The Marketing AI Show—the podcast that helps businesses grow smarter by making artificial intelligence approachable and actionable—is BACK with another episode.
You can listen now on your favorite podcast app, or keep reading for more on what to expect in this episode.
Episode 9: Britney Muller, Data Sci 101, on Key AI Use Cases for Search Engine Optimization
In this week's episode, show host Paul Roetzer, sits down with Britney Muller (@BritneyMuller), Founder of Data Science 101 and former senior SEO scientist with Moz. Britney is a prototype of what we call a next-gen marketer—or someone who understands what is possible with smarter technologies. Someone who knows that in order to deliver the personalized experiences modern consumers demand, marketing must become smarter. It must become marketer + machine. Britney started in the PR field, and is fully self taught in machine learning.
In the episode, Muller talks about machine learning from her perspective, and the different tools and resources available to help apply AI in other careers. She also discusses:
Specific AI use cases for SEO.
The impact of AI on marketing and SEO professionals in the next 3 - 5 years.
Advice for marketers who are just getting started with machine learning and AI.
Watch the Video
Read the Interview Transcription
Disclaimer: This transcription was written by AI, thanks to Descript.
[00:00:00] Paul Roetzer: All right. Welcome to The Marketing AI Show. I'm joined today by Britney Muller, founder of Data Sci 101 and former senior SEO scientist with Moz. In this episode, we're going to talk about what machine learning is specific use cases for SEO and tools and resources you can use to understand and apply AI in your career. Welcome Britney.
[00:00:26] Britney Muller: Thanks for having me, Paul,
[00:00:28] Paul Roetzer: I'm really looking forward to this. We're going to start off talking about your origin story, but I think, you know, I, I told you before we start recording, like our audience at the Institute is predominantly beginner to intermediate level in terms of their understanding of AI technology and largely non-technical.
[00:00:46] And so I know you were a PR strategic communications, major university of Minnesota. So I just, I feel like you are our audience, you are this person that came up with this liberal arts degree and has somehow found your way [00:01:00] into being an AI thought leader and SEO scientist. Tell us a little bit about your journey.
[00:01:05] How, how did you get into PR and how did that lead you into SEO and AI?
[00:01:10] Britney Muller: Yeah, it's sort of a weird, a string of events occurred. Really. So I graduated with the PR degree and I started out pre-med and my parents, me too. I failed. I failed pre-med in the first eight weeks. Yeah, my parents almost yanked me out of college.
[00:01:30] That's like, like, but yeah, I go ahead and I, but yes, I was so excited to do like the med route, but my parents, I remember my dad calling you Monday. I just mean like, Britt we just don't see you being in a lab for seven years. Like in a basement, you gotta like, go do fun stuff.
[00:01:51] Paul Roetzer: I just drank for eight weeks and like, didn't really have a choice.
[00:01:54] It was pretty much, my professor said you can ACE the final and you will not pass this class. And [00:02:00] I was like, all right. I, I guess it's time for a career change. So, okay, so pre-med to then you had to go find something else to do. So how did you land in PR and communications?
[00:02:12] Britney Muller: Yeah, so I applied to the school of journalism, which is interesting to me and they had two tracks.
[00:02:19] They had the advertising track and then the PR STRATCOM track. And I just thought the PR stuff was so fascinating to me and something that I could really kind of wrap my head around strategy-wise and have fun with. And I absolutely loved it. But what happened was, so after college, I didn't want an internship.
[00:02:42] I felt confident. I knew how to do these things. And I was applying for these jobs way above my skill level and couldn't get a job anywhere. Like I still actually have the folder. I would like to make wallpaper with all my rejection letters on day, the goal of mine and my [00:03:00] future home. Um, but yeah, could not find a job.
[00:03:03] And it was like starting to just like, get really kind of down and out about it. And, um, I've always loved snowboarding. And I was talking to my parents who are the best and my dad was like, why don't we just. You know, why don't you do that? Why don't we just like pack up your car and drive out to Colorado and find a spot for you to live?
[00:03:24] So my dad and I drove out to Colorado, made friends in Breckinridge at a bar one night, we went out, partied with them and hanging out the one, um, gal needed a roommate and the rest is history.
[00:03:34] Paul Roetzer: Oh my gosh.
[00:03:37] Britney Muller: It's so random. I was putting like a hundred days on mountain having so much fun out there. And waiting tables, which I had done since I was like 15 years old.
[00:03:48] And that gets old very quickly after you have this college degree and you're kind of like, wait, I should be doing something and I can actually connect in with the local realtor on Twitter. And so I was writing these local listings for him. He's amazing. Uh, Pete Deininger and he introduced me to SEO. And the second I tell you literally Paul the second I learned.
[00:04:10] That you can research and identify how many searches a month people do for something. My life was like, it was all I could think about. It would be out on a powder day and it would consume me of what, the number for Brittany Spears. Like what's the number for, you know, how many times are people searching this?
[00:04:30] Like I just, I became this obsessive, um, researcher about SEO. I was creating fake words, you know, quickly I could get them to rank. Um, needless to say, this got me introduced, like obviously HTML, I got into basic script, kitty hacking, which was really hilarious for awhile.
[00:04:49] Paul Roetzer: Script Kitty hacking?
[00:04:51] Britney Muller: Yes. So in the hacker world, it's like, it's very important for me to self identify as not a real hacker.
[00:04:57] I was literally just copy and pasting, like Tor packaging scripts to hack printers and wifi connections and silly stuff like that. But yeah. Um, yeah, the second I learned that you could manipulate search results. I mean, I just went crazy. I spent seven months preparing to rank number one for Burton, us open date of the event.
[00:05:20] And I ranked above burton.com and their marketing department called me. And I'm like, this is how you get a job. I got a show. Then they invited me to dinner and ask me all these great questions. And I told them everything, you know, I didn't know. Like it was really nice meeting you. And so after that, you know, I started an agency.
[00:05:41] We were very successful in the marketing space. Um, got super burnt out from that, got poached by Moz and absolutely loved it there. Um, and this interesting thing happened back around 2014 where I had sort of figured the SEO stuff [00:06:00] out. You know, there wasn't a whole lot changing in the SEO world and I was hungry for that, that kind of momentum and that intrigued again, something I could get really passionate and curious about.
[00:06:11] And in enters machine learning as the sexiest job of the 21st century. And I remember not knowing, you know, anything. I remember thinking, well, what are technical site websites that I know of in which I can look up machine learning? And the only thing I could think of was Git Hub I remember going to Git hub had no clue, what I was doing and typed in machine learning.
[00:06:33] And after a couple of pages of looking at different repos, I found to data science 109 course, and I don't think people were supposed to find it at the time. And they were releasing their classes. Every, I remember it was every Tuesday and Thursday and I absorbed them like a sponge. I lost my mind and I kid you not.
[00:06:55] I did all the homework assignments. I even emailed the TAs without a Harvard email address were questions helped me. So like, I feel so grateful. Like I stumbled across these resources. I got incredible support pretty early on. And just kind of became obsessed with this idea of you're able to feed a machine data and let it recognize patterns that, you know, most of the time were on identifiable to the human eye or to the human brain.
[00:07:26] So, um, yeah, just kind of went. All in after that.
[00:07:31] Paul Roetzer: That's awesome. Yeah. It's so funny. I mean, one pre-med to. I also then tried to get into journalism school and couldn't for a couple of years. So I ended up majoring in PR like I think you and I were like, meant to eventually like connect for this podcast.
[00:07:44] I had no idea that our backgrounds were so similar. And then like you, I kind of went down that discovery path. I, I, you know, eventually found courses through Coursera and Andrew NG's course, which I think you've taken too. Um, yeah. And it's just kind of that self-taught and I, you know, that's kind of what I wanted to dive in with you in this episode is.
[00:08:02] I, you know, I think you're, you're obviously more technical than the average marketer, but generally speaking, you are our audience, like, you know, this PR communications, liberal arts background, find a passion for AI. Go on this path of discovery. Self-teach because you can't learn it at university. Like the stuff at university, especially back in 2014 was all technical.
[00:08:25] Like you don't take a class like, you know, AI for marketing at a university. Even today it's very limited that it's taught outside of, you know, um, the technical side. So you have this passion for machine learning in part, because you went down this path of discovery. Like, why is it been so important to you to share that with the SEO community and the marketing community at a larger degree, but like, what is it you see in machine learning that makes you excited about what's possible now and moving forward?
[00:08:57] Britney Muller: Oh my gosh. There's so many things. I, I think it's one of the more powerful technologies I've ever become aware of. And the thing that I noticed early on is. The bottleneck and AI, it's no longer hardware and it's no longer data. It's people, it's literally people. It's, there's a very small subset of people who are able and willing to do these things.
[00:09:25] And I think that's a huge misfortune for. Different communities, different domain experts, whether it be healthcare or the non-profit space or marketing. I just see so much potential in these models that I get to play around with all day. So one of, one of the things that really started to kind of pique my interest was I would find a model that did something really well.
[00:09:50] So for example, I came years ago I came across the. Shakespeare model where all of Shakespeare and it creates new dramas, new content, new characters, new storylines, all stuff. And I thought, well, what if I fed it something else? Right? Like what if I just swap out the Shakespeare and put in marketing content?
[00:10:10] And it will, at the time I'm putting Beyonce lyrics and ran fish, but with like SEO raps and songs that learned how to rhyme, it was amazing how little it needed to learn that stuff. And it just completely blew my mind when I went into about transformers and how, you know, you're able to do those things, but.
[00:10:30] Again, back to what you're saying for me not being formally trained. Totally self-taught I sort of feel like this monkey see monkey do. I find and wrap my head around current models that are doing really powerful things and I just make them do something else. And a lot of times, you know, I'm on YouTube, literally watching someone and just, I copy, I copy it.
[00:10:53] And I swap out different stuff. I do a lot of things on Google code labs, or they walk you through step-by-step, which is super nice. And it gives you confidence to then do it again on your own differently and experiment and have fun with it. But I think. Um, you know, I hope more people get excited about the space and are willing to just try things.
[00:11:16] That's the most interesting thing too, that I've kind of come across in this field is, you know, I think it's quite easy to get excited about it and it's fun to see the opportunities, but few, few people in my opinion are willing to just press that return key or to just open the terminal and try to download Anaconda, you know, there's this.
[00:11:38] There's this interesting barrier to. Action in the more technical aspects of it, which I would love to break down.
[00:11:46] Paul Roetzer: Yes. It's funny. There was a, I think it was called runway ML. I'm just looking up. Yeah. So LA at 2019 was our inaugural marketing AI conference, MAICON and for my opening talk, it was like how to be a pioneer in AI.
[00:12:00 And so what I did at the end was the exact thing you're talking about. So this runway ML has all these pre-trained models and one of them was to turn any work of art into a classic. Like a, uh, a Picasso Monet, Monet. And so I actually took a drawing of a character. My daughter had created it's called super everything girl was a superhero that had every power and I fed it, the image. And then I turned it into a Picasso Monet. And on stage I showed how literally in three minutes, like I downloaded the capability on my computer, uploaded the image. And here was my daughter's work imagined through the eyes of the console.
[00:12:36] And it was just like, That was mind blowing to me. Like it was so cool as a non coder to be able to take a model. So for, for everybody out there, like what is a model like, let's, let's take a step back for a second and say, what is machine learning? Like, how do you define it? And then what is a model? So that people have an understanding what we mean when we say yeah, just find a model and repurpose it.
[00:12:56] Britney Muller: Yeah, that's a great question. Uh, so to break down machine learning, the best way that I like to describe it is how traditionally. You know, we've had programs for computers, for websites, for you name it, and programming requires literal programs of what to do upon the data to make this thing happen, to make this end-result machine learning, it's switched around.
[00:13:21] You take what the end result you desire. You move that up with the data. You're sort of training it upon the data in the model. The training creates the program. So it sort of swaps that around where we can now do, you know, unsupervised supervised where you're literally feeding, you know, one of the famous examples is the San Francisco housing price model.
[00:13:49] So that's a great example too, because you think about housing and predicting home prices. You have all of these feet, what are called features, which are the size of the home, the number of bedrooms, the location, the square footage, blah, blah, blah. And you reiterate in these, what's known as models where they're literally looking for.
[00:14:14] To assign weight to each feature, but not only that, they're doing a lot of kind of feature engineering in deep learning in and of themselves. So they might come up with new metrics. Like maybe they figure out that, you know, bedrooms divided by bathroom. You know, there, they do all this more complex analysis on the data that then.
[00:14:36] Results in fairly accurate testing and training upon new data. So that's the biggest thing is you want to play around with models and get comfortable enough with the process. Which I feel like I didn't just explain all that well, but it's, it's literally this iterative process where you're training a machine learning model to go over the data again and again and again, in different ways. And then you test that outcome on new training data to see how well it performs. How accurate is that information? Where is that loss what's known as a loss or loss curve? Um, and so that's kind of that's machine learning in sort of a nutshell. And then the model is once you're sort of happy with its performance, that is the end result that you would then use in action on a website in a predictive forecasting model and text generator, you name it, but the important part there, as well as there's this whole new field emerging with ML ops that helps create visibility around the model after it's sort of become this static model because. There's the data drift errors start to occur. Bias starts to pop up. You have to, it's not just a one and done process.t should be very iterative and kind of. Um, monitored continuously.
[00:16:05] Paul Roetzer: It's hard to know why it made a prediction or a recommendation or why, you know, for allowing it to make decisions, why it makes its own decisions. Because yeah, I mean, it's like a famous example of getting into the SEO space is, I mean, Google has been using machine learning for years to, to inform what search results show up.
[00:16:22] And obviously it's made, you know, it's evolved over time. But as a marketer, as an SEO specialist, you didn't always know why it was deciding what it was decided. There was all kinds of factors going into it. And in the end of the day, like it's, it's machine learning models like it, it's trying to learn the best result to show you.
[00:16:40] It's predicting what your intent is and trying to show you what it thinks is the best result for you. So for years, as marketers and SEO, people like. We were being dramatically affected by machine learning. We just didn't know that's what it was or what to call it.
[00:16:57]Britney Muller: Exactly.
[00:16:59] Paul Roetzer: And then they made, was it a Burt?
[00:17:00] I know it was like a, a big change in recent years where there was more context around the search. You were actually like, it was taking into consideration prepositions and actually using natural language understanding to not just look at keywords, but like, what is the actual intent of the question being asked and the search being conducted?
[00:17:18] So again, I feel like. For a lot of people, you know, want to be showing up in search results. We, we needed to understand the underlying technology, which was AI powered. And again, as an industry, we just really didn't know what that meant.
[00:17:34] Britney Muller: Exactly. It's a great way to do it. I love that.
[00:17:38] Paul Roetzer: So what are some other like you've, you know, I've watched some of your talks and like, you know, your whiteboard sessions and things, and I think you do a great job of giving very practical use cases in the SEO realm.
[00:17:50] So when you think about an SEO specialist, and again, like in our audience about 42% of our audience says they're involved in SEO. So we'll ask people like which areas of marketing you're involved in. So a pretty fair amount. They may not be specialists, but they're, they have some responsibility for how the site is ranking, how keywords are showing up.
[00:18:09] Um, what kind of traffic is coming in through organic. So what are some of the key use cases you think about an SEO that people should be using today? Or they could certainly be starting to explore the potential of AI in those areas.
[00:18:24] Britney Muller: Yeah. I see AI and machine learning really. Leveraging SEO on large websites specifically, just because at scale, it's next to impossible to account for every skew and shirt color and you name it.
[00:18:45] And just with a little bit of machine learning, it can be incredibly valuable to automate the writing of title tags. The writing of meta descriptions. Uh, there's also a lot of really interesting work that's being done around, you know, automating three Oh one redirects and automating some of these sort of the fundamentals that we have known and followed for ages so that we can then essentially level up and focus on kind of higher level strategy, higher level thinking as far as where we want to take the SEO efforts.
[00:19:25] Um, I've also recently become pretty obsessed with, uh, statistics and realizing how bad marketers are at like, we are terrible. I'm terrible. It's insane. Uh, it was, it was really humbling to take, um, statistics for data science, this MIT course recently. And one thing that kind of blew my mind was, you know, they don't start you out in the models, they start you out just simply understanding the data. And being able to describe it,] right? So being able to pull large amounts of data, whether it be from Google search console or Google analytics and have a statistical, just a 101 to clean that data to better identify how it's distributed to describe it more clearly and more accurately to a client who wondering why isn't this happening?
[00:20:23] How is this occurring? Um, and data science machine learning really allows you to leverage those insights on your own. So that's something I've also been really excited about is just. Planting the seeds of fundamental data science, things that you can do in notebooks to really power up your SEO and also be, um, more valuable in terms of, you know, on the fly insights, in a meeting, you know, this is fully capable to everyone out there with just a little bit of practice.
[00:20:57] So those are some things I think are. More immediately, um, kind of relatable. And then you go into, you know, the recommendation models. I think, you know, it was, uh, Lex Friedman and a lot of people who kind of think that. The recommendation models, the product recommendations, the different things on websites that recommend, you know, what to read next.
[00:21:21] All those things will likely be the more used at scale models for the next, you know, 10, 20 years. Um, and to put effort and focus into getting that right, I think is really.
[00:21:37] Paul Roetzer: Really strategic and amazingly, it's not like, you know, when you think about recommenders, Netflix, Spotify, like they're all based on that concept.
[00:21:45] And, you know, it's that it's predicting what you would want to consume next is basically what they're all doing. And yet, if you think about. A lot of the corporate sites, you go to, even a lot of the media sites, there's a really good chance they're using basic tagging to make those predictions that someone is writing an algorithm.
[00:22:03] It's not the machine learning is a human is actually saying if they looked at this or if they consume that, then they'll likely want to have this. And the site is actually recommending content that a human wrote. Uh, uh, uh, a decision tree for basically. And so the idea with machine learning is we suck at that as humans.
[00:22:21] Like it's impossible. If you have 10,000 or a hundred thousand visitors to a site, how, how is my little human brains supposed to know what those a hundred thousand people will actually want to do next it's it's incapable of it. And so I think that's, to your point, it's like, that's where machine learning is power lies is like making predictions about.
[00:22:40] Outcomes are about, you know, what happens next and then trying to recommend what to do as a result of it.
[00:22:47] Britney Muller: Yeah, exactly. Exactly. It's also interesting to think of like, um, The compute struggle side of things, you know, it's, it's essentially next to impossible to have customized models for every individual.
[00:23:01] So exactly. To your point, a lot of tagging is used. I know one of my favorite examples is a lot of people think that Spotify uses machine learning for, um, what to listen to next or your discover weekly. And what they actually do is what I heard at an AI conference from a Spotify engineer two years ago.
[00:23:20] Is that they take the song that you listen to and they. Look at the people who listen to those songs and look at what they commonly listened to outside of that. So what's the next most con brilliant, brilliant. Because now, you know, they, he was explaining, it would be super slow and next to impossible to run.
[00:23:41] An AI model for every individual. Instead we use this, you know, really interesting system that seems to perform extremely well. And so I also, like, I really learned to appreciate sort of the hacks like that, that don't necessarily require machine learning. That's something that I have been humbled with the last several years where I just, you know, everything is machine learning problem.
[00:24:04] Once you have this hammer and then you get into it and you're like, Oh, that's not the best. Maybe we could do this, but just even thinking like that, I think opens up a whole new world of tools and options and ideas and. That's what's so much fun.
[00:24:22] Paul Roetzer: There are so many applications. I mean, I love the couple of, you mentioned like tagging images, so you can use computer vision and train the model to know what's in an image.
[00:24:31] So if you have like, you know, apparel, for example, and you have a thousand skews, Yeah. The thought of having to write a product description and tag, those is mind numbing to the average person. I mean, maybe there's people who enjoy going through a spreadsheet and writing a thousand descriptions, but that is not something a human needs to do anymore.
[00:24:48] Like that the technology is there to do it. Um, made a descriptions, like you said, now you're going to want to do your primary pages. Like you still want human attention to the key things, but to your point, if you have 50,000 pages on the site, You don't need a human to write 50,000 page titles and descriptions.
[00:25:07] Like it's just not a good use of human.
[00:25:11] Britney Muller: Exactly. Exactly. It's funny. I was sent a demo yesterday, uh, by someone who just playing around with this conversion AI tool and they create content for a webpage based off very little information. And there was hundreds of tone of voice options. It was unbelievable.
[00:25:33] Paul Roetzer: Was that conversion.ai? I was literally looking at that yesterday cause they're using GPT three too, which yeah. copy.ai is another one that just raised like two and a half million, I think. Um, I believe they're both Y Combinator companies, but. Yes, GPT3. Maybe we should talk about GPT three for a minute.
[00:25:52] Cause I know, um, you've talked about GPT two in some of the talks you gave, but have you messed around with GPT3 outside of like exploring a conversion.ai? What they're doing with it? Is that something you've played with?
[00:26:04] Britney Muller: I'm worried I've been blacklisted somehow or something because I signed up for it way long ago. Like I was at the start of it and I haven't gotten access to it. I've played with it in very small amounts, but. Just through like hugging face when they had it and now no longer have it. Um, yeah, but have you?
[00:26:24]Paul Roetzer: I also applied to get access to it. And so we're, um, you know, it's, it's interesting. So, um, Conversion.ai is doing it copied out.
[00:26:35] AI is doing it. There are lots of these companies that have sprung up that are enabling a bunch of really interesting marketing use cases on the top of GPT3. Then you have Microsoft signed a licensing deal with open AI. I think it's like a billion dollars or something. So Microsoft actually, I believe controls the licensing rights to GPT three moving forward.
[00:26:55] Um, and so you're going to see it, I think, integrated into Microsoft. Products. Um, but again, like I'm, I'm not sure in the space is so quickly evolving because GPT too, I think you alluded this in one of your texts. Like they wouldn't even release the source code initially because they were so worried and yet GPT3 is like a thousand times more powerful than the previous version.
[00:27:19] And it's like, here you go. Like, let's just. Start writing copy and creating social posts and writing long form content. And yeah, it's, it's such a, the language generation space is so fast moving and it has so many implications on SEO because you have the potential in theory, to just give GPT3, a topic, a person's name, a theme, whatever.
[00:27:42] And it'll in again, in theory, write a thousand word article about it. And if, if a machine can just start writing whatever content you want. Wouldn't that just completely transform content marketing and SEO, like, do we, do we need SEO people in the future? Like, but the thing catches, it's not really what it appears to be.
[00:28:03] Like one uses a massive amount of energy. You would alluded to the required energy to personalize Spotify recommendations to every individual. GPT three uses a massive amount. Of energy and processing power to do what it does and it degregates over time. So the first 200 words may sound awesome. But then as you keep going, it, it loses its capability to continue to be human-like and sound.
[00:28:31] So it's such an interesting time in the industry because there's all these awesome models like you've been playing around with that. Do all these really interesting things to know pretty well. And then there's these transformational things that may change everything, or maybe not. Like I don't, I don't know the answers.
[00:28:48] Like I tell, I love talking to people. It's just like, what do you think? Like, where are you, where are you at with this technology?
[00:28:54] Britney Muller: Yeah, the thing that I have slowly been like wrapping my head around and it's taken me a while to get here is the idea of it being so dangerous. Like you said, and that's why they no longer were doing the whole open AI, which, you know, that was their mission was to make it open source and that they were going to, you know, build this and create this open community.
[00:29:17] And that never happened. But it's been interesting because at first, you know, similar to. Exactly what you were just saying. I got so excited about the STL implications and how it could sort of create these frameworks for marketers and for websites that you could then go in and kind of calm out and make it your own.
[00:29:39] Um, but where I do see the dangers after doing a little bit more digging is, you know, look at the political space, look at the conspiracy theories. Look at all of. The, you know, made up fake news sites, et cetera. Imagine if this got into the wrong hands and imagine if they wielded it in a very strategic way.
[00:30:00] I do think that that is frightening. And I mean, so it was sort of deep fakes, so all of these things, um, and they're becoming harder and harder to detect. Right. So I, I'm curious to, to hear your thoughts on that, like, do you think it's dangerous or how do you feel about it being more openly available to people
[00:30:20]Paul Roetzer: that is terrifying
[00:30:22] So I, you know, it's, um, So I, I did a lot in like 2016 to 2018, 19. I did a lot of private talks, like to groups of like 10, 15, really influential people at major organizations. And it was just like this introduction to AI and like opening minds of what's possible. But then. What you end up doing when you talk to really smart people who connect the dots very quickly, they start asking the questions.
[00:30:52] You just started bringing up like, well, wait a second. What does this all mean to politics, to society, to the economy. And then you [00:31:00] start going down a whole different path. And it's like, well, okay. Like, if you want to understand Surrey, which we all use daily, you have to understand its origins with DARPA and why Siri was created as a military.
[00:31:12] You know, function, and then you start to look at a lot of the innovations that exist and you realize the tech has actually been there for a decade or more, but in many cases it was innovated. In, in, in government that it was created for some other means than what it's now being commercialized for. And so that's one aspect is that many times they have origins outside of the obvious commercial use that we look at today.
[00:31:37] The other is to your point about politics, you can go look at the story of Cambridge Analytica. And the great hack on Netflix is one of the great documentaries, in my opinion, like what happened in 2015 and 16 and how they weaponize data to influence people's behavior. It was basically a psychological warfare and that was child's play compared to what's possible today.
[00:31:59] And to your point about disinformation campaigns, again, That's not new, no matter what side of the politics around we've been doing disinformation since politics was created, but the ability to do disinformation at scale and not have humans know if it was created by a human or a machine that's new. And that's like in the last few years and GPT3, definitely advanced what was possible there now, again, you can't just take the machine and say, okay, GPT3, just write me a thousand articles about whatever the topic is. Um, and they just go to print. You still need human oversight, these crazy cases where like the media, over-hyped what GPT three is doing, where they it's like, Oh, GPT3 wrote this article.
[00:32:44] It's like no GPT wrote a really shitty draft that your editors rewrote for like 10 hours. And so yes, it was used, but no, it did not do what you're claiming it did and the world isn't ending, but it's so hard. It's getting harder and harder to, to know that. And that's why like Facebook and Google and all these players are spending so much money building AI to detect AI.
[00:33:11] It's like, is this the deep Baker? Isn't it like, is this face I'm seeing. A made up face, like, you know, there's, I won't use names of apps, but like you can go in and actually create. Faces virtual faces and it's not anybody. It's not a photo of someone. It looks like a photo though. And you can do the deep fakes.
[00:33:30] Um, there was a cool one I saw talking about. Good. You just have it? I think it's, um, I don't know how to say the name sin easier. It's S Y N T H E S I A they just did a big thing with PepsiCo and lays potato chips where you can actually get a personalized message sent from Lionel Messi. And what they've done is pre-trained models that you just type in the text and he reads it and in multiple languages.
[00:33:57] And so I could send you basically a personalized message from Lionel, and you would have no idea it wasn't from him because it's his face trained as a deep, fake on any text. And so whatever I put in, it'll say, Hey, Brittany, it's fine. I'm like, hope you're enjoying your baby bag of Lay's potato chips. So that's fascinated, but like, To your point, there's always a dark side.
[00:34:22] So it's like cool commercial application. But if you know what it's capable of doing and how it works, which is where you've gotten to, it's like, you look at things, you're like, okay, I know they did that. And that scares me now because now I understand how it can be applied to other things. So, yeah, it's a long way of saying yes, I.
[00:34:40] I worry a lot about it. And it's part of why I went down this path a decade ago is I saw a bit of what I thought the future would look like. And I had little kids and I became very concerned for what their future would be. Um, and so I have part of what's driven me to understand AI is to understand the greater impact on society.
[00:35:01] Marketing is just my, what I happen to do for a living. So it was my way to figure it out.
[00:35:07] Britney Muller: Yeah, that's amazing. Have you read the controversial, uh, paper that Google ended up firing those two researchers?
[00:35:17] Paul Roetzer: I followed the story, but I have not read it, but I, the gist of what? So for people that aren't aware of it, what was, what happened? What was the topic?
[00:35:25] Britney Muller: Yeah. Okay. So it was, first of all, it's the most well-written brilliant paper I've ever come across today, right? It's on large language models and it talks about how, why larger isn't better and how now, you know, you look at Google and Facebook and Amazon, and they are making so much money off of these large models that I think, I think there's this position at these companies where they don't feel they can go back.
[00:35:55] Right. It's working, it's powerful. It's all of these things, but, um, Tim knit. Uh Gibro I don't know if I'm saying her name correctly, but she helped author, um, this paper called on the dangers of stochastic parrots can language models be too big. And they're artists a couple of things in here that are on believable that I've never thought about that are so smart.
[00:36:20] And one of the points is really social movements. And how, when you think of all of this text is being grabbed from some like a centuries worth. Of writing and literature and things have changed and there's no, um, really reconciling that when you've already fed a model, all of this data that is kind of no longer with the times it's going to lead to racist, uh, outcomes, it's going to lead to bias.
[00:36:52] It's going to lead to all of these things. Um, And yeah, there's just so many powerful takeaways. I mean, the effect on climates, uh there's I could go on and on, like, here's just a quote, feeding AI systems on the world's beauty, ugliness and cruelty, but expecting it to reflect only the beauty is a fantasy.
[00:37:17] Like there's, there's just beautiful things. Yeah. That's throughout that makes so much sense. This is such like a sound piece of, um, you know, work. I don't know. I, things like this definitely start to, you know, trigger a bunch of red flags and thinking about it and then experiencing it. Right. Like I play with these models all the time.
[00:37:38] I'm building Dom models all the time. Uh, just earlier this week, I. Um, use this program to delete me from a live camera view.
[00:37:47]Paul Roetzer: I saw that was on Twitter. I think you put that. He put it somewhere, but yeah. Yeah,
[00:37:50] Britney Muller: it was so much fun. But similar to that, you know, I was playing with PI Fu. Uh, this last summer, and I was speeding at images of my brother's wedding because this pie Fu model, it requires one like high definition image of a full body shot of a person.
[00:38:07] And it will create a 3d, like kind of like action figure of the person and it spins them around. It's pretty cool. And so I feed it, my brother. Um, I feed it, you know, Kate who's clearly wearing this big wedding dress and then I feed it a picture of me where I'm reading and I'm wearing a more fitted dress and it cuts my butt
[00:38:32] off. And puts in wallet and cell phone in my back pocket, because it's been trained on so many images of men and I. Believe it, I could not believe it. And I see stuff like that all the time. I mean, it's funny. We talk about this bias as if it's some, you know, debate it, but just look at like the T SNI models that Google has publicly available and type in engineer and look at the closest nodes.
[00:39:00] It's Michael man, James, all of the it's male dominated and. There's just, there's so many great cases and this paper touches on that too, you know, scale back and be more thoughtful about what it is that we're putting into these models, because at some point it does become increasingly difficult to go back.
[00:39:24] Right. And sort of fix things. Um, you know, their argument is it's next to impossible.
[00:39:30] Paul Roetzer: I said, that's what I love. Like the way you've approached learning. This is because you're not trying to be like an expert in machine learning or an expert in bias. What you're doing is absorbing the full story, like looking at all the different elements, taken a class in statistics, taking a class on machine learning, reading a paper on bias, like.
[00:39:51] That to me is what the future marketer looks like. It is this understanding of the bigger picture technology and what enables good and bad. And then what are the fundamental things you need to understand about it, to make informed decisions as a consultant and advising clients as a marketer within your own organization.
[00:40:10] Um, as someone who may affect the building of different technologies. Because to your point, like you can't just raise forward and build stuff because it's technically possible. And I think that's always been a fundamental flaw of Silicon Valley. Um, and, and at a bigger picture of engineering in, in many cases, it's like, well, we can build the bomb.
[00:40:29] Let's build the bomb. Like, unless, unless an ethicist shows up and says, we can't build the bomb, we're going to build the bomb. And I think in some ways, It's kind of where we are with AI and some of the bigger players, like Microsoft have been very front and center about what they're doing to try and build this in the right way to build AI for good.
[00:40:52] Um, and my big thing is we need more people talking about it. Like we have to have an understanding that bias exists, but how do you identify it as a market? Or how do you even know to ask the right questions? So you may find this amazing tech and it's like, Oh, this is going to save me a hundred dollars a month.
[00:41:09] If I buy this AI tech to do this thing for me in marketing. But if you didn't ask them, well, where'd the data come from? Or like, how, how is, how does the model learn? Like. If you don't even know to ask that question, then, then we can go down a slippery slope as an industry. And that's what worries me is uneducated people buying powerful technology and not understanding the ramifications of how they use it.
[00:41:33] Britney Muller: I think you're exactly right. Exactly. Right.
[00:41:38] Paul Roetzer: Well, I'm glad we got a chance to, and that was, I was actually going to ask you like things that were even excited you, but I mean, I think we, we kind of covered it. I mean, what I want to ask you, like this kind of my last question, and then we'll get into like the rep fair to end it, but.
[00:41:52] So having gone down the path, you've gone down the last decade and gotten where you are with this. You know, it was a really strong baseline understanding of a lot in AI. Like what is your advice to a marketer who's listening. And maybe for the first time thinking I really should pick up a book or I should take a course.
[00:42:08] So like, what is your advice to someone who wants to be better prepared for the future?
[00:42:15] Britney Muller: Yeah, that's a really good question. Um, the biggest piece of advice I can get is I do think this is sort of an industry that you can get a bit lost in. There is so much the more you dig and the more you learn, the more you realize how little you actually know.
[00:42:34] And this is something that I really struggle with. Um, and so I think being comfortable, not, not true expert and to your point, knowing enough to be dangerous, knowing enough to ask the right questions and also just knowing what's possible, right. Being able to surface solutions or ideas for applications of.
[00:43:00] Of models that have already been created to do something to a situation or project that you're working on. I think that in and of itself is unbelievably powerful. Uh, and then in terms of, you know, doing things on your own, just to not get too frustrated by it, You know, there's there, it's becoming more and more accessible, you know, back in 2014.
[00:43:23] I remember my first regression, when your regression model on TensorFlow, the original was over 175 lines of code. So proud of it, you know, I was like, this is, and now you can do it in six or seven. It's wild. How, how much more accessible these programs are becoming an end. You know, there's all sorts of options for individuals too.
[00:43:49] Test things out in an environment like CoLab notebooks, where you don't even have to, you know, download programs on your computer and, you know, wrestle through some of [00:44:00] those hiccups. You, these platforms are available at little to no cost already that are doing really powerful things. And so just to kind of encourage individuals out there that are curious, just to play around, that's all, that's all I would ask.
[00:44:15] And, you know, if my get hub machine learning for SEO. Repo is helpful. I would be honored to help facilitate and spark some of that interest because I know it can be really hard to kind of find notebooks or find applications. And so what I've basically done with that Rebo. I've written a couple of myself, but a lot of them I've just discovered and I've made it do things for SEOs.
[00:44:42] And so that's, um, kind of an easy way to kind of dip your toes in.
[00:44:47] Paul Roetzer: That's awesome. And you don't have to hack for Harvard classes anymore because Coursera exists on like probably when you started. So now it's just, it's all free. Like you could take a classroom, Harvard or Stanford or MIT or wherever you want to take a class from.
[00:45:00] [00:45:00] Britney Muller: It's a dream. Yeah. That's the other thing, those things, and to not be afraid to ask questions, you know?
[00:45:08] Paul Roetzer: All right. So we're going to wrap up the episode with rapid fire questions for Britney like we always do first, a quick word from one of our sponsors. HubSpots alright. Wrap it up here with a few questions.
[00:45:21] Um, think about SEO. What percentage of SEO tasks do you think will be intelligently automated to some degree in the next five years? So what will you know, what percentage of them will machines at least do part of the work that traditionally was done all by humans?
[00:45:39] Britney Muller: That's a tricky question because I actually forecast and predict that it will happen on the other end. I think that things will become no longer necessary because Google's machine learning will have gotten so good at better understanding things, the way that they are trusted. And what I mean by that, as you know, I don't know that the link graph will forever be totally necessary.
[00:46:03] I don't know that, um, You know, meta descriptions and we already see those kind of becoming extinct as Google grabs texts off of the page. Um, I think we're, we as an industry have to evolve with Google and I think in doing so in the next five years, I would predict anywhere from 30 I'll try and stay somewhat conservative.
[00:46:27] 30 to 40% of things will either no longer be necessary or have shifted to something else. Well, I, I would agree with you.
[00:46:37] Paul Roetzer: I mean, I think in SEO in particular, which actually leads to the next question and it might be your answer. So which marketing category will experience the greatest disruption from intelligent automation in the next five years, advertising communications PR, since we both have a background there, content marketing, email marketing, or SEO or other,
[00:46:56] Britney Muller: that's a good question.
[00:46:58] Um, And that's tough. I feel so biased. Cause I just there's so much. So I would say,
[00:47:05] Paul Roetzer: well, you just made a pretty good case. I mean, you're basically opposite leading 30% of not, you know, it's not just full automation, it's just get rid of it because you don't need to do it anymore. That's so hard to make that case for these other areas.
[00:47:16] I advertising maybe. I mean, I, I could see an argument that advertising is going to be really in trouble. Um, okay. Voice assistant use most Alexa, Google assistant, Siri Cortana. Don't use them.
[00:47:28] Britney Muller: Google. I love my Google home, but I do, you know, I cover the camera and I'm weird about it. And sometimes I'll just unplug it, but I do enjoy it.
[00:47:37] Paul Roetzer: Yeah. All right. Uh, more valuable in 10 years, liberal arts degree or computer science degree. Ooh.
[00:47:44] Britney Muller: Oh, that's so tricky. I would actually say liberal arts.
[00:47:48] Okay. I'm in the liberal arts category as well, also because like what a computer science is going to need to do in 10 years, if the machine does itOkay.
[00:47:57] Paul Roetzer: And then last one net effect over the next decade, [00:48:00] more jobs eliminated by AI, more jobs created by AI, or it won't have a meaningful impact.
[00:48:09] Britney Muller: I think there will be a shift. I do see it taking over jobs. That's just my honest opinion. Um, but I also see, you know, things changing similar to, you know, you think of when we, the auto industry happens.
[00:48:24] Paul Roetzer: Yeah.
[00:48:25] Britney Muller: So I think it will be redistributed somehow.
[00:48:30] Paul Roetzer: All right. And let our audience know the best place to find you.
[00:48:35] Britney Muller: So I'm probably most active on Twitter and it's just Brittany Muller.
[00:48:42] Paul Roetzer: Awesome. And we'll put that in the show notes. So I really appreciate your time. This has been awesome. I have many follow-up questions I'm going to hit you up with afterwards, um, but really appreciate your time and your insights for our audience.
[00:48:54] So thanks for being a part of this.
[00:48:56] Britney Muller: Yeah, this was so fun. Thanks for having me on.
[00:48:59] Paul Roetzer: alright this has been The Marketing AI Show. Thanks for joining us until next time. We'll see you soon.
Sandie Young was formerly the Director of Marketing at Ready North. She started at the agency during the summer of 2012, with experience in magazine journalism and a passion for content marketing. Sandie is a graduate of Ohio University, with a Bachelor of Science from the E.W. Scripps School of Journalism.