Like there are like these weird little places and stuff, but you had this great thing where you were like, you were like basically like images, content and code, I think were the three buckets you had. And it was like, like that’s your, that’s literally what WordPress is. It’s just like a big repository of those three things.
And I think for most people, probably coding, I would bet, is the one in the WordPress space maybe where they’ve maybe started or maybe been the most impressed. don’t know, would you agree with that one?
Keanan (04:26)
Yeah, I think the coding stuff for me is the area where it’s gotten the head the most improvement since I initially discovered it like early versions of Chad GPT were not very good at coding at all. But now, especially with the 01, the new Claude model 3.5 sonnet and a lot of the even coding specific sort of agent tools that are meant for that. You can really get pretty good output in a much shorter time. Just the other day was working on.
project and I needed just a little like quick CLI script to do some maintenance that’s going to run on a schedule and it’s something that you know I knew how to write and I could have written myself. It probably would have taken me an hour or two to put together, but I basically couple paragraphs of spec to Claude and it did the first draft for me. I noticed a couple areas where it was missing functionality that I needed and forgot to tell it about. So I mentioned that it updated the script and in about a half an hour we had something working that was able to just basically be tried and shipped.
Brian (04:54)
Mm-hmm.
Keanan (05:14)
And that’s just a huge amount of time saving. So obviously, you know, that was a case where the model really helped me because I knew what the correct sort of output looked like. And that that’s where it makes it a little easier to verify that what you’re getting is is accurate. But yeah, I think for people using it for coding, even for just small stuff, generating, you know, a call to a rest API, if you’ve done that 100 times in WordPress before, you know what the code looks like and you can very quickly verify the output and move on. That’s the kind of stuff we’re
Brian (05:37)
Mm-hmm.
Keanan (05:39)
I think people are seeing kind of huge benefits with this technology.
Brian (05:42)
Yeah, that’s, that’s, would say where I’ve mostly landed. It’s a lot of things, you know, how to do, and you just don’t want to do it again, especially if you’re doing like the co-pilot, like more like auto complete style where you start typing and it kind of finishes it for you. the other one too has been a lot of, I’m getting this error code. Tell me why this error code keeps happening. And, that one’s helped a lot. but you, yeah, it’s like, you really do still have to like understand coding. I don’t.
I don’t know that I’ve seen an example of somebody who has no coding experience or like technical knowledge open up a code editor and like come out with something great. I don’t know.
Keanan (06:16)
Yeah, it’s
very easy to, think, get fooled by the initial magic of this technology into thinking that like it really can do anything. but I’ve, you know, I tried, earlier this year, I wanted to, you know, make a really simple video game. And so I’m not a unity developer. I’m not a C-sharp programmer. And so I basically said, Hey, Claude, here’s what I want to develop. Like, can we work together? And even though I understand code, I don’t understand that particular language. And I noticed there were a couple of different times where I just caught, got caught in these like infinite loops where
Brian (06:31)
Hmm.
Keanan (06:43)
It was fixing one bug, but I didn’t realize it was actually introducing another one. And so it was just bouncing me back and forth. And same thing. I’ve had it, you know, help me debug like Docker configurations and other sorts of like infrastructure configurations that I’m not as familiar with. And I realized like, if I zoom out, we’re basically going back and forth and trying the same three steps over and over and we’re not getting anywhere. And it feels like progress because it’s constantly giving you new things to try. But if you kind of step back and think about it just a little bit critically and avoid getting sucked into the machine a little bit.
Brian (06:47)
Haha.
Yeah.
Keanan (07:11)
you’ll realize that it’s not actually giving you what you want. And that’s hard to notice if you’re not proficient in the technology you’re asking for.
Brian (07:19)
Yeah, I’ve definitely gotten caught in those loops before where it keeps, you keep saying, I have this bug. And then they’re like, okay, cool. We fixed it by doing this. you, after like three rounds, you’re like, no, we’re just fixing the same two bugs back and forth, but like we’re not doing it in a way that’s like meaningful. And it’s like, I feel like there’s been this theory in tech generally. And I know like in WordPress, like Matt Mullenwick says this a lot where like, you know, we’re going to start making software so much faster. Like WordPress is going to.
Keanan (07:29)
Yeah.
Brian (07:46)
be able to just like start progressing so much faster because AI is gonna be writing all the code and stuff like that. And it just feels like, don’t think, like I don’t, I think there’s definitely amazing productivity gains and I’m so happy I don’t have to write REST API endpoint stuff anymore and that sort of thing. But I find it really hard to believe that like WordPress is gonna start writing itself and we’re gonna suddenly see faster iteration on software. I don’t know, what do you think?
Keanan (08:10)
Yeah, I tend
to agree. I think that the speed is oftentimes overrated. Like it’s true that these models and with each successive version that comes out, this gets more true. They can give you output very fast, right? Especially the smaller models that are trained to do that sort of thing. They can give you a lot of output very fast. But at some point, like the speed needs to trade off for quality and correctness. This has been the same issue that people have had, you know, working with.
human programmers or any sort of automation system is like, you can build these things that give you results really fast. Someone still has to be checking them, verifying that they’re the quality that you want, that they’re correct in the way that you want them to be. And especially as we start giving these models harder and harder problems to solve, like we mentioned before, like if they introduce, you know, very subtle bugs that are hard to find, they may have done it quickly, but it’s, you you still have to spend that time kind of checking and QAing all that stuff. So I think that
there are efficiency gains for sure. But for me, like you mentioned, it’s like the things that will save me 45 minutes to an hour where I’ve already done that task a hundred times in my life and I could do it again. But if I can use these tools, get the benefit of speed, but also be able to very quickly verify the output is correct and quality. That’s, think where we’re going to find the sweet spot.
Brian (09:20)
Yeah. And the other piece too is like, like WordPress to use WordPress as an example. Like I actually don’t think WordPress’s biggest problem is that it needs more code. It’s probably that it needs more like product user experience, like expertise, you know, like that’s where it’s struggling. It’s like not that people aren’t writing enough code fast enough. It’s like the actual user experience. And for a lot of these situations where it’s like, I’m to go to a piece of code and I’m going to, or I’m going to go to AI and I’m going to say, generate a Python script to like do this data management for me. Perfect. Does it well?
exactly what you need. When you say, build like an app that’s going to do this. It’s like, well, now that’s a human user experience problem. And that’s, that is really writing the code is like the smallest part of that. When you have like the marketing, the design, the, know, the user interface, all that other stuff. I wonder if it’ll get to a place where it can solve those problems. or if, like you said, it’s just going to be like, you still need a human and you just like, it’ll help you a little bit in each of those places, shave off a few hours or something.
Keanan (10:18)
Yeah, I absolutely agree. think if you are arguing that WordPress needs more code, faster produce, like you can just go look at the pull requests tab on the Gutenberg repo. And it’s like, there’s so many PRs there of code that’s already been written, right? Like that’s, that’s not the problem at this point, but I’ve used a few of these automated coding tools. Replets agent is one that comes to mind, which is actually very good. And V zero from Vercell as well, which is like a, you know, it helps you design react components and builds them for you. And I do think that.
Both of those tools, unless you give it a lot of input with a lot of opinions about here’s kind of the color scheme that we should use. Here’s the style that we want to go for all of those things. They do still produce like very generic looking output, which in a lot of cases, like you mentioned with your Python data management script, generic reliable output is exactly what you need. But for things like building a product or building a website or anything that does require a little bit of like taste and kind of choices to be made about how it works and what it looks like.
These tools definitely aren’t there yet in terms of being able to provide that so far.
Brian (11:16)
And then like on the content side of it, there’s, I’m sure you’ve seen, it’s like a meme or a little cartoon or something. And it’s like the one guy is like, you know, like basically it’s like, hey, take this one sentence email and make it a paragraph and send it to that person. And the other person is AI, take the paragraph and give me back the sentence. And I’ve been doing a lot of like AI training materials, stuff like that. And a lot of what it’s really useful for, like data, spreadsheets, generate reports, read through all these spreadsheets. And it’s not nearly,
Keanan (11:29)
Yeah.
Brian (11:46)
as it’s a lot more boring office worker knowledge work stuff than I think the glamour switch is like, Sora is making videos and you know, like there’s cool stuff happening with like art and visual, but like, it really seems like the actual place where it’s like doing the best stuff is boring, mundane office work, even on the content side. Like, it’s, know, like your example is, it could write a poem or something. And it is really kind of crazy when you watch it do that. Like,
Sometimes your mind just goes like, how did you make that happen? But it’s like, I’m not going to probably read an AI generated novel at any point in my life. Like, I don’t think that’s ever going to happen. so, so what are your thoughts on like the content side?
Keanan (12:25)
Yeah, I agree. And I think this is a point that Dan Schipper from, from every makes really well is like even with content production, which is like the most, one of the most creative ways to do like that kind of work that you’re talking about. there is a lot of like mundane sort of drudgery type tasks that go along with it. Right. So like if I’m writing a blog post, I can have a really creative idea for something that I want to say. And that’s great. I still need to kind of think through like the structure of how it’s going to work. And that’s a, that’s a place where I use AI a lot is here’s my one paragraph pitch for this blog post. want to write.
Help me think through this outline and I’ll notice. OK, here are a couple points you missed or here’s something that you included that I don’t actually think we need to put in this post and so we’ll get to a point where it’s a really good outline and then I’ll do the writing. What most people think of is like the content production piece of it. But even after that and like I’m sure you’re kind of realizing this again as you’re starting up a new podcast, right? It’s like there are social media posts to write. There are like things to clip down and you want to pull out, you know, three interesting pieces from this in a form that’s.
you know, short enough for Twitter or X or and then one that’s a little bit longer for LinkedIn and like there’s all of this stuff that happens before and after the actual like creativity piece to make the thing you’re creating It’s most impactful, right? Like there’s all of that extra work as well. And so in my mind, like if AI can help with that, that helps more people kind of get their core ideas out there and more distributed. And you can make the argument that, you know, if we take the human effort away from that work, that means that.
social networks and everything else will just get flooded with all of this AI generated content. But I do think that at the end of the day, like there is all this work that surrounds the core work. Right. And that’s where I use AI a lot is to automate that sort of stuff.
Brian (13:57)
Yeah, I’ve had the same experience where it’s like, you know, I need to do a tutorial on this, like, give me like 10 different examples of what somebody would use this tool for. Give me like 15 block ideas that I could make a tutorial for. And it does really good brainstorming. It does really good outlining. It does a lot of that stuff pretty good in a way that like, you know, it’s probably now becoming dispensable to my workflow where I’m not sure if I…
I’m gaining or losing by not doing some of that brainstorming work myself, but I know that I don’t miss it, you know? but I think it’s the same problem. Like you said, with code, like the problem isn’t that like we’ll say WordPress needs more code. Like there’s plenty of code that is out there. And the other problem is like, there’s not a problem that we need more content. Like no one’s sitting there being like, there’s not enough content in the world. There’s not enough like podcasts and stuff like that. But like said, like taking out some of that, side work.
Keanan (14:28)
Yeah. Yeah.
Brian (14:50)
some of that stuff that like honestly like no you know how important is the tweet about the podcast episode like do you want to sit there and write you know all the tweets and stuff like that yeah i can kind of see that i can see why that’s the place to to sort of leverage it and where it’s becoming more useful for me i think
Keanan (15:08)
Yeah. And I think, you know, you can say like, how important is it for you to write that tweet or whatever, but I look at it both ways of like, it might not necessarily feel very important at the moment, but like the amount of podcast episodes that and brand new podcasts in general that I’ve discovered because someone posted about it to social media, either the guest did or the host did. And like, if that, you know, hadn’t existed, I would never have found out about it. So it is really important that like, if you have an idea that you want to get out in the world, you do have to tell people about it and tell people about it frequently and consistently. And,
People don’t always want to do that work or they feel bad kind of doing that self promotion in that themselves. And so these tools can be really good for that, especially if you have a strong opinion about like sample inputs and what sort of format you want it to be and that sort of thing. I think one of the reasons you’re successfully using it for ideas for blocks and things is because you have the WordPress experience and the perspective of what makes it interesting block and whether you’re kind of doing it explicitly or not, you’re giving the model information about what you want to see in a tutorial and.
Brian (15:49)
Hmm.
Keanan (16:04)
how you think about that. If I used it to brainstorm, you know, blog post titles or whatever, without giving it any sort of background knowledge or not having any knowledge myself in that particular area that I want to write a blog post about, I’ve noticed that the output is really pretty generic. But if I can give it, you know, good examples that I found online or just more information that comes from being me being involved in the topic, that’s where I found it’d be really, really helpful.
Brian (16:26)
Do you ever get concerned? Cause I know you have some experience doing like obviously teaching and like technical writing and tutorials, that sort of stuff. Like, I find myself going to chat GPT more often when I’m just needing some sort of educational tutorial thing that I know has been written a million times. And I know it’ll be able to like, give me a reliable answer versus like, you know, developers who, to write these sort of like technical blogs and things like that, where it’s, it’s pulling their
content and repurposing it, but it’s giving me maybe a more specific answer that’s like more useful to me. Do you get concerned that you’ll like lose the desire to write the post or like create the content or that will run out of that side of it or or do not worry about that?
Keanan (17:09)
Yeah,
think it’s a good question and I actually thought about this a while back and ended up writing a post on it of like should anybody blog anymore? Basically was something that’s the title was something around that and like I think it depends on why you’re doing it right. If you are just purely trying to. You know, right? Because you want to be a freelance writer, that’s I think going to be a lot tougher because like you said, one of the things that AI is great at and that I use it for all the time is to get a really specific tutorial. So it’s like I want to do this thing in WordPress, but I need to do it on this version of PHP.
Brian (17:18)
haha
Keanan (17:37)
and I need to use these three plugins, like maybe that’s a tutorial that doesn’t exist, but because of the collective knowledge of the model, it can generate that for me on the fly, right? So that’s a really good use case that is going to eat up a lot of the sort of SEO optimized, like generic writing. I think for folks who are more in like a consultative sort of role or who really want to show like, am someone who’s knowledgeable on this topic.
Brian (17:40)
Mm-hmm.
Keanan (18:02)
that’s still going to be important to write those sort of articles just to like show your expertise. And because you will have, you know, an opinion and a perspective on that topic that a generic AI model won’t have. I think the other piece of that that is only starting to now become important is that GENGBT and other models are starting to cite sources, which means you can actually get, you know, referral traffic that you might get from SEO in years past, like from these AI models now. And so the sooner you can like get your content in those models and the more comprehensive the
that you can have in terms of like coverage. I think you’ll still be able to benefit from, you know, even your work being synthesized by these models. Cause a lot of times people want to dig deeper and see sources and that’s becoming a more common pattern is people try to combat, you know, hallucinations and things like that.
Brian (18:44)
And I think like back to the beginning of the internet when people were just like, Hey, this is so cool. I can just put stuff on it. And I’m going to like put my recipe, you know, on my blog for fun and put all my recipes here because I want to, and everybody can do it to like now where when you try to read a recipe and it’s this like massive wall of text and ads and SEO stuff. And it’s just like the worst experience ever because that person’s trying to like, you know,
pay their mortgage based on like recipe blog posts or something like that. that long journey has been pretty brutal. But like you said, like I actually am optimistic that it’ll be like just another version of Google SEO. It’ll just be instead of SEO, it’ll be, don’t know, AI EO, no AIO or something like whatever, whatever it would be. Is there.
Keanan (19:09)
Right. Right.
Yeah, I’ve seen I’ve seen various acronyms floating around on LinkedIn. It seems like
people haven’t decided on one yet, but yeah, especially with like chat GPT starting to with search perplexity has already been doing that for a long time. Like there are tools that are really giving you the best of synthesizing and creating something custom, but also then citing sources and saying, hey, if you want to go further, here’s where you go. And so I think it’s, going to be really beneficial for folks to be there for sure.
Brian (19:49)
Do you have like, you know, since you have like the agency experience, do you have like tips or kind of like if you’re, you know, say WordPress agency, you’re a developer or on that sort of side of it and you feel like, everybody’s using AI and I’m not, or something like that. Do you have like kind of a good, you know, framework or like tips for people to kind of get their feet wet with that?
Keanan (20:10)
Yeah,
absolutely. I think about it in the kind of this is a common framework, but I’ve adapted it a little bit. It’s like the crawl, walk, run, you know, fly sort of framework, right? So if you’re not using any of these tools at all, there are a ton of like off the shelf tools that have AI built in that are already really useful without you needing to do anything yourself or customizing anything yourself. People started seeing that, you know, the proliferation of like AI meeting assistants, right? To sit in on your Zoom meetings and take notes. And that’s a really good example of a tool that you can just kind of download and use.
and get a ton of value from. I use granola personally. And basically just like listens to your meeting. You can take whatever notes you want to. And then when you’re done, it flushes out your notes with additional notes from the transcript. Let’s you link to the transcript and also ask any questions you want about the meeting. And for folks who are in a lot of like client meetings, especially agency folks like that, super helpful. You don’t have to think like, that they mentioned that there was something really important for this project that we’re kicking off. And like, I don’t know what it is. You can just kind of get the get the whole transcript.
There are also tools that help you, you know, take transcripts like that and turn them into text specs or technical requirements, documents, a lot of, again, the the work around the core agency work of producing whatever product you’re going to produce is you have to, you know, have meetings with clients, take good notes, turn them into tech specs for your developers or even for yourself. And there are a lot of just off the shelf tools that you do that. The kind of walk phase I think about as using tools like chat, GPT and Claude and these public chat bots where
Brian (21:12)
Mm-hmm.
Keanan (21:30)
Yes, there’s only so much kind of customization you can do, but you can upload your documents there. You know, you can ask it whatever questions you want to use its whole training data set to help give you good answers and automate some repetitive stuff. The run step is like taking those tools a little bit further with some advanced prompting techniques and you know, using things like Claude projects so that you can have, you know, a large language model that kind of has all your client specific context for one client or, you know, for a specific task you’re trying to take on.
Things like giving JGPT custom instructions so it knows how to respond to you specifically. Those are a little bit of like things where you can take that a little bit further. And then if you want to go further than that, right, this is the fly step. There are a ton of tools and I know you’ve been experimenting with a lot of them that are let you build your own kind of AI apps and AI enabled tooling. Things like Zapier and make and retool some of those. If you’re familiar with like code and able to code yourself like all of these large language models also have APIs and so you can.
You know, make direct calls to large language models from within your code as well. And that step gets a little more complicated because then you’re mixing the kind of fuzzy output of a large language model with the deterministic output of your code. And there are a lot of difficulties there, but you definitely don’t have to like jump into that at the very beginning. You can kind of start using these off the shelf tools and step your way up. And like we mentioned, because you don’t really know what these tools are good at until you try. That’s how I really recommend folks get into it and kind of just like up their exposure gradually so they can start to see.
Brian (22:32)
Mm-hmm.
Keanan (22:50)
This was a really good use case for this tool. We’re going to keep doing that or we spent two or three hours trying to get the AI to do this and we couldn’t really get it to work out, so we’re probably just going to shelve that for now. Maybe come back and try again when the models get better, but at least for now I can’t do that. That’s kind of how I think about it.
Brian (23:03)
Yeah, and I’ve been, I’ve been trying to ramp up my usage. so like one common thing has been like anything that’s like, we said like busy work, knowledge work, that sort of stuff. It’s like, even if I’m not sure what to do, I’ll just ask a model like, Hey, this is a workflow I have right now. Like where do, where can I automate it? Like automation really is a big part of it. Like it’s, it’s almost like the two really work together because it seems like it’s not just that you
want to use it to do this work, but you also want to automate processes and stuff like that. Half of the stuff I think I’m gonna use AI for, I’m actually just using Zapier to automate the thing, but it’s all part of just a different mindset of thinking about everything that you do and what you actually want to spend your time on and what is the busy work of your life that you don’t want to spend your time on.
Keanan (23:50)
Yep. Yep.
Brian (23:54)
The part you said earlier, which is about it being more ingrained in a lot of the tools, like we’re recording this on Riverside and when I’m done, it’s gonna do the show notes, it’s gonna do clips and stuff and I’ll edit everything, but it’ll give that first pass to me. Do you have a theory on, I was originally a strong proponent of the idea that like,
the operating system is going to handle AI for us in a way where you’re not even gonna need agents. you’re gonna be able to just, like, it’s gonna be so deeply baked into your computer that you won’t feel like you have to open up an app to use AI. It’ll just be part of the way you experience using a computer. And then I tried the Apple intelligence, like beta versions, and it’s by far the worst AI implementation I think I’ve ever seen. It is so…
terrible it actually makes you less productive with their stupid summaries of your texts and things like that. So now I’m not so sure about that. Do you have a like a theory on this if it’s going to be an OS thing or or what?
Keanan (24:50)
Yeah.
Yeah, I appreciate you mentioning Apple intelligence. just upgraded Mac OS on my work machine this morning and I got the prompt for it. So I clicked skip for now. I might have to set that up later, but after that not exactly glowing review, I may just skip it for now. But I think like there definitely is a large community of folks that want to run models on on their devices. And I think that unlocks a whole separate set of things that it can do. The trade off you’re really making is like, at least for now, the it’s really only the smaller models that can be run.
actual devices. And that’s why when you run chat GPT or when you run Claude, like it’s running on their infrastructure and you know, they’re there like the reason those more complicated models are able to run as quick as they are is because they have the specialized infrastructure for that. So I think that is one kind of current roadblock that will definitely be solved in the future. And bigger models will be able to run locally. But the other thing is that it’s still the case that people want, I think.
purpose-built tooling for a lot of these things. And so it’s like, if you have, you know, a sort of a granola like tool, right? That can listen to your meetings and take notes, they’re going to continue introducing features that are really good for that specific thing. Whereas an Apple intelligence, for example, doesn’t necessarily have the same incentive to introduce those features, right? And so, especially when you get to tools that are even more narrow than like take notes on our meetings, which is a pretty general purpose tool. I do think there will still be
Brian (26:06)
Hmm.
Keanan (26:14)
need for a lot of these tools that do one thing or a suite of related things very well. And the same reason that Chad GPT’s voice mode and things like that have had more initial traction than Siri, for example, or initial success maybe is because there are teams that are focused on producing that and not on producing an entire operating system, but also including some AI in there.
Brian (26:37)
Mm-hmm.
Keanan (26:37)
I think
it really comes down to like focus and the level of like detail you’re looking for from a solution like that.
Brian (26:44)
So the idea that one, I guess, universal AI that’s going to be able to specialize itself into all of these different areas, you think that’s probably far off and we still need human teams specializing these models to solve all these different use cases.
Keanan (27:00)
Yeah, I do. I look at it in the same way as like search engines a little bit like even though Google is the dominant search engine by far, we still have Bing. We still have doctor go. We still have all these other search engines, right? And we even have domain specific search engines, right? So there are legal specific search engines. There are scientific publication specific search engines and things like that. And so even though like one company is so dominant in that area, like there are still all these other tools and that calculus may change when we talk about AI models because.
the thought of like, well, if we actually achieve the artificial general intelligence that people say is coming, then maybe we won’t need any other platforms anymore because that tool will just be able to invent anything that it wants to at any moment. I think there’s an argument to be made there. But yeah, at least for the foreseeable future, think people like solutions that feel right to them or feel like they serve their needs pretty fully.
At least for now, like the only way we get that is by having teams that are focused on those problems and building solutions for them.
Brian (27:55)
Yeah, I think that makes a lot of sense. And, and like you said, like really understand, you can understand why the incentives at Apple are probably not. I’m just surprised. You know, I always thought Apple was not going to be one of those like, let’s stuff a feature in cause it’s popular, but it really does. Did come across as that where you’re like, if this is people’s first experience of these, of this like AI in general, I that’s not helpful at all. yeah, yeah. I.
Keanan (28:07)
Yep. Yep.
Yeah, yeah.
Brian (28:19)
There’s other things I wanted to talk to you about. Like I’m really curious about the concept of open source now, but I think we’ll have to save that for another conversation. Cause I think there’s, there’s a lot of other sort of issues around, you know, the way AI works. But for now, maybe you could just give me like where people can go to learn about, learn about you, follow you, learn about AI from you.
Keanan (28:41)
Yeah, it’s a good question. I think the kind of two best places would be LinkedIn. That’s where I’m sharing a lot of my stuff these days. I’m also on blue sky as of recently. So at K Copenhagen dot B sky dot social I think is what the kind of default handle is there. But yeah, those would be the two best spots. I post stuff there all the time and feel free to DM me as well anywhere anywhere you can find me.
Brian (28:55)
Hmm.
Awesome, man, thanks for this. You’re my favorite person to talk to about this stuff, so I appreciate it.
Keanan (29:07)
Yeah,
absolutely. Thanks for having me.
Keanan welcome. How are you doing today?
Keanan (00:01)
I’m doing alright. It’s good to be here. Thanks for having me.
Brian (00:04)
Yeah, for people who don’t know you, maybe you could give like kind of a brief bio background, how you got into WordPress, that sort of stuff.
Keanan (00:10)
Absolutely, yeah. So my name is Keenan. got into WordPress straight out of college. Actually, I started working at a WordPress VIP agency in Chicago where I live and worked there for a few years, just kind of getting interested in WordPress, starting to use it. They use it for most every project. so anything from, you know, small local grocery stores all the way up to some of our VIP clients. After that, just bounced around the WordPress world for a bit, worked for a pretty big tech blog that was all built on WordPress and then started my own.
WordPress agency. And so, you know, had lots of friends in the community and wanted to give it a go for myself. So with my business partner, we started our own shop a few years ago. And yeah, that’s basically kind of brings us almost to today. I’m no longer involved with that agency on a day to day basis, but very thankful to the world of WordPress and everybody, including yourself, who, you know, I’ve gotten to meet through that community.
Brian (00:58)
And I think of you as like also partially like agency buddy who understands the agency life, but now it’s much more like my go-to AI expert, like the person I am like, all right, what would Keenan say about this when it comes to AI? And I’m thinking because of WordCamp Phoenix, think, I don’t know if it this year last year where you gave a presentation and it was the first one that I saw where it was like you very neatly explained like, here’s what generative AI is good at.
Here’s what WordPress is good at. Here’s why you should pay attention to this, kind of. And then was like, maybe one of those first moments where I was like, okay, I kind of see it now. So like, maybe how did you get into AI from WordPress? And maybe a little bit about that talk if you remember it.
Keanan (01:41)
Yeah,
it’s a good question. So, I think my first kind of aha moment with AI, which I think a lot of people had as well was when Chai GPT came out. So that was the end of 2022. and I hadn’t really been following any of the, the GPT models or any of the generative AI work before that, but, I used Chai GPT and my, I think my first prompt was something like, you know, write me a poem in this style. and it just like kind of drove home for me, even though the models weren’t anywhere close to what we have now, that these new tools were really great at a separate
kind of problem solving that I knew how to do with code or any of the other technical tools that I had. like you mentioned, WordPress was my frame. And at the time WordPress, like in that technical work was really all I was thinking about. And so it kind of became obvious for me that like this is a content based language based sort of technology and blogging and web app building. And these language models were going to have a huge impact in all the areas of that. So just started kind of experimenting and thinking about how that would work. And, you know, again,
There are no like instruction manuals for this technology. There’s no very clear like with programming languages you have this language is good at these kinds of things, but not good at these kinds of things with a lot of these generative AI technology. They have weird knowledge gaps and skill gaps that like you would expect them to be able to do something very well and they can’t. And so really just like dove in and tried to start experimenting with can this write me an outline for a blog post or can this take something I’ve already written and help me edit it or can this write me a WordPress plugin or.
Brian (02:44)
Mm-hmm.
Keanan (03:02)
You know analyze code errors that I’m getting so that I don’t have to constantly like go back and forth debugging if it’s something simple. And so yeah, my my talk in Phoenix. It was February of this year was really about just like helping people. Have a little bit of a better idea. Just generally what these language models were good at and what they are good at and the frustrating thing about that talk is it’s very quickly out of date, so I like if I had to go back and do that right now it would no longer be accurate. There were things that I mentioned that the models couldn’t do.
Brian (03:23)
Yeah, yeah.
Keanan (03:28)
that they now can, especially with things like O1 coming out and stuff like that. But generally, you know, I try to give people that mindset of, know, if you want perfect factual recall, if you’re really looking for a search engine, like that’s not what you should use these tools for. But they do unlock a whole separate class of solutions that you there were no good ways to solve before.
Brian (03:32)
Mm-hmm.
Yeah, and it’s, I’ve been using it a bit for like, you know, a bit of a search engine where I’m like, is this something where I know there’s tons of good information already out there on the internet? Then it’s like, okay, it can kind of synthesize that or something like that. Like there are like these weird little places and stuff, but you had this great thing where you were like, you were like basically like images, content and code, I think were the three buckets you had. And it was like, like that’s your, that’s literally what WordPress is. It’s just like a big repository of those three things.
And I think for most people, probably coding, I would bet, is the one in the WordPress space maybe where they’ve maybe started or maybe been the most impressed. don’t know, would you agree with that one?
Keanan (04:26)
Yeah, I think the coding stuff for me is the area where it’s gotten the head the most improvement since I initially discovered it like early versions of Chad GPT were not very good at coding at all. But now, especially with the 01, the new Claude model 3.5 sonnet and a lot of the even coding specific sort of agent tools that are meant for that. You can really get pretty good output in a much shorter time. Just the other day was working on.
project and I needed just a little like quick CLI script to do some maintenance that’s going to run on a schedule and it’s something that you know I knew how to write and I could have written myself. It probably would have taken me an hour or two to put together, but I basically couple paragraphs of spec to Claude and it did the first draft for me. I noticed a couple areas where it was missing functionality that I needed and forgot to tell it about. So I mentioned that it updated the script and in about a half an hour we had something working that was able to just basically be tried and shipped.
Brian (04:54)
Mm-hmm.
Keanan (05:14)
And that’s just a huge amount of time saving. So obviously, you know, that was a case where the model really helped me because I knew what the correct sort of output looked like. And that that’s where it makes it a little easier to verify that what you’re getting is is accurate. But yeah, I think for people using it for coding, even for just small stuff, generating, you know, a call to a rest API, if you’ve done that 100 times in WordPress before, you know what the code looks like and you can very quickly verify the output and move on. That’s the kind of stuff we’re
Brian (05:37)
Mm-hmm.
Keanan (05:39)
I think people are seeing kind of huge benefits with this technology.
Brian (05:42)
Yeah, that’s, that’s, would say where I’ve mostly landed. It’s a lot of things, you know, how to do, and you just don’t want to do it again, especially if you’re doing like the co-pilot, like more like auto complete style where you start typing and it kind of finishes it for you. the other one too has been a lot of, I’m getting this error code. Tell me why this error code keeps happening. And, that one’s helped a lot. but you, yeah, it’s like, you really do still have to like understand coding. I don’t.
I don’t know that I’ve seen an example of somebody who has no coding experience or like technical knowledge open up a code editor and like come out with something great. I don’t know.
Keanan (06:16)
Yeah, it’s
very easy to, think, get fooled by the initial magic of this technology into thinking that like it really can do anything. but I’ve, you know, I tried, earlier this year, I wanted to, you know, make a really simple video game. And so I’m not a unity developer. I’m not a C-sharp programmer. And so I basically said, Hey, Claude, here’s what I want to develop. Like, can we work together? And even though I understand code, I don’t understand that particular language. And I noticed there were a couple of different times where I just caught, got caught in these like infinite loops where
Brian (06:31)
Hmm.
Keanan (06:43)
It was fixing one bug, but I didn’t realize it was actually introducing another one. And so it was just bouncing me back and forth. And same thing. I’ve had it, you know, help me debug like Docker configurations and other sorts of like infrastructure configurations that I’m not as familiar with. And I realized like, if I zoom out, we’re basically going back and forth and trying the same three steps over and over and we’re not getting anywhere. And it feels like progress because it’s constantly giving you new things to try. But if you kind of step back and think about it just a little bit critically and avoid getting sucked into the machine a little bit.
Brian (06:47)
Haha.
Yeah.
Keanan (07:11)
you’ll realize that it’s not actually giving you what you want. And that’s hard to notice if you’re not proficient in the technology you’re asking for.
Brian (07:19)
Yeah, I’ve definitely gotten caught in those loops before where it keeps, you keep saying, I have this bug. And then they’re like, okay, cool. We fixed it by doing this. you, after like three rounds, you’re like, no, we’re just fixing the same two bugs back and forth, but like we’re not doing it in a way that’s like meaningful. And it’s like, I feel like there’s been this theory in tech generally. And I know like in WordPress, like Matt Mullenwick says this a lot where like, you know, we’re going to start making software so much faster. Like WordPress is going to.
Keanan (07:29)
Yeah.
Brian (07:46)
be able to just like start progressing so much faster because AI is gonna be writing all the code and stuff like that. And it just feels like, don’t think, like I don’t, I think there’s definitely amazing productivity gains and I’m so happy I don’t have to write REST API endpoint stuff anymore and that sort of thing. But I find it really hard to believe that like WordPress is gonna start writing itself and we’re gonna suddenly see faster iteration on software. I don’t know, what do you think?
Keanan (08:10)
Yeah, I tend
to agree. I think that the speed is oftentimes overrated. Like it’s true that these models and with each successive version that comes out, this gets more true. They can give you output very fast, right? Especially the smaller models that are trained to do that sort of thing. They can give you a lot of output very fast. But at some point, like the speed needs to trade off for quality and correctness. This has been the same issue that people have had, you know, working with.
human programmers or any sort of automation system is like, you can build these things that give you results really fast. Someone still has to be checking them, verifying that they’re the quality that you want, that they’re correct in the way that you want them to be. And especially as we start giving these models harder and harder problems to solve, like we mentioned before, like if they introduce, you know, very subtle bugs that are hard to find, they may have done it quickly, but it’s, you you still have to spend that time kind of checking and QAing all that stuff. So I think that
there are efficiency gains for sure. But for me, like you mentioned, it’s like the things that will save me 45 minutes to an hour where I’ve already done that task a hundred times in my life and I could do it again. But if I can use these tools, get the benefit of speed, but also be able to very quickly verify the output is correct and quality. That’s, think where we’re going to find the sweet spot.
Brian (09:20)
Yeah. And the other piece too is like, like WordPress to use WordPress as an example. Like I actually don’t think WordPress’s biggest problem is that it needs more code. It’s probably that it needs more like product user experience, like expertise, you know, like that’s where it’s struggling. It’s like not that people aren’t writing enough code fast enough. It’s like the actual user experience. And for a lot of these situations where it’s like, I’m to go to a piece of code and I’m going to, or I’m going to go to AI and I’m going to say, generate a Python script to like do this data management for me. Perfect. Does it well?
exactly what you need. When you say, build like an app that’s going to do this. It’s like, well, now that’s a human user experience problem. And that’s, that is really writing the code is like the smallest part of that. When you have like the marketing, the design, the, know, the user interface, all that other stuff. I wonder if it’ll get to a place where it can solve those problems. or if, like you said, it’s just going to be like, you still need a human and you just like, it’ll help you a little bit in each of those places, shave off a few hours or something.
Keanan (10:18)
Yeah, I absolutely agree. think if you are arguing that WordPress needs more code, faster produce, like you can just go look at the pull requests tab on the Gutenberg repo. And it’s like, there’s so many PRs there of code that’s already been written, right? Like that’s, that’s not the problem at this point, but I’ve used a few of these automated coding tools. Replets agent is one that comes to mind, which is actually very good. And V zero from Vercell as well, which is like a, you know, it helps you design react components and builds them for you. And I do think that.
Both of those tools, unless you give it a lot of input with a lot of opinions about here’s kind of the color scheme that we should use. Here’s the style that we want to go for all of those things. They do still produce like very generic looking output, which in a lot of cases, like you mentioned with your Python data management script, generic reliable output is exactly what you need. But for things like building a product or building a website or anything that does require a little bit of like taste and kind of choices to be made about how it works and what it looks like.
These tools definitely aren’t there yet in terms of being able to provide that so far.
Brian (11:16)
And then like on the content side of it, there’s, I’m sure you’ve seen, it’s like a meme or a little cartoon or something. And it’s like the one guy is like, you know, like basically it’s like, hey, take this one sentence email and make it a paragraph and send it to that person. And the other person is AI, take the paragraph and give me back the sentence. And I’ve been doing a lot of like AI training materials, stuff like that. And a lot of what it’s really useful for, like data, spreadsheets, generate reports, read through all these spreadsheets. And it’s not nearly,
Keanan (11:29)
Yeah.
Brian (11:46)
as it’s a lot more boring office worker knowledge work stuff than I think the glamour switch is like, Sora is making videos and you know, like there’s cool stuff happening with like art and visual, but like, it really seems like the actual place where it’s like doing the best stuff is boring, mundane office work, even on the content side. Like, it’s, know, like your example is, it could write a poem or something. And it is really kind of crazy when you watch it do that. Like,
Sometimes your mind just goes like, how did you make that happen? But it’s like, I’m not going to probably read an AI generated novel at any point in my life. Like, I don’t think that’s ever going to happen. so, so what are your thoughts on like the content side?
Keanan (12:25)
Yeah, I agree. And I think this is a point that Dan Schipper from, from every makes really well is like even with content production, which is like the most, one of the most creative ways to do like that kind of work that you’re talking about. there is a lot of like mundane sort of drudgery type tasks that go along with it. Right. So like if I’m writing a blog post, I can have a really creative idea for something that I want to say. And that’s great. I still need to kind of think through like the structure of how it’s going to work. And that’s a, that’s a place where I use AI a lot is here’s my one paragraph pitch for this blog post. want to write.
Help me think through this outline and I’ll notice. OK, here are a couple points you missed or here’s something that you included that I don’t actually think we need to put in this post and so we’ll get to a point where it’s a really good outline and then I’ll do the writing. What most people think of is like the content production piece of it. But even after that and like I’m sure you’re kind of realizing this again as you’re starting up a new podcast, right? It’s like there are social media posts to write. There are like things to clip down and you want to pull out, you know, three interesting pieces from this in a form that’s.
you know, short enough for Twitter or X or and then one that’s a little bit longer for LinkedIn and like there’s all of this stuff that happens before and after the actual like creativity piece to make the thing you’re creating It’s most impactful, right? Like there’s all of that extra work as well. And so in my mind, like if AI can help with that, that helps more people kind of get their core ideas out there and more distributed. And you can make the argument that, you know, if we take the human effort away from that work, that means that.
social networks and everything else will just get flooded with all of this AI generated content. But I do think that at the end of the day, like there is all this work that surrounds the core work. Right. And that’s where I use AI a lot is to automate that sort of stuff.
Brian (13:57)
Yeah, I’ve had the same experience where it’s like, you know, I need to do a tutorial on this, like, give me like 10 different examples of what somebody would use this tool for. Give me like 15 block ideas that I could make a tutorial for. And it does really good brainstorming. It does really good outlining. It does a lot of that stuff pretty good in a way that like, you know, it’s probably now becoming dispensable to my workflow where I’m not sure if I…
I’m gaining or losing by not doing some of that brainstorming work myself, but I know that I don’t miss it, you know? but I think it’s the same problem. Like you said, with code, like the problem isn’t that like we’ll say WordPress needs more code. Like there’s plenty of code that is out there. And the other problem is like, there’s not a problem that we need more content. Like no one’s sitting there being like, there’s not enough content in the world. There’s not enough like podcasts and stuff like that. But like said, like taking out some of that, side work.
Keanan (14:28)
Yeah. Yeah.
Brian (14:50)
some of that stuff that like honestly like no you know how important is the tweet about the podcast episode like do you want to sit there and write you know all the tweets and stuff like that yeah i can kind of see that i can see why that’s the place to to sort of leverage it and where it’s becoming more useful for me i think
Keanan (15:08)
Yeah. And I think, you know, you can say like, how important is it for you to write that tweet or whatever, but I look at it both ways of like, it might not necessarily feel very important at the moment, but like the amount of podcast episodes that and brand new podcasts in general that I’ve discovered because someone posted about it to social media, either the guest did or the host did. And like, if that, you know, hadn’t existed, I would never have found out about it. So it is really important that like, if you have an idea that you want to get out in the world, you do have to tell people about it and tell people about it frequently and consistently. And,
People don’t always want to do that work or they feel bad kind of doing that self promotion in that themselves. And so these tools can be really good for that, especially if you have a strong opinion about like sample inputs and what sort of format you want it to be and that sort of thing. I think one of the reasons you’re successfully using it for ideas for blocks and things is because you have the WordPress experience and the perspective of what makes it interesting block and whether you’re kind of doing it explicitly or not, you’re giving the model information about what you want to see in a tutorial and.
Brian (15:49)
Hmm.
Keanan (16:04)
how you think about that. If I used it to brainstorm, you know, blog post titles or whatever, without giving it any sort of background knowledge or not having any knowledge myself in that particular area that I want to write a blog post about, I’ve noticed that the output is really pretty generic. But if I can give it, you know, good examples that I found online or just more information that comes from being me being involved in the topic, that’s where I found it’d be really, really helpful.
Brian (16:26)
Do you ever get concerned? Cause I know you have some experience doing like obviously teaching and like technical writing and tutorials, that sort of stuff. Like, I find myself going to chat GPT more often when I’m just needing some sort of educational tutorial thing that I know has been written a million times. And I know it’ll be able to like, give me a reliable answer versus like, you know, developers who, to write these sort of like technical blogs and things like that, where it’s, it’s pulling their
content and repurposing it, but it’s giving me maybe a more specific answer that’s like more useful to me. Do you get concerned that you’ll like lose the desire to write the post or like create the content or that will run out of that side of it or or do not worry about that?
Keanan (17:09)
Yeah,
think it’s a good question and I actually thought about this a while back and ended up writing a post on it of like should anybody blog anymore? Basically was something that’s the title was something around that and like I think it depends on why you’re doing it right. If you are just purely trying to. You know, right? Because you want to be a freelance writer, that’s I think going to be a lot tougher because like you said, one of the things that AI is great at and that I use it for all the time is to get a really specific tutorial. So it’s like I want to do this thing in WordPress, but I need to do it on this version of PHP.
Brian (17:18)
haha
Keanan (17:37)
and I need to use these three plugins, like maybe that’s a tutorial that doesn’t exist, but because of the collective knowledge of the model, it can generate that for me on the fly, right? So that’s a really good use case that is going to eat up a lot of the sort of SEO optimized, like generic writing. I think for folks who are more in like a consultative sort of role or who really want to show like, am someone who’s knowledgeable on this topic.
Brian (17:40)
Mm-hmm.
Keanan (18:02)
that’s still going to be important to write those sort of articles just to like show your expertise. And because you will have, you know, an opinion and a perspective on that topic that a generic AI model won’t have. I think the other piece of that that is only starting to now become important is that GENGBT and other models are starting to cite sources, which means you can actually get, you know, referral traffic that you might get from SEO in years past, like from these AI models now. And so the sooner you can like get your content in those models and the more comprehensive the
that you can have in terms of like coverage. I think you’ll still be able to benefit from, you know, even your work being synthesized by these models. Cause a lot of times people want to dig deeper and see sources and that’s becoming a more common pattern is people try to combat, you know, hallucinations and things like that.
Brian (18:44)
And I think like back to the beginning of the internet when people were just like, Hey, this is so cool. I can just put stuff on it. And I’m going to like put my recipe, you know, on my blog for fun and put all my recipes here because I want to, and everybody can do it to like now where when you try to read a recipe and it’s this like massive wall of text and ads and SEO stuff. And it’s just like the worst experience ever because that person’s trying to like, you know,
pay their mortgage based on like recipe blog posts or something like that. that long journey has been pretty brutal. But like you said, like I actually am optimistic that it’ll be like just another version of Google SEO. It’ll just be instead of SEO, it’ll be, don’t know, AI EO, no AIO or something like whatever, whatever it would be. Is there.
Keanan (19:09)
Right. Right.
Yeah, I’ve seen I’ve seen various acronyms floating around on LinkedIn. It seems like
people haven’t decided on one yet, but yeah, especially with like chat GPT starting to with search perplexity has already been doing that for a long time. Like there are tools that are really giving you the best of synthesizing and creating something custom, but also then citing sources and saying, hey, if you want to go further, here’s where you go. And so I think it’s, going to be really beneficial for folks to be there for sure.
Brian (19:49)
Do you have like, you know, since you have like the agency experience, do you have like tips or kind of like if you’re, you know, say WordPress agency, you’re a developer or on that sort of side of it and you feel like, everybody’s using AI and I’m not, or something like that. Do you have like kind of a good, you know, framework or like tips for people to kind of get their feet wet with that?
Keanan (20:10)
Yeah,
absolutely. I think about it in the kind of this is a common framework, but I’ve adapted it a little bit. It’s like the crawl, walk, run, you know, fly sort of framework, right? So if you’re not using any of these tools at all, there are a ton of like off the shelf tools that have AI built in that are already really useful without you needing to do anything yourself or customizing anything yourself. People started seeing that, you know, the proliferation of like AI meeting assistants, right? To sit in on your Zoom meetings and take notes. And that’s a really good example of a tool that you can just kind of download and use.
and get a ton of value from. I use granola personally. And basically just like listens to your meeting. You can take whatever notes you want to. And then when you’re done, it flushes out your notes with additional notes from the transcript. Let’s you link to the transcript and also ask any questions you want about the meeting. And for folks who are in a lot of like client meetings, especially agency folks like that, super helpful. You don’t have to think like, that they mentioned that there was something really important for this project that we’re kicking off. And like, I don’t know what it is. You can just kind of get the get the whole transcript.
There are also tools that help you, you know, take transcripts like that and turn them into text specs or technical requirements, documents, a lot of, again, the the work around the core agency work of producing whatever product you’re going to produce is you have to, you know, have meetings with clients, take good notes, turn them into tech specs for your developers or even for yourself. And there are a lot of just off the shelf tools that you do that. The kind of walk phase I think about as using tools like chat, GPT and Claude and these public chat bots where
Brian (21:12)
Mm-hmm.
Keanan (21:30)
Yes, there’s only so much kind of customization you can do, but you can upload your documents there. You know, you can ask it whatever questions you want to use its whole training data set to help give you good answers and automate some repetitive stuff. The run step is like taking those tools a little bit further with some advanced prompting techniques and you know, using things like Claude projects so that you can have, you know, a large language model that kind of has all your client specific context for one client or, you know, for a specific task you’re trying to take on.
Things like giving JGPT custom instructions so it knows how to respond to you specifically. Those are a little bit of like things where you can take that a little bit further. And then if you want to go further than that, right, this is the fly step. There are a ton of tools and I know you’ve been experimenting with a lot of them that are let you build your own kind of AI apps and AI enabled tooling. Things like Zapier and make and retool some of those. If you’re familiar with like code and able to code yourself like all of these large language models also have APIs and so you can.
You know, make direct calls to large language models from within your code as well. And that step gets a little more complicated because then you’re mixing the kind of fuzzy output of a large language model with the deterministic output of your code. And there are a lot of difficulties there, but you definitely don’t have to like jump into that at the very beginning. You can kind of start using these off the shelf tools and step your way up. And like we mentioned, because you don’t really know what these tools are good at until you try. That’s how I really recommend folks get into it and kind of just like up their exposure gradually so they can start to see.
Brian (22:32)
Mm-hmm.
Keanan (22:50)
This was a really good use case for this tool. We’re going to keep doing that or we spent two or three hours trying to get the AI to do this and we couldn’t really get it to work out, so we’re probably just going to shelve that for now. Maybe come back and try again when the models get better, but at least for now I can’t do that. That’s kind of how I think about it.
Brian (23:03)
Yeah, and I’ve been, I’ve been trying to ramp up my usage. so like one common thing has been like anything that’s like, we said like busy work, knowledge work, that sort of stuff. It’s like, even if I’m not sure what to do, I’ll just ask a model like, Hey, this is a workflow I have right now. Like where do, where can I automate it? Like automation really is a big part of it. Like it’s, it’s almost like the two really work together because it seems like it’s not just that you
want to use it to do this work, but you also want to automate processes and stuff like that. Half of the stuff I think I’m gonna use AI for, I’m actually just using Zapier to automate the thing, but it’s all part of just a different mindset of thinking about everything that you do and what you actually want to spend your time on and what is the busy work of your life that you don’t want to spend your time on.
Keanan (23:50)
Yep. Yep.
Brian (23:54)
The part you said earlier, which is about it being more ingrained in a lot of the tools, like we’re recording this on Riverside and when I’m done, it’s gonna do the show notes, it’s gonna do clips and stuff and I’ll edit everything, but it’ll give that first pass to me. Do you have a theory on, I was originally a strong proponent of the idea that like,
the operating system is going to handle AI for us in a way where you’re not even gonna need agents. you’re gonna be able to just, like, it’s gonna be so deeply baked into your computer that you won’t feel like you have to open up an app to use AI. It’ll just be part of the way you experience using a computer. And then I tried the Apple intelligence, like beta versions, and it’s by far the worst AI implementation I think I’ve ever seen. It is so…
terrible it actually makes you less productive with their stupid summaries of your texts and things like that. So now I’m not so sure about that. Do you have a like a theory on this if it’s going to be an OS thing or or what?
Keanan (24:50)
Yeah.
Yeah, I appreciate you mentioning Apple intelligence. just upgraded Mac OS on my work machine this morning and I got the prompt for it. So I clicked skip for now. I might have to set that up later, but after that not exactly glowing review, I may just skip it for now. But I think like there definitely is a large community of folks that want to run models on on their devices. And I think that unlocks a whole separate set of things that it can do. The trade off you’re really making is like, at least for now, the it’s really only the smaller models that can be run.
actual devices. And that’s why when you run chat GPT or when you run Claude, like it’s running on their infrastructure and you know, they’re there like the reason those more complicated models are able to run as quick as they are is because they have the specialized infrastructure for that. So I think that is one kind of current roadblock that will definitely be solved in the future. And bigger models will be able to run locally. But the other thing is that it’s still the case that people want, I think.
purpose-built tooling for a lot of these things. And so it’s like, if you have, you know, a sort of a granola like tool, right? That can listen to your meetings and take notes, they’re going to continue introducing features that are really good for that specific thing. Whereas an Apple intelligence, for example, doesn’t necessarily have the same incentive to introduce those features, right? And so, especially when you get to tools that are even more narrow than like take notes on our meetings, which is a pretty general purpose tool. I do think there will still be
Brian (26:06)
Hmm.
Keanan (26:14)
need for a lot of these tools that do one thing or a suite of related things very well. And the same reason that Chad GPT’s voice mode and things like that have had more initial traction than Siri, for example, or initial success maybe is because there are teams that are focused on producing that and not on producing an entire operating system, but also including some AI in there.
Brian (26:37)
Mm-hmm.
Keanan (26:37)
I think
it really comes down to like focus and the level of like detail you’re looking for from a solution like that.
Brian (26:44)
So the idea that one, I guess, universal AI that’s going to be able to specialize itself into all of these different areas, you think that’s probably far off and we still need human teams specializing these models to solve all these different use cases.
Keanan (27:00)
Yeah, I do. I look at it in the same way as like search engines a little bit like even though Google is the dominant search engine by far, we still have Bing. We still have doctor go. We still have all these other search engines, right? And we even have domain specific search engines, right? So there are legal specific search engines. There are scientific publication specific search engines and things like that. And so even though like one company is so dominant in that area, like there are still all these other tools and that calculus may change when we talk about AI models because.
the thought of like, well, if we actually achieve the artificial general intelligence that people say is coming, then maybe we won’t need any other platforms anymore because that tool will just be able to invent anything that it wants to at any moment. I think there’s an argument to be made there. But yeah, at least for the foreseeable future, think people like solutions that feel right to them or feel like they serve their needs pretty fully.
At least for now, like the only way we get that is by having teams that are focused on those problems and building solutions for them.
Brian (27:55)
Yeah, I think that makes a lot of sense. And, and like you said, like really understand, you can understand why the incentives at Apple are probably not. I’m just surprised. You know, I always thought Apple was not going to be one of those like, let’s stuff a feature in cause it’s popular, but it really does. Did come across as that where you’re like, if this is people’s first experience of these, of this like AI in general, I that’s not helpful at all. yeah, yeah. I.
Keanan (28:07)
Yep. Yep.
Yeah, yeah.
Brian (28:19)
There’s other things I wanted to talk to you about. Like I’m really curious about the concept of open source now, but I think we’ll have to save that for another conversation. Cause I think there’s, there’s a lot of other sort of issues around, you know, the way AI works. But for now, maybe you could just give me like where people can go to learn about, learn about you, follow you, learn about AI from you.
Keanan (28:41)
Yeah, it’s a good question. I think the kind of two best places would be LinkedIn. That’s where I’m sharing a lot of my stuff these days. I’m also on blue sky as of recently. So at K Copenhagen dot B sky dot social I think is what the kind of default handle is there. But yeah, those would be the two best spots. I post stuff there all the time and feel free to DM me as well anywhere anywhere you can find me.
Brian (28:55)
Hmm.
Awesome, man, thanks for this. You’re my favorite person to talk to about this stuff, so I appreciate it.
Keanan (29:07)
Yeah,
absolutely. Thanks for having me.