Accidentally building SOTA AI

0:06Welcome to Practical AI. If you work in artificial intelligence, aspire to, or are curious how AI-related technologies are changing the world, this is the show for you. Thank you to our partners at Fastly for shipping all of our pods super fast to wherever you listen. Check them out at Fastly.com. And to our friends at Fly, deploy your app servers and database close to your users. No ops required. Learn more at fly.io.

0:42Welcome to another episode of Practical AI. This is Daniel Whitenack. I'm a data scientist and founder at PredictionGuard. And I'm joined as always by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How are you doing, Chris? I'm doing fine. It's another exciting day in the AI world in 2023. Probably the most exciting year in AI ever. Yes. And our episode today is very timely. It's not Lately, although we are going to talk about Lately. Oh, God. You actually said that. That's the best one yet, honestly. I've heard a lot of them. Okay, good. Well, we have with us Kate Bradley-Chernis, who's CEO at Lately.ai.

1:30How are you doing, Kate? I am now tickled, so well done. I'm impressed with you. We can be friends. Okay, great. I'm very happy about that because, yeah, you've done some amazing things. You're doing some amazing things lately and can't wait to hear more about them. Before we jump into more specifics about what you're working in specifically, I'm curious because you have had experience for a number of years working in this sort of generative AI space and especially on social media and content generation. I'm wondering what your feeling is about kind of this general moment that we're in with AI and like how it's shaping.

2:12Has it reshaped some of your thinking around AI? Has it proven things true that you've always thought about AI or what's on your mind right now? Yeah, I mean, it's so funny because we actually didn't even know that we had built AI back when we had started building it. A mentor had to kind of give us the clue. And then they got us a grant with IBM Watson. This is in 2018. So, you know, my perspective on it has certainly changed over the years in so many ways. I mean, it was so hard for us to explain what we did for so long. And suddenly, everybody thinks they know what we do now. So they're coming to us with this like whole different perspective that we still have to educate.

2:51But what's exciting to me is there's been there's three waves we've seen already happen. And there's another one that I believe is about to happen. The first one is like, holy shit, AI, everybody, right? Wow. You know, this is amazing. Everyone just had a freak out. And then the second wave has been all around the legalities, a pullback, you know, copyright issues. You know, what are those kind of court liabilities that are going to be up and coming? And then the third wave is around the voicings, right? Like, OK, well, now that everybody has cliff notes, essentially, like, how do you make it your own?

3:24And then the fourth wave is, and it's already here, really, we're seeing employees, like employee job descriptions, the need for prompt experience and expertise. So, and lately, by the way, has been ahead of the curve on all of these, like since the beginning, we were nine years old at this point. So it's funny and weird to be in the position of spending years trying to communicate to people what we were doing, why it mattered, what the value was, and then suddenly be riding at the top of this wave where we've already built the future. Right. And now everybody's just cottoning onto it. And you mentioned like having so I'm reminded of Primer.

4:05I don't know if you've ever seen that movie where they like build a time machine in their garage. I don't think they're like totally intending to do what they're intending to do. So you build AI without realizing what you were doing. So what was your original intention or like where was your motivations when you stepped into doing what eventually turned into what is lately.ai? Well, some of it is very boring, but I'll start with the exciting part, which is that so I used to be a rock and roll DJ. Nice. My last gig was broadcasting to 20 million listeners a day for XM Satellite Radio. And my uber power is turning listeners into fans or customers into evangelists.

4:49Right. And I would leave stations and then I'll never forget my program director calling me six months later and saying, hey, the Arbitron book came out and you were number one. How did that happen? Because it was shocking. I was in a format called AAA. We're always like 20 or 21. rock and pop are number one. And evenings, you know, it was like totally unheard of. And I said, I threw your playlist out the window, which was the truth. But I also was the production director. So I was in charge of all the sound in between the songs. And so it was like the me show for, you know, four hours a night.

5:25And I started thinking about, I had written thousands of commercials. I was a fiction writing major. I really was excited about the theater of the mind. I'm going to vomit on you guys here a little bit. This is the place to do it. Bomb it away. Okay. You're doing a great job going with me. So the theater of the mind is when your imagination has to play a role in the act of the storytelling, right? So when you are reading, it happens. When you are listening, it happens. When you're watching, it doesn't happen because all the pieces are there in front of you. So there's nothing for you to fill in the blanks, right?

5:55So that parallel was really interesting to me. And one of the things I read that book, you know, This Is Your Brain on Music, where it dissects the neuroscience of music listening. So when your brain, Chris, listens to a new song, it must instantly access every other song you've ever heard before in that moment. All right. And so what it's trying to do is to find familiar touch points so it knows where to index that new song in the library of the memory of your brain. And it's tugging on nostalgia and obviously memory and emotion, all the things that build trust. Trust is why we buy. Now, similarly, Daniel, if you're writing me an email or a Slack message or a text message, I'm going to hear your voice in my head.

6:34And your voice has that same idea, right? Like there's this sound component to it. So if you're doing a great job, you're going to try to figure out how to write in a way that's tugging on nostalgia and memory and emotion and trust. So I took these ideas out of radio. And there's another story for when we're having a beer in the middle here. but suddenly I had a marketing agency and my first client was a little company, you know, called Walmart. This is the boring part. So I built Walmart, one hell of a spreadsheet system that took these ideas and translated them into writing. And I got them 130 % ROI year over year for three years.

7:11When we built Lately, Lately was designed to replicate the spreadsheet system I had built for Walmart. And at that time, the industry where we're in was called Marketing Resource management. How exciting. Yeah. I mean, really like the name of our company was Cloud MRM. And so we were building an organizational system for marketing is what it was, right? That's like about as boring as you can get. I love that. I spent a decade in the marketing industry. I'm so with you on this. I loved it. Oh, really? So, you know, spreadsheet hell, you were there. A waving hi. So almost wrapping up this very long story.

7:47So we had built marketing resource management, we had built a feature for every spreadsheet that I had built Walmart. And there's this one feature that everybody was using more than the rest. And the idea was you pasted in a URL of a blog, you clicked a button, and we would instantly atomize it into dozens of social posts, right? That was the foray of the AI we had built. Now we do much more, which we can talk about later, but that's how we got at least to the beginning of the nine-year trip. It's super interesting that you say that because I think maybe people kind of think that like they can easily reuse content on all of these different platforms.

8:27And over time, as I've found out, mostly through failing, if I'm if I'm honest, like the same content that you produce for like for your blog, like it's not a trivial task to just take that and like create a really compelling LinkedIn posts or a really compelling tweet or like what it just doesn't work the same way. And also, like, of course, the algorithms are all changing all the time and that sort of thing. So is that part of part of the education or do you feel that maybe marketers probably feel this pain most and that's who you're talking to? So like your audience, are they mostly wondering, like, how does a computer do this for us?

9:08Because they know the problem and they see it as a really tough problem? Or are they mostly like, oh, you know, I could do this if I really sat down and wanted to do it. But you know, why? So why do I need a computer? Which direction kind of do you find people in most of the time? Yeah, there's so many good thoughts in there. So the first one is like, it's like math, right? There's a reason you go to school and you learn algebra. Because when you're using a calculator, you need to be able to know if you've pressed the wrong button, because sometimes you do and you get the wrong answers, you need to have the background in your mind so that you can use the technology to do the hard work for you, right?

9:45So similar with us, we found that, you know, if you're a complete idiot, I can't change that. Right? You have to have some... Unfortunately, that AI does not exist yet. Right, right. True AI does not exist yet. Magic doesn't exist. I hate to burst a bubble there because, you know, I'm a huge Harry Potter fan. In fact, I saw a car at the Target the other day and his, I assume it was a man, but the license was Avada Kedavra. It's like literally telling everybody to go F themselves like all day, all the time, this guy. Wow. I know, it was pretty bold. So with Lately, let me explain what we do and then answer your question.

10:24So Lately is able to learn your unique voice, Daniel or Chris, or the unique voice of your brand, and you can customize that voice by region, say, for example, or any kind of subset, it also is able to tell you exactly what words, ideas, phrases, even the sentence structures that make up the highest possible performing social media messages for you, specifically customized for your target audience on any specific channel. Now, we're able to build that model in about a couple of seconds for you. And then once you have the model, you have to train it and you train it with long-form content. So we just mentioned a URL of a blog.

11:05So it could be any kind of text, like a newsletter or anything from a Word document. It could also be any kind of audio, like a podcast. It can also be any kind of video. So an interview with a CEO, it could be a webinar, a Zoom call, whatever you want. In the case of audio and video, we will also give you dozens of audio sound bites and miniature video clips that go with the text version of the social post. And then we teach you, and you can do this automatically, how to drip feed that content over time because the long tail has payoffs that are exponential as opposed to the one-off. So we're also working with marketers to educate them, as you were asking about, on post-mo versus promo, right?

11:51So promo is great, but it's really hard to get butts in seats live, But the after the fact is much easier and you get exponential eyeballs. You know, it becomes pretty much everything is evergreen content nowadays. And if you think about that, which I'm sure you guys already have, like before you went into this interview, yes, it's timely, but generative AI is going to be around for a long time here. Right. And these ideas, once they're living online, will the SEO will be, you know, over the roof, over the moon, so to speak. So the who is your question. At first, we thought our target person was me, small agencies.

12:25Like, you know, I was charging Walmart 140 grand to do four months of work. And the idea was, let's give you 140 bucks in your one person. You know, so we were trying to service SMBs. But the mistake we made was we had built this massive, robust platform that was very much an enterprise platform. You know, we didn't know that yet. And again, we met another kind of mentor. we were working with SAP and they were like, hey, we think you don't understand the power of what you built here. Let us help you out. And I'll pause the story there. There's more, but does that answer your question? Yes. It was a good story.

13:01I was all in it. You like cut it off right there in the middle. I was like, what are you doing? Well, we'll get the rest. I mean, like the, you know, it's funny because trends change, markets change, obviously. And you want to be, you don't want to be, you know, loosey goosey, but you also want to be flexible and you want to be able to foresee those things and then turn on the dime when you need to. And with us, because we had built marketing resource management, that window of sexiness was pretty small. Actually, there was a company called Percolate that was dominating the enterprise space. And here we are, this little company trying to be the SMB version of something that no SMB could really understand, honestly.

13:36And we were trying to sell to marketers, specifically like CMOs. And we discovered after a lot of pain, actually, that they were very much threatened by being replaced, you know, which is kind of reasonable. And so we had to learn to reposition the product. We learned to automate it and create self-service. We learned to pitch to CROs. We built in a feature lately where you can actually syndicate the content. One button can push out months worth of social posts across every employee's channel that they have connected that you can imagine. Right. So you can literally do all the influencer marketing for your entire business in like an hour.

14:16So you mentioned voice quite a few times, which is definitely a loaded term within like the AI space. So there's one part of voice, which is like we've had Josh Meyer from Koki on the show talking about like voice cloning. And like, you know, style transfer of voices or using your voice and in another language, that sort of thing. So certainly there's that element of that is relevant to content. Right. But I if I'm understanding right from a very non-marketing person's perspective, what you're referring to as voice is somewhat inclusive of that, but is much more. Could you kind of help those maybe without as much of a marketing background understand what what do you mean when you say voice?

15:01And what do you mean when like an AI system learns your voice? What is learnt, I guess, in that process? I like it. So because we're able to surface you literally a word cloud of the words that make up the messaging that gets you the highest engagement, the most clicks and likes and comments and shares, we can see how you write and we can see how when you're writing well, what's the DNA that makes up those messages? and we can learn how you write for your brand or for yourself. Like even when I write social posts on LinkedIn, I get 86 ,000 views because I'm really good at writing and I use a very specific quote voice style of writing.

15:46For example, in real life, I swear like a sailor, I'm just so foul. And online, I try to do that less. So I'll make up hyperbole like holy hot pickled jalapeno peppers, for example. Now, when Lately is trying to write in my voice, it will insert things like that because it knows me well enough to do that. So part of me kind of coming into this conversation, like I see so many of these people that are making posts that are very formulaic, right? Like, I don't know how many times a day on Twitter, I see some tweet that says, what a week in AI. Here's seven things that you shouldn't miss. Tweet thread.

16:29And like, oh my gosh, how do I like, I should be able to get through these, but they're not using lately.ai. I can tell you. Right. They are not. Yeah. Could you speak a little bit towards like generative content and how that could be, or maybe is in certain cases formulaic and how this sort of approach can go beyond that to maybe be more creative, certainly more personalized. I think you're talking kind of about this personalized voice. I assume we're not just talking about like generate the listicle for or the tweet thread for Twitter type of thing, right? Yeah. So what's interesting about your point is, I mean, this is why humans are always relevant, because you quickly can see that there is a formula there and it begins to wash over you.

17:21It's not tantalizing anymore. And the instinct that you have is not the instinct of the people who are still publishing those. They're still sticking with the old thing because they haven't taken a moment to research really, but even just use their instincts of like what's working, you know, you have to experiment. You have to. It's the same way in radio. Like, I mean, I spoke to a black room of no one for 12 years. I know how you feel. Yeah, you know how I feel, right? Right. And what they advise you, what my mentors advised me back then was you imagine the listener in your mind, whoever it is, right?

17:55Imagine who the audience is. Now, online, I generally just try to entertain myself, you know. And so let me back up. So that was sort of one thing is like humans are relevant and we'll get into that a little more why. And they will remain relevant always in sales and marketing that will never go away. here I'll tell you why real quick when we were talking about the theater of the mind what happens there if you are a good author or you're good on the microphone you're actually letting someone in so I'm going to make you Daniel feel as though you're part of this story now I've written it this way so that you have to fill in I know you have to fill in the blank I have to allow for a third character that is you that's your imagination in your brain that's why when you read a really great book it's so powerful and then you see the movie you're so pissed off right?

18:43They're taking that ownership away from you. So it's the difference between a one-way street and a two-way street. And this is how you build fans. That mysterious thing is the human element in marketing. Marketing is often unexplainable. It is. Like people are always trying to science it to death and you cannot. This is what makes it magical, you know, which is why human training is something we actually worked into our algorithm from the beginning. That was a fascinating point there is that kind of that unknowable aspect a little bit of what tickles a psyche and that evolves and people get used to the thing that was the past thing a moment ago.

19:18and whatever you're doing now will change in another moment, you know, as we accommodate that in our brains and start ignoring it. And so, you know, you've spent these years working on your algorithms to try to hone in on that and, you know, using your method and try to get to voice of what the messaging is in the different mediums. So now in the last year, all this technology change has been thrown at you and your company, generative AI and large language models and all this stuff, all these cool tools. How has that changed how you're approaching the problem of kind of finding that magical nugget of voice in there?

19:58Because you have a whole new set of tools available to you now that you're presumably integrating in to all this stuff. That has to have turned your world upside down a little bit just because of the vast amount of capability you now have in addition to what you had before. A couple of things. is number one, there's more than one kind of generative AI. Now the world is really only familiar with one, which is ChatGPT. We were in the closed beta four years ago of ChatGPT2. So we're OGs with them. But I built an engine that's mine. It's proprietarily mine. So we like to think of lately as a fully loaded ice cream sundae.

20:36There's a banana and hot fudge and a bunch of different flavors of ice cream and whipped cream. And I made this sundae all myself. On the top, there's a couple of tiny, tiny chocolate sprinkles. I like chocolate. They could be rainbow. And that's IBM Watson and Google Pegasus and MeaningCloud and ChatGPT. Now around the data sets, what's been exciting is, you know, there's this huge wave of, oh my God, public data, copyright infringement, you know, all that. And we only use a private data set at lately, your data, and we don't share it. And so we've been getting greenlit by legal. And so that kind of explosion of information is obviously really helping us a lot because a lot of companies are now banning generative AI company-wide because people are cheating at work and they're worried about data being put out into the Google's public, you know, database, for example.

21:27But the other thing too is, so generative AI, as we all know it, not me, but the world, is type in a few things, get a big long thing out, right? That's the deal. Lately works the opposite way. You have to give us a very long thing first, like a video or a blog or audio file. And then we're going to take that unique model we built from you and clip it up into dozens and dozens of very small social posts. And the way we're able to capture your voice, which ChatGPT can't do because they don't have, and we're not competitors with them in any way, but they don't have a data loop. But I do. You connect your social channels to lately.

22:06I can read your analytics. I can see what's performing best. And I'm learning all day long, every day long, whether you're publishing through lately or something else. And then every change you make to anything you publish, if you edit it, if you add a picture, if you delete it, if you don't publish it, I'm learning. Right. So it's a tight loop. The other thing here is, you know, now that there's been this explosion of words, the problem still remains. How do you cut through the noise? This is the problem always of all social marketing and all sales enablement, you know, whatever. Guess what the answer is?

22:40It's the same answer. Be more human, right? Be more human. So when you build that into your algorithm, so I'll give you an example. We worked with Philips Electric, which is they've rebranded to Signify. And we got them on average 124 % more engagement on LinkedIn. We saved them 84 % of their time. And I think it was 82 % of their budget. And it's not just on creating content, but it's on creating content that works, right? This is the thing. Now, you can ask generative AI to create social media posts for you, but they'll have no way to know if it's the right ones to create because there's no information to check that data.

23:22Let me follow up on something you said there. I'm going to ask the dumbest question you've ever heard. You just said, be more human. What does that mean when you say that? So when you're saying that's the thing to do, I hear you. And I think I know what that means, kind of like, but I probably don't because I think it means something very specific to you in the context. Can you share that a little bit, what that means to you? Yeah. Thanks for asking. No one has ever asked that. They really should, right? I'm very good at dumb questions. I excel at them. No, no, no, not dumb. Very smart question.

23:56So it's using that instinct, right? The, because you can't science everything to death, it's predictive and it's a little bit indescribable. So you have to know whether that joke is going to land, right? Like a comedian, that's the deal. Sometimes the atmosphere is ready and sometimes not yet ready. You can make a joke about that just yet, right? That ability to read the room, which is again, something I did in the dark for years is a, I don't think that can be taught to be honest to you. um however it does come from making mistakes right so the more mistakes you make the more you learn from them um being human on social can be it's funny you see people experimenting on linkedin a lot right now doing like more personal things i do incredibly personal things on linkedin and people are like what you do yes they're usually having to do with animals so yes a lot.

24:51Great. Is it working? You're getting engagement from it? You know, the truth is I'm not measuring it, but it definitely is different from what everyone else is doing because I'm just for anyone out there, because people hit me, I'm really bored with everyone self-promoting on LinkedIn all the time. And so like when people hit me up with those messages or they're in my feed, I just turn off everything. So yes. There you go. Yeah. So you're experimenting. Anything that's not the usual may capture my interest. If it's the same old stuff, I'm asleep as my eyes glaze over your post. I love that. And that's a key thing to think about.

25:28So say if I want to promote something that I'm doing, sometimes I'm lazy, of course, but there's only two objectives on social media, click and share. That's it, right? So if you back into it, knowing that that's the deal, I need to write copy that's either shareable or clickable. It's very hard to get to the clickable point because that's a lot of trust you're asking from people and time and all that. But sharing is easy because sharing is all about the ego. And if I give you content that's worth sharing, you look good, right? Kind of like if someone in college brings you the latest record from whoever, and then you play it for a friend, now you're the tastemaker.

26:01Same idea. So shareable content, like Gary Vee has created it. You can just spread joy with positive messaging. I find that any content that includes God does really well. Certainly negative stuff. Like I've posted, you know, in tears, getting a no from an investor. I mean, I was experimenting with all these things. I'm doing it on purpose. You know, I'm trying these things out. I often will do things like I'll say if there's an interview coming up with Chris Daniel and moi, like I'll refer to myself as moi because I like Miss Piggy and like you might not know that about me, but I didn't, but I have a raccoon named Miss Piggy for real.