Podcast audio transcript

[0:00] Music.

[0:05] To the DrupalEasy Podcast. This is season 17 episode number one and I am Mike Anello, your host. First off, I really hope everyone enjoyed our April Fool's Day episode. If you missed it, you might want to take a few minutes and check it out and get a really good laugh at my expense in today's episode. I'll be talking with Jamie Abrahams, one of the maintainers of the AI module. We'll be talking about all things AI in the Drupal community. But most importantly, we'll be diving in a bit into the new AI module that's bringing together many of the various AI module maintainers in the Drupal community. But before we get to the interview, I would be remiss if I and at least quickly mention DrupalEasy's long form professional module development course. Our online semester begins on August 13th and runs twice a week for 15 weeks. And honestly, if you really want to learn Drupal module development, I encourage you to check out the course at DrupalEasy.com/pm.

[1:08] Music.

[1:13] Welcome to the DrupalEasy podcast, Jamie Abrahams. How are you? I'm good. How are you? Good, good, good. It's great to meet you. I don't think we've met previously. So it's always nice to meet someone for the first time when we're doing the podcast. So, Jamie, you're here today, you're going to talk to us about uh kind of the state of AI integration modules in Drupal. And we're going to specifically focus on a new initiative around a module simply titled AI that actually has the AI name space on Drupal.org as well, which is kind of nice. But before we do that, I am not even going to attempt to pronounce your Drupal.org user name.

[1:53] So go ahead and pronounce that and give us a little back story where that name comes from. Yeah. So, uh it's Yaz Satan. So, uh you know, Predator with Arnold Schwarzenegger, if it has punching in it, then it's a movie I've seen.

[2:05] Yeah. So, uh I was obsessed with a video game Alien versus Predator when I was younger and the predators in their own language and in the comics are called Ya Jars and one of their names is Satan.

[2:15] So I had that since I was about, I don't know, 10, it stuck as a teenager. And, uh, yeah, I got involved in the Drupal community pretty, from a, you know, pretty young age as a teenager. And so that's became my handle and in most business contexts I don't use it. But, uh, when I'm doing Drupal, that's still my name. Yeah. Your account's over 15 years old. So you've been around Drupal for a long time. How did you get started? Yeah. So we got started, me and my co founder, Andrew Belcher, we were building, we were aiming to build a company around church management software. But I was always excited by online communities. Ever since I was about 10 or 11, I was playing online like AM O so I was excited by online communities and the impact it could have on culture. So we started this company and we thought the CRM aspect of what we needed would take about three months to build. And uh 15 years later, we're still working on native Drupal CRM to a, to a degree. Yeah, we've diversified since then. So originally it was mainly churches and church organizations, but we've got quite a diverse portfolio of clients now. So our largest is a uh train management delay repay system for train operating companies. So it's a system where if you're, if your, you know, train is delayed, you can upload your ticket into it. It validates that your ticket was valid for that journey. Calculates how delayed it was cross referencing a very large database we built, it's got fraud checks. Customer service repays you the amount and we provide that directly for the train companies.

[3:42] Yeah, and that's why we got into AI because as you can imagine, there's a lot of scope for AI automating things there. Absolutely. So the the companies freely give. Yeah, and you're the director slash co founder. Correct? Yeah, me and Andrew. Um we co-founded, well, we started working together a long time ago just over 20 years ago and I did some coding when I was 16. It was like VB script, CS S was this new thing to me at the time. Everything was tables based layout. I got paid to build some websites, but the websites would have gifts of like snow floating down that kind of stuff. But really early on Andrew could just do everything that I needed. And I actually had this team of like, 14 to 16 year olds at our church building things. We were building this, we're going to try and make an open source online community platform for churches at the time based around PHBBB. Yeah. So, uh, that's, that's where we started. Didn't VB script, didn't that become a SP?

[4:37] Yeah. Yeah. So I was doing a SP Yeah. So I was doing a SP that's where I got my start as well. That's what you, well, actually, the very first thing I did in web development was and this goes back, I, I don't want to even count how many years was and this is back in like the mosaic days. That's how, that's how old I am. But creating CG I script using object Pascal via Borland Delphi. And then from there, I got into VB script and a SP and a sp.net and Cold Fusion and blah, blah, blah, blah, blah. And here we are talking about Drupal so many years later. But let's, let's start talking about the past.

[5:18] Let's fast forward to today to the future because really the main topic that I want to talk to you about is this new AI module? And I was super excited when um you know, you wrote a blog post I didn't even know what the date was that in June, maybe when you wrote that or May or? Yeah, I think so. Yeah. Yeah, fairly recently and I was very excited to see it because the goal of the AI module and correct me if I'm wrong is to basically provide kind of low level common functionality for AI modules in Drupal. Yeah. Right. OK. Very good. And so.

[5:53] If you're kind of, you know, before your announcement about the AI module, there's more than a few AI modules on Drupal.org in the contrive space. And that's always kind of my, whenever I see stuff like that, we saw that like with Whizzy Wig, editors back in Drupal seven days and the gazillion different ways to lay out pages, you know, even today for me, it's, it was great to see. OK, a lot of these AI module maintainers are saying, OK, hey, before we go too far down the road here, let's team up, let's team up and put common functionality in one place. That way we can all build our stuff, you know, faster, better, stronger on top of the super strong foundation. So what was kind of your view of the current state of AI modules leading up to you and the other maintainers getting together and saying, hey, let's, let's join forces.

[6:51] Yeah. So we, we'd actually been talking a lot. So everyone who had been doing AI stuff, we'd been trying to organize stuff on hash AI on Slack. We'd been trying to do regular meetups for A good year and a half. So it's been, it's been a while that we have been collaborating behind the scenes.

[7:07] The big thing was this abstraction layer and about the timing of that. So to some degree, I actually um I pushed back a little bit. So, you know, Marcus really wanted to do it. A lot of developers really wanted to do this abstraction layer. And for me, I was really keen on finding genuine uses of AI that would work and impact people first before worrying about like underlying technology because a year and a half ago AI was very cool in theory, but there were so many demos that seemed magical, but when you tried it out, it just didn't get things right, just that just slightly below that level for it to be useful. It was just this interesting demo that you wouldn't want to try out. But gradually over the last year, we've seen real use cases of it making a big difference. We've seen a lot of downloads of the open AI module. People were very excited by the CK editor stuff, Marcus with AI Interpol was doing amazing work and RAA I search stuff was starting to take off. People were really interested in it. So it happened around about, yeah, June, just before June that given these things were really taking off, Kevin approached us. He also noticed that the AI name space was something that someone else was sitting on it for quite a long time and there wasn't anything happening with it so we could take over that AI name space and we thought if we have that name space now's the time.

[8:27] So you mentioned a couple of names. I just want to introduce some of the players that we're talking about Marcus and his, I forgot, I'm blanking on his last name. All of a sudden, Johan Johansson, he works for freely good with you. He is the maintainer of the AI interpret module which. If you haven't checked this module out, I'm not talking to you. I'm talking to our listeners. It's amazing. And Marcus has done a series of videos using AI interpret which is just jaw dropping, you know, some of the demos he does. So that's Marcus that you mentioned, Kevin, Kevin Quillen, who is the maintainer of the open AI slash chat GP T module, which, I'll go out on a limb and say that's probably the, the AI module that's gotten the most attention so far.

[9:16] That does a lot of generative stuff, you know, mainly on, you know, entity, a edit pages, you know, generating tags and, you know, providing AC K editor button for you to give a prompt and it will generate some content for you in CK editor. Yeah. OK. So I just, I wanted to make sure that we weren't short changing any of the players or the modules that you referenced there. Yeah. Yeah. Mar Marcus worked for a large pr firm for a while building this AI stuff. So he's been also working on AI stuff for a good year and a half and he's new to freely.com but he's new to freely give. Yeah, so AI interpreter was his thing that he's really been pushing for quite a long time and then workflows of AI is that place with all the videos. And yeah, seeing that is the speed at which you can build really cool AI applications actually do something. It's sort of like 20 minutes, half an hour, a couple of hours in the browser or in the browser.

[10:10] So you kind of, you know, set the table for us that there were and there's more than just, you know, you know, the open air module or the open AI interpolated module. There are also odd mentor and a bunch of others, but there's a few things that all of those modules have in common and that's the functionality that is initially going to move into the AI module for, for, you know, in some respects it's already there. So I just kind of want to cover that real quick and just talk about how.

[10:39] You know, if you're going to be making calls to AI providers, you need to authenticate with them. So every module kind of had to build their own method for doing that. And so that's one of the things that has been brought into the AI module. So can you talk about that real quick? Like what that looks like? Yeah. So the core, what we started with was an abstraction layer because uh everyone was focused on open A I. And whenever we'd show off these cool demos, the first thing people would ask us is can we do all of open source? Meanwhile, Claude has been just getting so good like Sonic 3.5 is amazing. So whereas it, whereas before open AI was the only player really. Now, it's not the case. There's a lot of different providers, Mistral if you're in Europe, especially if you're French speaking is a lot better than open AI and Claude in a lot of situations.

[11:25] So we started with OK, we need to do this abstraction layer and we, we built on top of a module called LM provider made by someone called Michael Gao, Seo Gao. And so he's part of this team and Marcus, he took his module sort of made it a bit more Drupal, made it something that everything else could work on. And that was step one that we built of, we built on top of alpha alpha, one was being able to do the core of what the open AI module could do. But now you have this abstraction layer where you can choose which provider you want, set up all your keys, set up all your authentication, but you can now swap them out. So anything you build on top of the AI module, it will be able to support all of the models that have a provider. Yeah. So the so kind of the goal or an example of that is, you know, maybe today you build a solution that uses one provider to, you know, for some AI bit of AI functionality, but maybe six months from now, there's another provider.

[12:19] Comes out that has more powerful functionality for what you want to do. You don't really have to redo anything other than just swapping out the provider if you have, you know, a configuration for on your site. Exactly. Yeah. And it's really powerful for from an ethical standpoint as well. Like if, if as the models go in a direction that you prefer, if there's more environmentally friendly models, if there's ones that are more and more open source, you can now choose that. And a big reason why now's a good time is the open source models. They get really good. They're actually genuinely useful. They weren't a year and a half ago. Whereas now there's plenty of situations where you would be totally happy using a completely self hosted, have the model on your own server. There's no privacy issues, it's all open source, there's no censorship issues or control or worried about being banned. So it's very powerful by being able to build things on an abstraction layer.

[13:09] So the other side of that abstraction layer is that to make a call to one of these providers, you know, they all have their own API so like if you want to send, you know a chunk of text to one provider and say, please extract you know some keywords from that, you know, each provider might have a slightly different API call to their service in order to get that data back. So there's abstraction layers for both providing the the data to the provider, you know, in a common way. So that to make everything swappable as well as receiving that data back, so that the data that comes back from the provider is then translated into a common interface. So that what you get back is in a consistent format, regardless of the provider that is giving you that data.

[13:58] So it's my understanding that those two bits, the plug in architecture for the providers and those common interface layers, that's kind of the the MVP of the AI module. Do I do I have that about right? Yeah, you do. And and one of the reasons why AI Interpol is so powerful is because it takes what you've just said and then and then pushes it further for not just AI but all of these services. So with Lang chain, one thing Marcus was struck. So lang chain for is sort of the starting point of everything interesting in AI about a year and a half ago, Lang chain was this library where everything cool that anyone was building, you'd see it in lang chain. So it became a big inspiration for all of us doing anything with AI Python library very code heavy. And I was really into this idea of chaining prompts and services together to build complex applications by doing things step by step. But one thing Marcus found was every single service you use would send data back in a completely different format. So you'd have to write a lot of code to make it pass it on to the next step. But by putting everything in fields, by making it so that every single time you send something to a service or an AI module, A model, etc you put it in a field like the date field, it's now in a very consistent format. So the AI module and the abstraction layer, we're doing that for just standard prompts, then we're extending that with interpolate AI automator to do that with training things and putting fields things into fields.

[15:17] And then similarly with AI search, we're doing the same kind of thing for using AI to search things, using something called rag or vector search or embeddings. Yeah. And even just hearing you talk about it now, like it gets me excited all over again from the first time I really kind of understood this module is going for because it is, it's really exciting.

[15:36] You know, because it allows you and if you go to that workflows of a i.com site that we talked about earlier that has all the youtube videos about AI interpret. It's really cool. Some of those demos, the way it'll take some, you know, some, some author input, send it to one AI provider to do something, get that response back and then send that response to a different AI provider to do something slightly different or add on to that. And so just the fact that you can like chain these calls to different AI providers and have it all work somewhat magically right now. It seems in the browser is just amazing. So, yeah, I do want to mention um we're gonna have a lot of the links to stuff we're talking about in the show notes I want to take. So let's take another step back. So we kind of talked about actually, it's not really a step back yet. We talked about kind of the MVP for the AI module, but that doesn't really cover like the functionality that you see in the open AI chat GP T module. That exists today or the stuff in AI interpolated augmenter module? So how far will the AI module go, will it just go so far to provide this base level functionality? And then it will become a dependency of open AI chat EP or, or Interpol or is the plan to bring some of the functionality from those other modules actually into the AI module? Like where's the line?

[17:05] Yeah. What, what our goal was is it is kind of like um an API module, but we wanted to make it so that when you install the AI module, it would give you all the functionality you'd expect from an AI module. So most of the functionality for open A I, the open AI module, we are porting that in to the AI module. So already in Dev Frederick W has been really helpful. He works at caliber at the moment and he's helped move a very simple AI translate module that existed and he's put that into the AI module. Uh A lot of the CCA says stuff that we had in open A I, they're now in the DEV version of the AI module. So we're trying to bring a lot of that in. Um We're currently working on AI search so that used to be called open AI embeddings.

[17:48] And the big thing we're working on at the moment should be around any week. Now really is being able to have like a chatbot where you can talk to your content and find things in your content and that will be there. So we're gonna try and include most of the kind of simple but obvious functionality you'd expect from an AI module in the A module. So when you install it, it will actually do something. We, we're kind of using, we're using like search API as our inspiration more than C tools as our inspiration. So search API out of the box, it, it does do something, it works. You probably wanna have solar elastic search for it to really work well. But out of the box, you have facets, you have the ability to do search, you have all the U I to control what goes into your index and we're trying to do something like that. So we'll have a chat bot out of the box, we'll have it plu.

[18:37] So in contrib, there'll be loads of different US and UXS and approaches to doing chatbots like the little chat thing at the bottom, right of your screen, a full screen one in a sidebar, all of these things can be done in contra. But out of the box, you'll see it.

[18:51] And the reason why is because although it is an API module interpolate, uh automator will be useful for building things AI search will be useful for making complicated applications. We wanted to include some examples of how to do it and also have them useful straight away in the core AI module.

[19:09] Yeah, you just mentioned something which reminded me and I think you actually just confirmed it for me as well. The AI interpolated functionality that's moving into the AI module, you're changing the name to Automator. Is that correct? It's changing to Automator and it's pretty much already there. So, in, I think it's either an alpha in the DEV, it's been moved and then all of the providers, almost all of them, apart from a couple of more niche ones have been moved across. So it's definitely something we're, we're internally using it for some of our clients. So it's something that it's still, you know, need to iron out some bugs. But we're pretty close to being able to release an alpha for that. I want to circle back to providers in a couple of minutes. But before we do I, I'm really curious about.

[19:51] You know, you mentioned that you've been talking to the other, you know, or not the other but other AI module maintainers contributors over the past year and a half. Has it been an easy sell to get everybody on board with, with the AI module and this common functionality? And I'm not looking for you to name names obviously, but is everybody on board with this or are there AI modules or like uh we're just gonna, you know, go on our own for now and see how it goes.

[20:17] So I think because I didn't write a lot of the code, I think I can be a little bit arrogant about how good the abstraction layer is. So I think it is really good. And so what we found is it's been a lot easier a cell now than it was six months ago because a lot of people will look at it and you can tell when I'm talking to some other organizations, they don't really want to work together, maybe they want to do things in house, they've been experimenting themselves and then the developers will like look at the abstraction layer and be like, oh, ok. We would have done everything like that. And then, oh, you know, there's lots of little things we've done to help things out, like make it really easy to expose the module selector in your forms, put in loads of settings. We've done a lot to help with security to make it so that the moderation layers on by default to make it so that you're less likely to get banned. And we've thought about models that don't have moderation. So there's all these things that we thought about that the more people have looked at it, they've gone. OK. There's no reason why I do my own thing. We were initially getting less excitement with the more advanced uh sub modules. We've got like AI search and AI automator. Uh We've also got integration with EC A but I think nobody's gonna rebuild EC A from scratch now. So I think when you look at these things, a lot of these developers as they see how they've built and we have really built them to be tools that other people can use. I think it won't take long before everyone collaborates around AI search like they're already collaborating around the AI protraction layer. Yeah, I, I actually purposely did not put anything about EC A in a rundown.

[21:46] Because I, I'm a big fan of the EEC A module and I think the integration between what the different AI providers can bring to the table as EC A actions like that. That is going to be, it's going to be fantastic. All right. And now I'm thinking about it. I'm pausing because I'm like, I really want to talk about that, but I know that that's going to be a deep, deep dark. Well, just one thing about that is we've actually got a Google Summer Code project right now and we're mentoring someone called and his thing he's working on is particularly EC A and AI and he's particularly been trying to work on the UX side of things.

[22:22] And so he had a couple of ideas but his, I, we'll see if it, I mean, we're halfway through and this is quite crazy idea. It's quite wild. It's quite interesting. We'll see if he can get to the end of it. But his current dream is to make a chatbot that allows you to describe a EC, a style process and then the AI will just create the whole thing for you. And I actually know someone who he was telling me about some software that he's been working on. Where when you do these workflows, you can draw a diagram of a, of an automated workflow that you want just on a whiteboard or piece of paper, upload it. And AI will like create the whole thing. And we're hoping to be able to do something similar with the ETA. So not just use EC A to use AI within an AC A thing, but use AI to help you create these flows. I have seen similar similar wish list items when talking about Star Shot and experience builder. Like, wouldn't it be amazing? You just draw the layout and then have AI generate that layout in Drupal for us, which, I mean, it sounds like that there's a lot of, there are providers out out there that provide pieces of that and who knows, maybe in a couple of years that'll, that'll.

[23:30] Well, I mean, a couple of years or maybe a couple of months uh cos it is definitely something that we want to build by September. So there has been, we've, you know, I've been speaking to Laurie a lot. Um We've been speaking with the star shot team for Drupal Com. Barcelona, my dream. I don't know if we'll be able to achieve it. But my, we, we Marcus has worked on some of the core basic building blocks of this. So he's got some demos. If you go to the free to give youtube channel, you can see some of these demos where he's drawn components and it's created components using layout paragraphs. Uh So we've got like the basic building blocks of a lot of these things at Drupal DEV days. We did a demo where you upload a PDF of a whole bunch of restaurants and it created, it used AI agents to create the content types, the content views had views, filters. And then meanwhile, there's uh Drupal Ninja. He's been working on a Drupal AI module on github that is all about generating code. And he's been looking at things that have been you upload like just a design and it creates a story books. So it creates like, you know, it creates layouts but it's all cs S javascript, et cetera. So there's all these pieces there together. It's hard not to get excited about this stuff.

[24:39] Yeah, I think it needs funding. So, one of the things that we're exploring is trying to get funding for this before we can go ahead with it. But I reckon it will be possible by September October, you'll be able to like put in your URL of a wordpress site, your own, click a button, whole thing will be moved to Drupal. I, I think that is within reach of like months. Oh my gosh. So I uh I, so when you say that is the thing, it's like, yeah, well, that's what, that's what I want to build. I just need to get it funded, but we're, we're ready for it. We've got plans, I've got write ups of how we're going to do it. The crazy part is, is we're actually recording this in July of 2018. So that's half now.

[25:17] Um Yeah. Ok. Let me, I'm gonna get us back on track because I, I could, I could ask about that stuff a bit more, but I kind of want to talk about move to the business side of things.

[25:29] Because I feel like when like the average person on the street, Here's, you know, the phrase AI or things about AI in July of 2024 they're thinking about generative A I, you know, the, the, the chatbot or mid journey. But I, I actually think, and I have yet to implement any AI solutions for any of my clients yet. I'm actually getting pretty club talking to one of them for a while about what we want to do. I actually think there's much more potential for Drupal developers. Um kind of the, the non generative side and kind of the stuff you mentioned before that the rag stuff where, you know, actually parsing site content or, or an organization with content, and then making that available, you know, to their customers as part of a chat bot, or even doing this, you know, simple kind of behind the scenes, stuff like sentiment analysis of post and, and things like that. So I'm just curious for, from your perspective, as someone who's been using AI with clients, you know, for definitely more time than I have. Like, what are your thoughts there and what you're seeing as far as like, where do you think the business is gonna be in 369, 12 months.

[26:47] Yeah. So from a um we've been thinking about this a lot and, and we're using this word like primitive to describe like what are the things that we know? AI is so good at that. If you have a situation that requires this, it's really likely to help. And one of the best primitives is making unstructured content structured and it's something that we've tried with code before with things like RED X, you know, you could kind of do things there a little bit you can do searching for specific words and looking at patterns, but it's never really been very good. Usually, if you're gonna do anything with code, anything in the programming space, you need structure content first. So we've been really excited by JSON XML, all these things to try and have everything in the world already structured so that we can do stuff with it. What's exciting about gen AI and L MS is they can take this unstructured stuff and very quickly and easily put it into some structured format.

[27:39] And then once you've done that, there's a whole space of problems that you never thought you could solve that you can suddenly now solve with A I. So a really simple example, a lot of our AI automator use cases start with web scraping. So you're doing a research project. My wife did an hr research project where she was researching HR awards and she needed to find a whole bunch of hr awards. Look at like, how much do they cost? What's the process for applying for them? How likely would they succeed by comparing it with her company's like successes over the last years. And it all started from a web search or a PDF she downloaded, read, put things into a spreadsheet. There are web scraping tools that end users can use, but they're very complicated because you have to go to the website, decide which is the bit of the website that has the useful content. What do I ignore?

[28:27] But with LL MS, you just scrape, you can sometimes scrape raw html, you can ignore everything. Uh You can scrape a AAA picture of the site and pull it off. And so I think that's something that's really powerful. So when you say non gen A I, things that like, um you know, people have done with machine learning like sentiment analysis, the thing is sentiment analysis is quite expensive. It takes quite a long time. You need quite a lot of data. You need a lot of knowledge of how to figure out how to train an ML model to do sentiment analysis on your comments. But actually if you treat chat G BT or Claude as an ML model that you've trained and just make it ask it to pretend to be one, it's pretty good. So you can get it to do sentiment analysis for you and it will be, it will be pretty good out of the box. It won't be maybe not as good as a, a well trained ML model, but it will get you going really quickly. So I think that's one of the big uh use cases is there are so many areas of automation, like, you know, the train one that we talked about of automating claims. So many of them that start with some unstructured content. It might be an email, it might be a file piece of paper.

[29:32] That if we have access to that, that we can then put it into some structured format in Drupal Json or whatever, you can suddenly do loads of things with it. So that's the area I'm most excited about from a business point of view is helping organizations automate a lot of things with that starting point. There's at least a couple of Marcus's videos that are related to that. I know that there's one of them. There's um I don't know, I don't remember the name of the module, but it's an AI Interpol scraper module of some type where you can either scrape the whole page or you can actually pass it HTML element on the page that you want to scrape. But then there's also one and I don't remember the exact provider or the recipe that he used. But the example was you take a picture of a receipt and the provider actually would parse that and then it would, it would actually create structured data out of it just based on.

[30:27] You know, there'd be a line that has tax colon and then the amount of the tax and you know, the LLM would basically associate the, the amount of the tax with the word tax. And that's how you kind of get your structured data. So a lot of that stuff is already possible. It's just a matter of, you know, using these building blocks. And that's kind of where I wanna, I wanna go next with this discussion. This will be our kind of our last big topic. But for me, one of the biggest eye openers, as I'm starting reading more about AI integration Drupal and all these modules was just the sheer number of providers.

[31:04] You know, businesses out there start ups that are providing API s to their models, that are specific to certain tasks. And so I guess, you know, if someone is brand new to, you know, not, not brand new to A I, but they're looking into like how I want to start integrating AI to some of my Drupal solutions. Like where do they go to start learning about all of the potential for, you know, all of these providers and all the different things they do? Like a lot of what I've learned so far quite honestly has been by watching Marcus videos. Like, I mean, I think that's the answer. Yeah. Yeah, I mean, I mean, there's, there's an, there's a model out there and an API I think it's perfect prompt.

[31:54] And you literally send it a prompt like an AI prompt and maybe give it some extra data. And it will send you back a better prompt that you can use with a different AI model. So learning about all the available models, I mean, that's like one level of knowledge and that integrating them is another level. But like where would someone start from a, from a non Drupal standpoint?

[32:19] I actually think, yeah, workflows of AI is probably the best place I've seen. It's one of the reasons why I really wanted to work with Marcus is I would speak to all these people who are doing AI stuff and they would tell me about embeddings and vector search and transformer models and neurones. And they would talk about all this tech. And then I saw Marcus and he was just like he just made so many, there was like 60 videos of him. I can build this, I can build this, I can build this and that to me is the exciting thing about AI is not like how it works, but that it works like all the ways in which you can just use it for these things very quickly. And because you can build things so quickly, it doesn't matter if it doesn't work perfectly. The first time around it took you 20 minutes. So you just tweak it all the time. And so I, every almost everyone I've spoken to when they're trying to get into A I, I actually think that's a really good, particularly if you're a developer work flows of AI is one of the best things I've seen that's out there because everything else starts with, like you can, you can look at tutorials on line chain, you can look at tutorials on Lama index and they are really good. Like if you look at Llama index, Llama index is sort of, I would say it's sort of beating lang chain as the go to library. It's sort of, it's better built, it does more advanced stuff. What you can do with Llama index is way cooler and way more advanced than what you can currently do with AI module in Drupal. It's really, really, really cool and cutting edge, but you have to write a lot of code. You have to do a lot of work before you get going. If you're going to use it to like index all your content, it's gonna take you hours and hours to figure out how to store anything before you can try something out.

[33:47] You try workflows of A I. You can follow one of those tutorials. Uh We did 11 recently uh that we would recommend is with Meranti Tech Talks. Meranti is enterprise hosting and in the Drupal community, they own amazing. A lot of people know of amazing, right?

[34:03] And so me and Marcus, we did a podcast there recently where we built an AI application that was, it aimed at like uh an event organizer, like a corporate event organizer and they want to research venues. And it's this tool that you can put in the URL of a venue and it will extract. True information, like how many people, the capacity, uh number of rooms categorizes it, but also can extract things from Google reviews and then summarize them and give you strength and weaknesses depending on the use case that you've, you've told it.

[34:33] And we've had this thing of builder AI application in 20 minutes because it's a 20 minute podcast. I think it more takes about an hour because there's a lot of fields that you have to click. But if you go to that podcast, Marcus has recorded a 3.5 hour one where he goes from the beginning, like you have to install D DEV yourself, but you've got D DEV on your machine, you've never used Drupal before. You know, nothing about it. And he takes you through all the steps to build this very complex a application. And we've got a uh non code uh uh project manager on our team who she started working through it and has found it really helpful. So I actually genuinely think that the, I know I work with him, but it's a really, really good starting place everywhere else will always get you having to worry about getting your DEV environment set up in this complicated way and learning about really underlying stuff about how the open AI API works that you don't really need to know to get going. I think there is a certain level of like vocabulary that is very helpful to understand when working with A I.

[35:33] But I'm past that. II I think I, you know, I understand exactly what you're saying where it is very helpful just to see stuff happening. Yeah, there is certainly a year ago it was true that you needed to know all this vocabulary. Um One of the things I think I really like about the AI automator is, is it does all this training stuff in a way that's very intuitive. That doesn't, it's just like I've got a field and it's a summary of reviews. So I'm gonna use AI to write a summary of reviews and where do those reviews come from? Well, I had another field that scraped it all from Google reviews. So I'm gonna insert a token and you are chaining and you are doing prompt engineering and you are thinking about context limits and all of these terms, but you don't actually need to think about those things to get it to do what you want. Um And there's a really cool, I, I used to give these talks before and there was this quote by someone at open AI who said that everyone's worried about what do they have to do now to have a job working with A G I. And he said all of us are used to working with HG I human general intelligence. And we're all having to figure out how to communicate what we want, ask people to do things for us evaluate if the thing they did was good enough.

[36:40] And that's a thing that you can be better at or worse at. You can be better at knowing how to use language. You can be better at evaluating how something has went. And this ex post was saying, like, actually focusing on that I think is more important than knowing about the specifics of AI because as AI gets cleverer and cleverer and cleverer you need less and less of that stuff. So I feel like, uh, yeah, that's, that's my view.

[37:03] That's really interesting. Yeah, that is basically, it boils down to communications, right. Being able to communicate what you want in a, in a way, give it, you know, clear parameters, clear goals. Yeah, that's, that's pretty interesting. So the last thing I want to bring up is, you know, we've been talking about a lot of, you know, different AI providers, most of them do have a free tier, not all of them though. So what we haven't really talked about yet is, you know, how much is all of this going to cost? And don't worry, I'm not gonna ask you, you know, to answer that. But I mean, that's, that's a valid question for someone who is going to, you know, watch work flows of AI or someone who wants to e even as you mentioned, you know, out of the box AI module, you want to have a chat bot that works with the, you know, the organization's content.

[37:59] I mean, is, is everything that you're envisioning in the, the, the AI module going to be available at a free tier without using a paid tier of, you know, of a, of a provider or like how, how do you envision that kind of plan out over the next few months? Yeah. So to start with, um when we, one of the tools we built, uh we wrote a prompt with open AI and it was about six cents per prompt. And we were expecting to do maybe something like 60 to 70,000 a year. So that would be an idea of like the costs associated with. It was just an example to give you kind of indicative costs if you, when it comes to free tier, most of the API S, you do have to pay all the time. The only one that is free is hugging face. So we do have the hugging face provider and that gives you loads of open source ones and that is free, but it breaks all the time like it's you, you, you, if you want to use it, you have to just keep clicking, run, run, run, run, run and to try and get in because they limit how many things they'll process per like second almost. But the best way to get started is to get it on your own machine. So there's a tool called Llama LM Studio that you can download, get it on your Windows machine. Mac Linux, it is proprietary, but it's built on some open source software that allows you to download a model onto your own machine.

[39:20] And we have a LM studio provider. We also have something called O Llama that is similar, but it's purely open source. The U I isn't quite there yet. So that's why I'm recommending Llama Studio LM studio at the moment. But there you, when you're building things, you can test out really good open source models all on your machine. And that doesn't cost anything apart from obviously the energy costs of your laptop running it. And there are some models out there like Gemma seven B that will run on machines that don't have a graphics card. So you can do basic things there. But yeah, if you're gonna use open AI or clawed, you're probably gonna have to pay something per yeah, per thing that you run. But usually when you're, when you're testing things, you probably don't need to spend more than like five or $10 to try stuff out. Uh So, but it does mean you, unless you have a good business model, you don't really want to be putting a chatbot on your website that is open to anyone to use because then the cost could go up.

[40:15] Yeah, I think that's going to be a challenge for, you know, there's a, there's a technical challenge for Drupal developers of just learning all this stuff and understanding how to implement it. But then there's also the business challenge of like, how do you estimate for your clients, how much all this stuff is going to cost? Not from a building standpoint, but just from a, you know, an API standpoint, I think that's, that's a bit of a challenge. It is a challenge as functionality. We plan to build an AI in the AI module. So we've started, we've started plans for that. It's all about prioritizing what's important. But one of our goals is to have a price evaluating tool. So when you do your test runs, you can do a test run and it will tell you how many tokens it use, how much it costs. And then you can use that to extrapolate when you do it in real life. So it's one of those things that we're hoping to work with at the, in the AI module.

[41:08] All right. Well, this, you know, Jamie, this has been fantastic, you know, every time, uh you know, I try and spend a few hours every week, uh more lately, just playing around with this stuff and learning more. And I get excited every, every time I, I dive into this again, I get more and more excited and talking to you. The feeling is the same. There's so many other kind of avenues of discussion in this area. But, you know, I'm, I'm just super excited for, for, you know, the future of, of AI and Drupal. I know that, you know, there have been, I think your, your blog post might mention that you're hoping that there could be some type of AI integration with star shot eventually in the future. You know, even what you talked about earlier as far as drawing the layout and having that automatically just come to fruition is, is just amazing. So I, I can't thank you enough for your time. I can't thank you, you know, for your support of the AI module and, and Marcus's time as well. And uh yeah, just thank you. I can't wait to see, you know how the AI module develops. Cool. Oh, thanks for having me. It's been, uh, it's been a lot of fun. Super duper. Yeah, I don't know if I should also mention, yeah, the star shot stuff and what I want to do by September. Those are just my opinions and ideas. Yeah. No one's gonna hold you to them. Yeah. No one's, yeah, this is not an official, uh, this podcast is not an official document of any type. So you can, you can say whatever you want.

[42:34] All right. Well, cool. Thank you very much and hopefully we'll get you on another podcast in the future.

[42:41] Thank you for listening to the DrupalEasy Podcast. Don't forget to check out all of our long form Drupal training courses at DrupalEasy.com and stay tuned for the next episode of the DrupalEasy Podcast.

[42:56] Music.