March 26, 2025 • 37 mins
What happens when a seasoned digital marketer embraces AI not just as a tool, but as a foundational part of their entire business operation? Justin Lane of Adalane Media Group pulls back the curtain on his agency's remarkable AI transformation in this eye-opening conversation.
Justin reveals how he's built a sophisticated system of custom AI agents that have fundamentally changed his marketing approach. "I fire up a new agent for everything I do," he explains, detailing how he trains specialized models for different marketing tasks—from analyzing Google Ads performance to generating platform-specific content that doesn't sound AI-generated.
The results are striking. By analyzing client phone calls, Justin's AI discovered keywords that customers mentioned but never searched for online, leading to 800 targeted LinkedIn posts addressing these hidden pain points. His systems can personalize websites for individual visitors, rewrite content to match stakeholders' preferences, and even generate code so effectively that Justin admits, "I'm done writing code. I literally just have it write it."
What makes this conversation particularly valuable is Justin's candor about both capabilities and limitations. He explains the technical concepts of LoRas and RAG in accessible terms, distinguishes between hallucinations and creative output, and acknowledges that despite these advances, human creativity remains essential. "All of this AI was trained on things that creative people built," he notes.
For marketers wondering where to start their own AI journey, Justin offers practical advice: begin with accessible tools like ChatGPT or Claude, focus on asking the right questions, and systematically build your knowledge base. The competitive advantage, he suggests, belongs to early adopters—at least for now.
Ready to explore how AI might transform your marketing approach? Listen now and discover what's possible when you embrace the AI revolution happening right under our noses.
This podcast is proudly sponsored by USC Annenberg’s Master of Science in Digital Media Management (MSDMM) program. An online master’s designed to prepare practitioners to understand the evolving media landscape, make data-driven and ethical decisions, and build a more equitable future by leading diverse teams with the technical, artistic, analytical, and production skills needed to create engaging content and technologies for the global marketplace. Learn more or apply today at https://dmm.usc.edu.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:01):
Welcome to Mediascape insights from digital
changemakers, a speaker seriesand podcast brought to you by
USC Annenberg's Digital MediaManagement Program.
Join us as we unlock thesecrets to success in an
increasingly digital world.
Speaker 2 (00:22):
It is truly a privilege and honor to have
Justin Lane of Adeline MediaGroup on the podcast.
We have known each other almostsince when your business
started in 2019.
We met shortly after that.
Justin, again, I'm so thrilledto have you here and I'm really,
really excited about what we'regoing to be talking about today
.
You've been in digital forquite a while, so I wanted you
(00:47):
to share a little bit about yourbackground, google and AdWords
and all of that stuff, and thenwe'll get into the really
exciting stuff, which is howyour agency has even transformed
from the free gen AI world intothis new world.
That's not just generative AI,but all of the really, really
cool things that you're doingthat I must have in my life
(01:08):
today.
Speaker 3 (01:09):
Let's do it.
Thank you for having me.
I'm excited to be here.
So yeah, if we go back 20 years, my first entry into
advertising was when Google cameout with their organic search
engine and I was a programmerdeveloper at the time and I was
tasked with developing hundredsof websites programmatically to
game that and take the toplistings, the top 10 search
(01:32):
results of everything, and I'vebeen playing that game.
How do I generate enoughcontent?
How do I understand analgorithm and then generate
enough content to get in it?
That just transformed into thepaid side as well.
So how do I do really good atpaid and what systems?
Legalzoom was my first entrythere, when they started writing
systems to help them managetheir big budgets and moving
(01:52):
through.
So incorporating technology toreally amplify my output has
always been the core of what Ido and I think, like seven years
ago, openai started providingtools to semantically process
data.
So I've been doing that for SEOetc.
And then when all of thesemodels came out for us to access
, that's where my whole worldchanged.
(02:15):
I quit writing code.
I have AI that writes the codeand remembers my code and keeps
iterating over my code base towrite the code for me so I don't
have to write code anymore,just quit programming almost
altogether.
And then for the marketing sideof it, it's been critical in
understanding signals that Icouldn't detect that the
algorithms themselves, likeGoogle's AI or other systems,
(02:37):
have made shit.
Like I move my budget this muchand all of a sudden two ad
groups have a higher CPC.
It understands like maybe therewas too much inventory and now
Google is focusing on two adgroups have a higher CPC.
It understands like maybe therewas too much inventory and now
Google is focusing on two adgroups and the other two it's
just going for the top clickLike really interesting theories
are coming out of it and thenwe test the measure against that
.
So AI has just I fire up a newagent for everything I do
(02:59):
honestly on my local computer.
Speaker 2 (03:01):
Okay, this is yes, so let's talk about that.
You're in marketing tech,advertising operations.
When you started the company,what did it look like compared
to now?
And I know that's a verylengthy question answer perhaps,
but let's get into it.
Speaker 3 (03:17):
Yeah.
Well, when I started it it wasalmost out of necessity.
The company I was working atpreviously was an affiliate
management company and they werelarge and they wanted to move
into the measurement ofe-commerce space.
They were acquired and thatdigital measurement platform the
new company didn't want anymore, so my whole division was laid
(03:39):
off.
So a good friend of mine wentto W Promote and wrote Polaris.
He was on the team with me andthen so I kind of went off and
said, well, I was testing onthese large clients at the time
MusclePharm, kings, hawaiian,testing this stuff for free.
I was like I'll just say, hey,now you got to pay me and I'll
be your agency of record.
So that's kind of how itstarted.
(03:59):
The transition from being thetalent in the company to being
accountable for everything andjust trying to humble down as
fast as possible was probablythe most brutal thing ever.
So after I went through a fewof those things, I really got to
(04:20):
the core of like what does theclient need?
And I'm still have frictionbecause there's levels of
optimizations that I know I needto apply.
And then there's business logicwhere they're like but we feel
like it should be this and I'mjust like.
Well, me and these systemsdon't agree with you.
Historically that's been tabledata and charts and
visualizations to explain it andthey're kind of like we don't
get it.
(04:41):
Ai has bridged thatcommunication gap and said
here's a bunch of complicateddata and this guy is super poor
at explaining it to you.
Let me try.
And it just there it is.
So I've leaned into AI foreverything.
I've had to account for thehallucinations and the training
of specific things like thetemporal memory of it,
(05:03):
remembering our lastconversation, or even going back
let's talk about two weeks agowhat we did.
So I've had to incorporate thatkind of temporal knowledge into
the context.
But now I basically take amodel like a llama or something
and I'll just build a Lauraspecifically to handle Google
ads and I'll attach that Lauraand it's that is the model
(05:24):
parameter training that was doneagainst that model, and then I
can let's talk about Google ads,where do we leave off, et
cetera.
And now I'm deploying that foreach client.
Like their website is an agentthat understands its marketing
and all that stuff that comesthrough and it writes content is
pretty wild, yeah.
Speaker 2 (05:41):
Wow.
Now a question on that, becauseI've talked to other people who
are.
You know, personalization issuch a big thing now, whatever
that means, right?
Does that mean, though, thatsomebody could come to a website
and that the website, as anagent, would have input if that
person is an existing customer,on kind of their purchasing
(06:02):
habits, color waves, whatever itis that they might purchase
from this company, and then beable to personalize the website
for each person?
Speaker 3 (06:08):
100%, wow, 100%.
Yeah, that is the new temporalknowledge graph that they're
putting out.
So there used to be.
The RAG concept was like justfeed me information that
basically, in simple terms, getsprepended to my prompt.
Speaker 2 (06:24):
Wait before you go further for the audience.
What does RAG stand for?
Speaker 3 (06:28):
I have to look it up.
It is RAG and Laura I have to.
I think Laura is low rankingRetrieval, augmented Generation,
and then Laura is the otherimportant one and that is low
rank adaption.
So if you think about when yougo to chat GPT, you need to feed
it a document and then say,okay, we're going to reference
(06:50):
this document.
I'm going to ask you a fewquestions about it.
That's basically rag.
So imagine, like the actoranalogy I like to use and I
don't know if it's great is likean actor, they will have to
learn a British accent to play aBritish role.
So you're altering theparameters of how they speak in
(07:10):
general.
You're not giving them newknowledge.
You're saying do what you do,but do it differently.
That's altering the parametersof the model.
And then when you feed them thescript, that's the rag part of
giving them context.
This is what I want you to dowith it.
The model, however, doesn'ttake up any of your token space
for the prompt.
So that's why we would trainthe model to be intimately
(07:30):
familiar with the high levelconcept, topic, website, client,
whatever all the products andservices they have, so we don't
have to keep reminding it in thecontext.
Let's talk about this product.
And, by the way, this is whatthis product is.
Don't forget, we've alreadykind of modeled the parameters
in the altered, the parametersin the model for the LoRa.
Speaker 2 (07:50):
For that, that was a lot Is a poor man's way of
saying of this.
I'm just thinking with Claudebecause I like to use Claude.
I have projects set up right,so in the projects I've fed them
here is everything about thisproject that I'm doing.
Here's some language that otherpeople use that I like, and
then I can ask it for certainthings based on different
(08:12):
prompts that I give.
So it's that, but on a muchbroader scale.
Speaker 3 (08:17):
No.
So I'm going to sue.
Claude's probably going to hateme and there's probably going
to be some AI people out therethat know how Claude's actually
doing this, but I'm going tosuper simplify it.
So Claude, basically, is takingall of that information and
summarizing it into theappropriate tokens that it needs
.
So it's super condensing it andprepending it to all of your
(08:40):
queries, saying talk like this.
So it's just taking everythingit needs to know about all of
your documents, but giving abrief enough summary about it
and tokens to say and we need tobe this different when we
respond back.
And it's stacking all of thatbefore you write your prompt.
Then you write your prompt andthen it's like, based off of all
of this, I will answer thatquestion this way.
Okay, that is still retrieval.
It is still going out andcapturing context each time you
(09:03):
query it.
It's going out and grabbingwhat you did and coming back.
It's not inherently trainingitself to always understand that
whenever you come and justwrite a prompt with no context,
to say this is what I understand.
That's what the LoRa trainingis is to make it fully
understand.
Speaker 2 (09:21):
One thing that I know that you and I have talked
about recently was the fact thatyou stack all these models.
So I'd love for you.
You know I don't want to getyou to give away your secret
sauce, certainly but I do feellike what you're doing is so
different than what a lot ofpeople in traditional agencies
are still doing and that youknow, while they might use some
(09:42):
Google Analytics AI tools, youknow other tools and different
programs.
You have a full system.
Speaker 3 (09:50):
Yeah, and that came from like trying to to prompt
everything my way through it andI found like there's times like
I'll go through.
Here's an example of a client.
They have a blog, they have 600posts and they every day
there's a business update.
We would like to be more aboutAI, so all of those posts need
to now incorporate AI into it.
(10:12):
So the system reads all 600blogs every day and gets the
context of the business goalswhat is the business about and
what is it and it will rewritethose blog posts and they just
sit there as an MD file.
We don't use them.
We could, but then from there ittakes all of that content and
says if you were to rewrite thishere's a LinkedIn post, here's
a meta post, here's a Reddit adit builds all of the content you
(10:35):
would need for distribution.
That gets tricky because all ofthose different places require
different text lengths for theirdistribution, et cetera.
So I have another model thatI've loaded the web page of the
requirements.
It says go get the requirementsof a Reddit ad and then I put
it in what's called an open APIformat.
So it's almost like a schema,if you will and it basically
(10:57):
says it needs a headline and itshould be this long, et cetera.
Right Running against thatschema.
Even then the AI doesn't reallyget character lengths right for
building an ad.
It's somewhere too long.
So I'll add context and say becareful to do this and be
careful, so all of that contextis stored in that model.
I don't have to go to theoriginal model and say rewrite
(11:18):
this article and then write anad and also don't forget about
this ad.
I don't have to prompt all that.
I basically just hand off thecontent to this other one and
say let's write a Reddit ad andits only job is to be good at
writing Reddit ads.
It's its only job.
Same with Google, same withMeta, et cetera.
So that's when I just shipthese over to do that.
Speaker 2 (11:36):
Wow Okay, amazing.
Speaker 3 (11:45):
So it sounds like your workflow, your work life,
has changed a lot, even in thelast two, three years.
Yes, in the last few weeks I'vebeen working on this stuff for
clients in a silo, and not eventelling people.
I do it Like clients, because Idon't publish.
By the way, this was allwritten by OpenAI, right?
I don't tell anybody that it'shappened, I'm just like here's.
One of the big ones was all thephone call transcripts from one
client.
I consumed all of that and weanalyzed it and we said what
(12:07):
keywords are we buying andwhat's the gap?
And then what we found out wasno one searches for a particular
keyword, but they alwaysmention that that's an issue on
the phone call.
We should probably talk aboutthis particular product offering
that we have and incorporate it.
So then that wrote a bunch ofLinkedIn content, 800 posts of
(12:27):
problem solutions.
Cause it from the phone calls.
It identified from all of thosephone calls where they
mentioned that product there wasa problem and then how we're
the solution or this client isthe solution.
And then it crawled all of theirblog posts and ranked them and
said which of these blog postsis the best suited to present to
someone to explain how we solvethe problem in this area.
And then did all the publishedall of those.
(12:49):
Here's your LinkedIn post,here's the problem, here's a
solution.
Read more at our blog thattalks about it.
So they had 800 posts to choosefrom, to put it anytime they
wanted to, to pump out.
And they didn't know that wasall AI.
They're just like you must havea bunch of content writers.
I'm like, well, we use someprograms, right.
So I've been shooting down thisAI tunnel rapidly, thinking I'm
(13:13):
way behind and everyone else isway ahead, and I am, in the
last two weeks, just now, comingto the surface of like, hey,
there's probably some cool toolshere that people could use.
Speaker 2 (13:23):
Yeah, exactly.
There is also the question Ihad this discussion earlier
today about AI creating images,creating graphics.
You know creating our posts,and are they too generic?
How do we make sure that theyare relevant for our audience,
that it's not just you knowmaking a sea of sameness when it
(13:45):
comes to what we see from.
Okay, this, you know, from thiscertain category, like all
vacuum cleaners now have theexact same copy, or the, you
know, everybody who's a servicewriter in this area have the
exact same copy, the exact sameimages.
So how do you counter that?
Speaker 3 (14:00):
Yeah, so this is.
There's probably going to be AIexperts that are like that's
not quite how it works, butbasically there is one wise old
man and that is the AI thatconsumed all of the knowledge
and that's the chat GPT that youtalk to.
It has one tone and since mostof the content that it is
consumed has been storytellingand fantasy that kind of stuff
it will always start with in aworld and be that dramatic.
(14:23):
Right For both images and themodels for generative AI.
You just need to take a Lauraon both of them and say I know
you want to talk like this fellabecause that's your default
mode, but if you can just feedit some tokens tokens is the
language that they speak in feedit some tokens to say, when you
(14:44):
talk, just be this differentyou will completely alter the
conversation.
So I have one of my core agentsthat I use and I probably
shouldn't say this until I getsued but it is trained on David
Ogilvie, so it writes just likehigh impact fact hitting.
This is how we write content toget conversions coming through
(15:04):
and it is completely differentthan what you would get Like.
I'll run them both.
I'll run the normal one andthen the David Ogilvie one, and
they're just night and daydifferent.
And then clients sometimes havetheir own authors.
So if they have an author I'lldownload their stuff to talk
like them.
And then I don't really trainoff of this because it's not a
lot of data, but when I publishsomething to like a Google
(15:27):
document for the client toreview, each individual, each
stakeholder will leave commentsand I'll take those comments and
use AI to summarize them andturn them into some type of
context to lean into when wereview.
So it's like what would thisstakeholder think of this piece
of content?
Or alter this content tosatisfy the stakeholder?
Either one right.
So it's almost like I'll writea piece of content, or AI will
(15:50):
write a piece of content andthen I'll get feedback from the
other AIs of the stakeholders tosay this is what they would
think.
And before I present it to theclient I can kind of say, yeah,
I think they're all going tolike it, that kind of thing.
Speaker 2 (16:02):
Yeah, this is far above and beyond, I think, where
most people are right now intheir AI journeys.
So how are you sharing thiswith clients, or what are you
doing and how are you gettingmore clients?
Are you using these sametechniques?
Are you doing any advertising?
Is it all word of mouth at thispoint?
Speaker 3 (16:22):
Nothing.
You're probably the secondperson that even knows that I do
this, even through my clients.
So no, I tried on my websiteand I was like I'm going to
build these agents.
But all of these other systemsstarted coming out where like,
oh, we have these AI agents andyou upload your data and
understand your context and I'mlike, yeah, but read the end
(16:43):
result Even like SEO seems to bethe go-to example for all of
these agents how to crawl thingsand do this.
And when you read the end result, it is like everybody else's,
like it's just what an ai wouldsay.
There's no like we need toreally heavily modify the output
of this to have the a bettertone, like the client has.
So I got a little bitdiscouraged on saying I just
(17:06):
don't want to be another onethat does AI because I actually
download the models and actuallytrain them for the client.
So I just haven't said anythingat all, and even the clients
that I have don't even reallyfully understand what I'm doing.
So there's no testimonials oruse cases of us actually using
AI, right.
Speaker 2 (17:24):
But there are testimonials for the results
they've gotten and the work thatyou're doing for them.
Speaker 3 (17:30):
Yes, but those, oddly enough, are in the form of like
now we had this expectation tomeet of providing this much
content for social media right,and now we're hitting that goal
consistently and our team isn'tstruggling to make it, and all
of the content we're generatingis really thoughtful and
contextual to the environment.
If there's a trade showhappening, I also crawl.
(17:51):
So for every phrase or keywordthat a client has, I crawled a
Google search engine that Ibuilt for them through Google
and I get the top two results orthe top two pages of results.
So I'm always understandingwho's ranking where and what's
happening.
So if news comes through of likea conference or something, the
AI will understand that oh, itlooks like there's a conference
(18:12):
coming and we have content aboutthis, and so its recommendation
would be let's publish thisLinkedIn article versus all
these other ones, because theconference is coming.
And so just to be ahead of thecurve, like to say I'm going to
write about this, is that cool?
And they're like, oh yeah, weforgot that was coming up, like
that's a good idea.
Just to be ahead of the curvefrees up so much resources for
(18:33):
folks to just get their contentout, not some AI.
Let's pretend like we're anauthority at something right.
It's their content what they'regoing to do, but just get it
out in a low friction highlydistributable way.
Speaker 2 (18:54):
So your workflow and your life has changed a lot.
You're also saying somethingthat I think is really important
, which is you know, there's oneschool of thought that you need
to know AI tools because you're, when you're, you know, the
person who knows the toolsversus the person who doesn't
know the tools is going to havethe job.
There are other people say,well, ai is probably going to
replace a lot of these jobscompletely, because the
technology that you're sharingright now, but you had to
(19:17):
program that you had to put allof these together to, in essence
, make sure that all of thesedifferent models are doing
exactly what you want for eachspecific client and for each.
In essence, make sure that allof these different models are
doing exactly what you want foreach specific client and for
each.
You know each goal of eachclient as well.
Speaker 3 (19:31):
Yes, I coded all of that.
It's probably that component ofmy software itself is probably
10,000 lines of code.
I say I coded it, but I haveall like Claude coding it for me
Well take the credit.
Yeah, yeah, take the credit.
But the thing that you have tobe careful of and I think a lot
of people will talk about this,especially as you get into the
more agentic world, where youhave agents talking to other
(19:52):
agents and doing stuff and theagent is calling other agents is
they all still hallucinate alittle bit.
So if they don't know something, they'll make something up.
So I have one of theinstructions.
When I write code is I have atremendous amount of output.
I did this and here's why.
So almost the reason why we didthings, if I can get reasoning
logic out of it, this is whywe're coding this and this is
(20:14):
the response.
So, as a human, I can review itand say that's not that far
fetched.
That's cool.
Fortunately, I'm in advertising, so hallucinations make it more
creative.
But if you were doing somethingsuper complex that required
precision and a deep knowledge,I don't know that AI is the
place for that, but yeah, I'vecoded all of this up sitting on
(20:38):
a server at DigitalOcean,believe it or not.
Yeah.
Speaker 2 (20:42):
So, yeah, agents are here.
Right, I listen to a lot of AIpodcasts and a lot of the
advertisers are like the agentsare coming.
Well, they're clearly here.
It's not a reason utilizingthem to the full effect.
And especially people like me,who I'm not a technical user
right, I'm learning from thebusiness perspective.
I do know a lot of people inthe ai world.
I know a lot of founders ofsome of the smaller tools, so
(21:05):
I'm getting to learn them.
But what else do you thinkpeople who are in marketing
advertising, specifically on thedigital side, what are the most
important skills that they needto have?
Because it sounds like it'sbeyond just knowing how to use
ChatGPT or Cloud or Notebook, lmor Gemini or Copilot.
Speaker 3 (21:25):
It doesn't have to be .
You could use those.
So if you really break it downand take away, like the actual
Laura training that I'm doing,I'm just doing really
sophisticated prompt chaining,right.
So that's even the context.
Like I'm taking big sentenceslike instead of saying I got 15
drops on me today, it's wetoutside, I can just say it's
(21:47):
raining today and that token isvery small and I can put that in
and still have a tokenallowance in the GPT prompt if I
want to for context, right.
So I think a hundred percent,like OpenAI, ChatGPT is moving
to an enterprise model wherethey want to store all your
information.
Like just understandingfundamentally how the AIs work
(22:07):
in the models but then gettingreally good at prompting them.
You can't go wrong.
Any engine you cannot go wrongor any model going and prompting
I think is tremendouslypowerful and Clod is amazing
because it just has so likemillions of tokens you can put
in.
So just dump a graph in and say, tell me about this advertising
and generate a graph.
(22:27):
Right.
Like, definitely go, do that Ahundred percent.
You could get through your dayto day and make yourself worth
10 of you just using these freeand open source tools, Right?
So I would say Claude Lama andopen, like Chad GPT, play in
them equally.
Right Lama is freaking awesome.
I guess Glock too, but inTwitter, but I don't use that
that.
(22:47):
I guess Grok too, but inTwitter, but I don't use that
that much, but I want it.
I have nothing against it.
So definitely keep promptingright, Because how you ask it
doesn't matter if you built yourown custom solution or whatever
.
How you ask these questions isgoing to be critical because
it's going to take the tokens ofthat question and run it
against its token database ifyou will and say how should I
(23:10):
answer this?
So that's critical.
And then just a fundamentalunderstanding of if you go to
chat GPT, you're probably goingto get the voice of chat GPT,
that model, how it trained, likethe grandfatherly voice of I
know everything.
How do you make it sounddifferent?
You can do that with promptingor you can use LORAs for both
the images to make the imagescome out different when you
(23:31):
generate them, as well as thetext.
So just understanding that andhow they get context and
temporal data.
And honestly, you can go askChatGPT like, hey, can you
explain to me LORA and RAG andtemporal data and how that would
all be used to do this, andjust read the answer and know it
, and then keep using those todo your job in media, because
(23:51):
there's still the interfacesthat you have today that are
free for the most part areincredibly powerful.
Incredibly powerful for it, soI'd say, keep going.
Speaker 2 (24:03):
Okay, nice.
Yeah, I will say that most ofmy students know maybe not
everybody listens to thispodcast, but I'm also in school
right now and for some of myclasses I will put in.
Can you explain this concept tome, exactly what you're saying?
Because, to be perfectly blunt,claude explains it a lot more
clearly to me than the videos.
(24:23):
The textbook sometimes theprofessor, sometimes the
professor, you know it breaks itdown so that I can really.
Okay, now I understand how touse Tableau appropriately, or
how to you know, because I justfinished a data and business
analytics class and had to workin Solver and Power Query and
Tableau and like all thesedifferent things to create
(24:45):
graphics and images.
But so I didn't ask you to dothe work for me, but I asked it
just step by step.
Tell me you know I'm gettingthe wrong answer.
What should I do differently?
Can you look at what I put infor my codes?
And so it's been a reallyamazing asset.
I sometimes call it my bestfriend.
Speaker 3 (25:01):
I would challenge you in two areas there.
One, before you ask it toexplain something, give it some
context and maybe things you'refamiliar with Say, I'm used to
Excel, Right, and I know how todo this in Excel, and also I
need to understand this becauseI'm going to teach it to other
people and that willdramatically change its response
(25:24):
out and give you scenarios onhow to help you do that.
And then the second thing wouldbe if it can do it, let it do
it.
Let it write the code.
I haven't written a SQL queryin over a year.
I dumped the database table andI say I need to know this about
this and it will write a SQLquery and I'll paste it in and I
don't even think about itanymore.
Nice, Like it's perfect.
So as a developer who writescode and was like and I have an
(25:47):
advantage in this ecosystem ofadvertising because I can write
code to do things faster, I'mdone writing code.
I'm literally done with it.
I just have it write it, Icompile it.
If it works, I'm like great.
Could it be better?
I don't know, but it works.
So we're done.
It's pretty wild.
Speaker 2 (26:03):
So, justin, what does success look like to you today?
Because I'm sure it's changedalso since you started your
agency.
Speaker 3 (26:10):
Yeah, I don't.
I really don't know.
This is probably the first timewhere I really don't feel like
I have an advantage, becausethis is becoming accessible to
people so rapidly.
So everything that I've beenworking on, in a year's time
everyone will have access to.
So this is almost like the wildwild west the first to deploy
(26:30):
it and be the agency of record,if you will, to help people get
integrated.
They're going to win, but in ayear from now, me to come in and
say, no, use me.
Instead, there will be noadvantage.
That I have right, because AIis the same as AI.
If they already have an AIconsulting agency, I won't win
their business unless they justabsolutely suck, and I doubt
(26:52):
they will.
You know what I mean.
So it's almost like be thefirst one there, but then after
that, once everybody has an AIagent, there's no room for
people to build AI agentsbecause they have AI agents that
will probably build the new AIagent.
So I'm decently terrified thatI won't have a home in the world
in the future and I don't likewhere I would make money.
(27:14):
It would still be inadvertising, but again, like
bots will just see.
I just had a conversation todaywith someone and they were like
we have 3 000 domains in theirprime real estate, domains like
razorbladecom.
That wasn't one of them, but itcould be and I was like that
doesn't matter to me anymore.
Like, yeah, you're going to gettype in traffic, but ideally
what you would do with that isyou would put up razorbladecom,
(27:38):
you would put a model behind it,train it on everything it needs
to know about razor blades andjust have a prompt and say you
could go to chat GPT and it'lltell you what it knows about
razor blades.
But if you come torazorbladecom, this thing will
tell you, up to date and withprecision, everything you wanted
to know about razor blades.
Right, and it's just a promptedwebsite.
(27:58):
It's not even what you justcome ask it.
I just don't know.
I have no idea where I'm goingto be in six months, no idea.
Speaker 2 (28:08):
Let's talk further about this.
The last prompt that I gave tomy students for the first
quarter of the program for DMM510, which is all about
advertising you know dataanalytics.
You know how to measure ethics,privacy, implications.
We got into mad tech, but Ialso like to also incorporate AI
(28:31):
tools into every class.
And the last prompt I said,instead of doing the
participation prompt, read thispress release from a guy who I
was on his podcast.
He's in, I think, sweden.
You know about Web 4.
And here's what he's trying todo and here's what he thinks Web
4 is going to look like.
And it was very puppies andsunshine and butterflies and
rainbows oh my goodness rightlike it's going to solve all the
(28:54):
world's problems, and thisisn't this.
This will happen out of it forme.
I just get excited to thinkabout and I'm sure I can do this
now and you're gonna have toteach me how you know having an
agent of my own, like mypersonal assistant agent, like
go through these emails.
You know, having an agent of myown, like my personal assistant
agent, like go through theseemails.
You know, put them intocategories by X, Y, Z, put them
in this folder.
(29:14):
These emails answer this waybecause I get 20 million guest
requests for my podcasts on adaily basis and you know the
people follow up if I don'trespond right away and I don't
usually have time to respondright away but to yeah, okay,
book this appointment for mebecause I don't have time to
look through their calendar,find time that syncs up with all
(29:36):
of that kind of stuff, and thenalso being able to talk to
other people's agents, as youmentioned.
Speaker 3 (29:42):
Yeah, that's here right now and all that needs to
happen there.
I think you'll start to hearpeople more talking about an API
driven web versus like a dub,dub, dub web.
And that's because if, like,you could go in and tell your AI
agent, your specificpersonalized agent, this is what
I ate today and now I'm feelingthis, like I feel sick.
(30:04):
If you tell it everythingyou're doing, the pattern
recognition will be tremendousand it can tell you.
On a Tuesday, you typically eata lemon today, but every time
you do it you don't feel wellafter.
Maybe don't do that today, andhere's the reason why.
Right, but you have to feedthat data.
But if everything's an API yourcalendar, all of your emails,
your health apps, all of that itcan just consume it and detect
(30:29):
these patterns without youhaving to prompt it.
So you will be the blocker forhaving a great AI agent because
it just won't have enough ofyour data to make really good
decisions for you.
So, yeah, I would say definitely, build your own AI agent and
figure out how to get all of thedata that matters to you into
it and it can have temporalknowledge of, like let's just
(30:51):
review yesterday, what wasyesterday?
Like.
Or you can say let's go backand review all the times I ate a
lemon Is there a pattern thereof me hating that or something
like that?
Right, and it'll just like.
You can just talk to it and saylet's talk about lemons.
I eat them.
Why are they great for me?
I no.
Every time you eat it, you feelhorrible and here's probably
(31:11):
five reasons why.
And here's some substitutes onwhat you should do, game changer
.
But it needs that data, thatcontext about you.
Speaker 2 (31:19):
Yeah, well, and for the lay person right, a person
who's new to AI, maybe they'rein the field of marketing
digital media already, maybethey're not, because I do.
We have a lot of students whocome to the program who already
are working in the field.
We have a lot of students,equally, who you know have done
other careers or adjacent things, and they know that they need
(31:41):
this knowledge.
So what are some tools that youwould recommend for them to
start exploring so they canunderstand agents, understand
how to create their own.
Speaker 3 (31:51):
Oh, my goodness, Honestly, I don't think I would
leave ChatGPT or Claude.
I wouldn't leave to an externalresource because all of these
external resources I've built AItools in the context of how
they want to monetize AI right.
But, you could almost say thatClaude and Gemini and ChatGPT
(32:12):
are like their own, like they'regoing to be honest with you and
tell you the evolutions ofwhat's happening.
So if I were brand like I don'tknow anything about digital
marketing and I want to be adigital marketer, right, I would
first what do I need to knowabout digital marketing?
I would ask that, put that in aGoogle doc right, Start a
folder.
And then I would go researchall of these topics.
So what does the Googleplatform say about media buying
(32:33):
and what are its best practices?
Load it.
Let's summarize that.
What does Google say about it?
Now, let's come up with somephrases and keywords that we can
search in Google and let's gosee what Reddit and Quora, what
people say specifically LikeI've tried that Put.
Say specifically, like I'vetried that, put that in.
And by the time you're done, youkind of have a really good view
(32:54):
of like, well, if I buy aclick-based, maximized click
campaign, this is what Googlesays it does.
This is the experience peopleare having with it.
Like, maybe, if I were to trysomething for a client, this is
what I would try.
And then you ask the cloud orthe GPT, like, hey, I'm thinking
about trying this for thisclient 's a website.
(33:15):
Here's the thing.
What do you think?
Oh yeah, you should totally dothat, or don't do that.
Or here's some other ideas.
You could be a media buyerinstantly and be as
knowledgeable as me.
I'm not even gonna lie like my,the competition for media
buying if you had the confidenceto trust ai enough to say the
most professional media buyer isat my keyboard's end.
I just have to ask it the rightquestions.
You would win.
(33:35):
You would win all day.
Speaker 2 (33:38):
Okay.
Well, going back to the otherside, you still have team
members, so it's not all AIdriven.
You still have copywriters,right?
You still have web designers.
You still have other people.
Speaker 3 (33:47):
Yeah, so I mean that's the other part of it is
like all of this AI was trainedon things that creative people
built, inspiring people built.
Like it's not making up copy,so you still need people to
generate original things.
And the thing about AI movingforward if we talk about like
(34:10):
okay, you can't, like searchengines will go away when people
go to find a solution forsomething, they're going to go
to an AI.
That AI is going to be lookingfor general, like inspiring
responses that people havecreated.
Like I don't know the answer tothis, or I do know the answer,
but I want a very updated orcreative answer.
This person wrote it.
I'm going to feed them in andsay this is what they wrote and
(34:32):
I will describe it.
So we've gone from likeeveryone using AI to write copy
to now AI is going to be.
If it were a human desperate forfresh content, like I'm tired
of reading my own work on theweb, someone give me something
fresh and exciting.
So I think there'll be atransition into that of creators
really having a space.
(34:53):
Now that might be creators thatspecifically copywriters that
write specifically for AI.
It might look different thanwriting for a particular brand,
but, yeah, they're necessary andAI, like I said, has its.
No matter how much you promptit, it will always say the thing
you don't want it to say.
Right, imagine a world or put arocket ship emoji, like it's
(35:16):
always going to just do that andsometimes you just can't tell
it not to.
It just won't listen.
Speaker 2 (35:22):
Yeah, Wow, we've covered so much and I feel like
it's daunting but also exciting.
Speaker 3 (35:29):
Yes, so much and I feel like it's daunting but also
exciting.
Yes, and I would.
I would challenge everyone wholistened to how I explain things
.
Go find two more people thatexplain it as well, because I am
also every day.
This changes.
So, as much as I follow thisstuff technically and read like
all the models and what they'redoing and what they're good at,
like this is hard to like have aday job leveraging AI to do
(35:51):
good things and then alsounderstand it enough to tell
people.
So take the way I said it, butdo your own research and say is
that really how that works?
Hopefully we're close.
Yeah, but, chad, you think I'mpretty smart?
Speaker 2 (36:05):
Well, that's good.
Speaker 3 (36:06):
Yeah.
Speaker 2 (36:07):
Fantastic.
You've given us a lot to thinkabout, a lot to explore.
I really appreciate everythingthat you've shared and prompted
us to do.
Speaker 1 (36:18):
Yeah.
Speaker 2 (36:19):
Yeah, any last words that you want to leave the
audience with?
Speaker 3 (36:24):
I would say, just like, the world needs explorers,
and now we have better tools soeveryone can go explore almost
equally.
Let's just see where it goes.
Keep exploring.
Speaker 2 (36:34):
Great.
Thank you, Justin.
I will have Adelaide Media inthe show notes for anybody who
wants to learn more about Justin.
I do know that we have onestudent in the DMM program who's
interning for you right now.
Speaker 3 (36:49):
Yes.
Speaker 2 (36:50):
So there may be some other opportunities open for
those of you intent in theprogram.
Who actually listen to thepodcast?
Speaker 3 (36:58):
Yes, yes.
Speaker 2 (36:59):
Yeah.
Speaker 3 (36:59):
Fantastic, I will be AI heavy, but we will go through
all the old school media buyingtechniques to get history so
you can have that advantage.
Speaker 2 (37:07):
Fantastic, and we're going to probably do it.
We'll have to do a live demo ora recorded demo at some point
as well.
So, counting on that.
Absolutely Fantastic.
Thank you to everybody wholistens to Mediascape Insights
from Digital Changemakers.
Please leave us a rating orreview, or both on your favorite
platform or as many as you'dlike to, because it really does
(37:30):
help us get more discovered,find more listeners and get more
engagement for our show.
With that, I'll be back againnext week with another amazing
guest.
Speaker 1 (37:40):
To learn more about the Master of Science and
Digital Media Management program, visit us on the web at
dmmuscedu.
Welcome to Mediascape insights from digital
changemakers, a speaker seriesand podcast brought to you by
USC Annenberg's Digital MediaManagement Program.
Join us as we unlock thesecrets to success in an
increasingly digital world.
Speaker 2 (00:22):
It is truly a privilege and honor to have
Justin Lane of Adeline MediaGroup on the podcast.
We have known each other almostsince when your business
started in 2019.
We met shortly after that.
Justin, again, I'm so thrilledto have you here and I'm really,
really excited about what we'regoing to be talking about today
.
You've been in digital forquite a while, so I wanted you
(00:47):
to share a little bit about yourbackground, google and AdWords
and all of that stuff, and thenwe'll get into the really
exciting stuff, which is howyour agency has even transformed
from the free gen AI world intothis new world.
That's not just generative AI,but all of the really, really
cool things that you're doingthat I must have in my life
(01:08):
today.
Speaker 3 (01:09):
Let's do it.
Thank you for having me.
I'm excited to be here.
So yeah, if we go back 20 years, my first entry into
advertising was when Google cameout with their organic search
engine and I was a programmerdeveloper at the time and I was
tasked with developing hundredsof websites programmatically to
game that and take the toplistings, the top 10 search
(01:32):
results of everything, and I'vebeen playing that game.
How do I generate enoughcontent?
How do I understand analgorithm and then generate
enough content to get in it?
That just transformed into thepaid side as well.
So how do I do really good atpaid and what systems?
Legalzoom was my first entrythere, when they started writing
systems to help them managetheir big budgets and moving
(01:52):
through.
So incorporating technology toreally amplify my output has
always been the core of what Ido and I think, like seven years
ago, openai started providingtools to semantically process
data.
So I've been doing that for SEOetc.
And then when all of thesemodels came out for us to access
, that's where my whole worldchanged.
(02:15):
I quit writing code.
I have AI that writes the codeand remembers my code and keeps
iterating over my code base towrite the code for me so I don't
have to write code anymore,just quit programming almost
altogether.
And then for the marketing sideof it, it's been critical in
understanding signals that Icouldn't detect that the
algorithms themselves, likeGoogle's AI or other systems,
(02:37):
have made shit.
Like I move my budget this muchand all of a sudden two ad
groups have a higher CPC.
It understands like maybe therewas too much inventory and now
Google is focusing on two adgroups have a higher CPC.
It understands like maybe therewas too much inventory and now
Google is focusing on two adgroups and the other two it's
just going for the top clickLike really interesting theories
are coming out of it and thenwe test the measure against that
.
So AI has just I fire up a newagent for everything I do
(02:59):
honestly on my local computer.
Speaker 2 (03:01):
Okay, this is yes, so let's talk about that.
You're in marketing tech,advertising operations.
When you started the company,what did it look like compared
to now?
And I know that's a verylengthy question answer perhaps,
but let's get into it.
Speaker 3 (03:17):
Yeah.
Well, when I started it it wasalmost out of necessity.
The company I was working atpreviously was an affiliate
management company and they werelarge and they wanted to move
into the measurement ofe-commerce space.
They were acquired and thatdigital measurement platform the
new company didn't want anymore, so my whole division was laid
(03:39):
off.
So a good friend of mine wentto W Promote and wrote Polaris.
He was on the team with me andthen so I kind of went off and
said, well, I was testing onthese large clients at the time
MusclePharm, kings, hawaiian,testing this stuff for free.
I was like I'll just say, hey,now you got to pay me and I'll
be your agency of record.
So that's kind of how itstarted.
(03:59):
The transition from being thetalent in the company to being
accountable for everything andjust trying to humble down as
fast as possible was probablythe most brutal thing ever.
So after I went through a fewof those things, I really got to
(04:20):
the core of like what does theclient need?
And I'm still have frictionbecause there's levels of
optimizations that I know I needto apply.
And then there's business logicwhere they're like but we feel
like it should be this and I'mjust like.
Well, me and these systemsdon't agree with you.
Historically that's been tabledata and charts and
visualizations to explain it andthey're kind of like we don't
get it.
(04:41):
Ai has bridged thatcommunication gap and said
here's a bunch of complicateddata and this guy is super poor
at explaining it to you.
Let me try.
And it just there it is.
So I've leaned into AI foreverything.
I've had to account for thehallucinations and the training
of specific things like thetemporal memory of it,
(05:03):
remembering our lastconversation, or even going back
let's talk about two weeks agowhat we did.
So I've had to incorporate thatkind of temporal knowledge into
the context.
But now I basically take amodel like a llama or something
and I'll just build a Lauraspecifically to handle Google
ads and I'll attach that Lauraand it's that is the model
(05:24):
parameter training that was doneagainst that model, and then I
can let's talk about Google ads,where do we leave off, et
cetera.
And now I'm deploying that foreach client.
Like their website is an agentthat understands its marketing
and all that stuff that comesthrough and it writes content is
pretty wild, yeah.
Speaker 2 (05:41):
Wow.
Now a question on that, becauseI've talked to other people who
are.
You know, personalization issuch a big thing now, whatever
that means, right?
Does that mean, though, thatsomebody could come to a website
and that the website, as anagent, would have input if that
person is an existing customer,on kind of their purchasing
(06:02):
habits, color waves, whatever itis that they might purchase
from this company, and then beable to personalize the website
for each person?
Speaker 3 (06:08):
100%, wow, 100%.
Yeah, that is the new temporalknowledge graph that they're
putting out.
So there used to be.
The RAG concept was like justfeed me information that
basically, in simple terms, getsprepended to my prompt.
Speaker 2 (06:24):
Wait before you go further for the audience.
What does RAG stand for?
Speaker 3 (06:28):
I have to look it up.
It is RAG and Laura I have to.
I think Laura is low rankingRetrieval, augmented Generation,
and then Laura is the otherimportant one and that is low
rank adaption.
So if you think about when yougo to chat GPT, you need to feed
it a document and then say,okay, we're going to reference
(06:50):
this document.
I'm going to ask you a fewquestions about it.
That's basically rag.
So imagine, like the actoranalogy I like to use and I
don't know if it's great is likean actor, they will have to
learn a British accent to play aBritish role.
So you're altering theparameters of how they speak in
(07:10):
general.
You're not giving them newknowledge.
You're saying do what you do,but do it differently.
That's altering the parametersof the model.
And then when you feed them thescript, that's the rag part of
giving them context.
This is what I want you to dowith it.
The model, however, doesn'ttake up any of your token space
for the prompt.
So that's why we would trainthe model to be intimately
(07:30):
familiar with the high levelconcept, topic, website, client,
whatever all the products andservices they have, so we don't
have to keep reminding it in thecontext.
Let's talk about this product.
And, by the way, this is whatthis product is.
Don't forget, we've alreadykind of modeled the parameters
in the altered, the parametersin the model for the LoRa.
Speaker 2 (07:50):
For that, that was a lot Is a poor man's way of
saying of this.
I'm just thinking with Claudebecause I like to use Claude.
I have projects set up right,so in the projects I've fed them
here is everything about thisproject that I'm doing.
Here's some language that otherpeople use that I like, and
then I can ask it for certainthings based on different
(08:12):
prompts that I give.
So it's that, but on a muchbroader scale.
Speaker 3 (08:17):
No.
So I'm going to sue.
Claude's probably going to hateme and there's probably going
to be some AI people out therethat know how Claude's actually
doing this, but I'm going tosuper simplify it.
So Claude, basically, is takingall of that information and
summarizing it into theappropriate tokens that it needs
.
So it's super condensing it andprepending it to all of your
(08:40):
queries, saying talk like this.
So it's just taking everythingit needs to know about all of
your documents, but giving abrief enough summary about it
and tokens to say and we need tobe this different when we
respond back.
And it's stacking all of thatbefore you write your prompt.
Then you write your prompt andthen it's like, based off of all
of this, I will answer thatquestion this way.
Okay, that is still retrieval.
It is still going out andcapturing context each time you
(09:03):
query it.
It's going out and grabbingwhat you did and coming back.
It's not inherently trainingitself to always understand that
whenever you come and justwrite a prompt with no context,
to say this is what I understand.
That's what the LoRa trainingis is to make it fully
understand.
Speaker 2 (09:21):
One thing that I know that you and I have talked
about recently was the fact thatyou stack all these models.
So I'd love for you.
You know I don't want to getyou to give away your secret
sauce, certainly but I do feellike what you're doing is so
different than what a lot ofpeople in traditional agencies
are still doing and that youknow, while they might use some
(09:42):
Google Analytics AI tools, youknow other tools and different
programs.
You have a full system.
Speaker 3 (09:50):
Yeah, and that came from like trying to to prompt
everything my way through it andI found like there's times like
I'll go through.
Here's an example of a client.
They have a blog, they have 600posts and they every day
there's a business update.
We would like to be more aboutAI, so all of those posts need
to now incorporate AI into it.
(10:12):
So the system reads all 600blogs every day and gets the
context of the business goalswhat is the business about and
what is it and it will rewritethose blog posts and they just
sit there as an MD file.
We don't use them.
We could, but then from there ittakes all of that content and
says if you were to rewrite thishere's a LinkedIn post, here's
a meta post, here's a Reddit adit builds all of the content you
(10:35):
would need for distribution.
That gets tricky because all ofthose different places require
different text lengths for theirdistribution, et cetera.
So I have another model thatI've loaded the web page of the
requirements.
It says go get the requirementsof a Reddit ad and then I put
it in what's called an open APIformat.
So it's almost like a schema,if you will and it basically
(10:57):
says it needs a headline and itshould be this long, et cetera.
Right Running against thatschema.
Even then the AI doesn't reallyget character lengths right for
building an ad.
It's somewhere too long.
So I'll add context and say becareful to do this and be
careful, so all of that contextis stored in that model.
I don't have to go to theoriginal model and say rewrite
(11:18):
this article and then write anad and also don't forget about
this ad.
I don't have to prompt all that.
I basically just hand off thecontent to this other one and
say let's write a Reddit ad andits only job is to be good at
writing Reddit ads.
It's its only job.
Same with Google, same withMeta, et cetera.
So that's when I just shipthese over to do that.
Speaker 2 (11:36):
Wow Okay, amazing.
Speaker 3 (11:45):
So it sounds like your workflow, your work life,
has changed a lot, even in thelast two, three years.
Yes, in the last few weeks I'vebeen working on this stuff for
clients in a silo, and not eventelling people.
I do it Like clients, because Idon't publish.
By the way, this was allwritten by OpenAI, right?
I don't tell anybody that it'shappened, I'm just like here's.
One of the big ones was all thephone call transcripts from one
client.
I consumed all of that and weanalyzed it and we said what
(12:07):
keywords are we buying andwhat's the gap?
And then what we found out wasno one searches for a particular
keyword, but they alwaysmention that that's an issue on
the phone call.
We should probably talk aboutthis particular product offering
that we have and incorporate it.
So then that wrote a bunch ofLinkedIn content, 800 posts of
(12:27):
problem solutions.
Cause it from the phone calls.
It identified from all of thosephone calls where they
mentioned that product there wasa problem and then how we're
the solution or this client isthe solution.
And then it crawled all of theirblog posts and ranked them and
said which of these blog postsis the best suited to present to
someone to explain how we solvethe problem in this area.
And then did all the publishedall of those.
(12:49):
Here's your LinkedIn post,here's the problem, here's a
solution.
Read more at our blog thattalks about it.
So they had 800 posts to choosefrom, to put it anytime they
wanted to, to pump out.
And they didn't know that wasall AI.
They're just like you must havea bunch of content writers.
I'm like, well, we use someprograms, right.
So I've been shooting down thisAI tunnel rapidly, thinking I'm
(13:13):
way behind and everyone else isway ahead, and I am, in the
last two weeks, just now, comingto the surface of like, hey,
there's probably some cool toolshere that people could use.
Speaker 2 (13:23):
Yeah, exactly.
There is also the question Ihad this discussion earlier
today about AI creating images,creating graphics.
You know creating our posts,and are they too generic?
How do we make sure that theyare relevant for our audience,
that it's not just you knowmaking a sea of sameness when it
(13:45):
comes to what we see from.
Okay, this, you know, from thiscertain category, like all
vacuum cleaners now have theexact same copy, or the, you
know, everybody who's a servicewriter in this area have the
exact same copy, the exact sameimages.
So how do you counter that?
Speaker 3 (14:00):
Yeah, so this is.
There's probably going to be AIexperts that are like that's
not quite how it works, butbasically there is one wise old
man and that is the AI thatconsumed all of the knowledge
and that's the chat GPT that youtalk to.
It has one tone and since mostof the content that it is
consumed has been storytellingand fantasy that kind of stuff
it will always start with in aworld and be that dramatic.
(14:23):
Right For both images and themodels for generative AI.
You just need to take a Lauraon both of them and say I know
you want to talk like this fellabecause that's your default
mode, but if you can just feedit some tokens tokens is the
language that they speak in feedit some tokens to say, when you
(14:44):
talk, just be this differentyou will completely alter the
conversation.
So I have one of my core agentsthat I use and I probably
shouldn't say this until I getsued but it is trained on David
Ogilvie, so it writes just likehigh impact fact hitting.
This is how we write content toget conversions coming through
(15:04):
and it is completely differentthan what you would get Like.
I'll run them both.
I'll run the normal one andthen the David Ogilvie one, and
they're just night and daydifferent.
And then clients sometimes havetheir own authors.
So if they have an author I'lldownload their stuff to talk
like them.
And then I don't really trainoff of this because it's not a
lot of data, but when I publishsomething to like a Google
(15:27):
document for the client toreview, each individual, each
stakeholder will leave commentsand I'll take those comments and
use AI to summarize them andturn them into some type of
context to lean into when wereview.
So it's like what would thisstakeholder think of this piece
of content?
Or alter this content tosatisfy the stakeholder?
Either one right.
So it's almost like I'll writea piece of content, or AI will
(15:50):
write a piece of content andthen I'll get feedback from the
other AIs of the stakeholders tosay this is what they would
think.
And before I present it to theclient I can kind of say, yeah,
I think they're all going tolike it, that kind of thing.
Speaker 2 (16:02):
Yeah, this is far above and beyond, I think, where
most people are right now intheir AI journeys.
So how are you sharing thiswith clients, or what are you
doing and how are you gettingmore clients?
Are you using these sametechniques?
Are you doing any advertising?
Is it all word of mouth at thispoint?
Speaker 3 (16:22):
Nothing.
You're probably the secondperson that even knows that I do
this, even through my clients.
So no, I tried on my websiteand I was like I'm going to
build these agents.
But all of these other systemsstarted coming out where like,
oh, we have these AI agents andyou upload your data and
understand your context and I'mlike, yeah, but read the end
(16:43):
result Even like SEO seems to bethe go-to example for all of
these agents how to crawl thingsand do this.
And when you read the end result, it is like everybody else's,
like it's just what an ai wouldsay.
There's no like we need toreally heavily modify the output
of this to have the a bettertone, like the client has.
So I got a little bitdiscouraged on saying I just
(17:06):
don't want to be another onethat does AI because I actually
download the models and actuallytrain them for the client.
So I just haven't said anythingat all, and even the clients
that I have don't even reallyfully understand what I'm doing.
So there's no testimonials oruse cases of us actually using
AI, right.
Speaker 2 (17:24):
But there are testimonials for the results
they've gotten and the work thatyou're doing for them.
Speaker 3 (17:30):
Yes, but those, oddly enough, are in the form of like
now we had this expectation tomeet of providing this much
content for social media right,and now we're hitting that goal
consistently and our team isn'tstruggling to make it, and all
of the content we're generatingis really thoughtful and
contextual to the environment.
If there's a trade showhappening, I also crawl.
(17:51):
So for every phrase or keywordthat a client has, I crawled a
Google search engine that Ibuilt for them through Google
and I get the top two results orthe top two pages of results.
So I'm always understandingwho's ranking where and what's
happening.
So if news comes through of likea conference or something, the
AI will understand that oh, itlooks like there's a conference
(18:12):
coming and we have content aboutthis, and so its recommendation
would be let's publish thisLinkedIn article versus all
these other ones, because theconference is coming.
And so just to be ahead of thecurve, like to say I'm going to
write about this, is that cool?
And they're like, oh yeah, weforgot that was coming up, like
that's a good idea.
Just to be ahead of the curvefrees up so much resources for
(18:33):
folks to just get their contentout, not some AI.
Let's pretend like we're anauthority at something right.
It's their content what they'regoing to do, but just get it
out in a low friction highlydistributable way.
Speaker 2 (18:54):
So your workflow and your life has changed a lot.
You're also saying somethingthat I think is really important
, which is you know, there's oneschool of thought that you need
to know AI tools because you're, when you're, you know, the
person who knows the toolsversus the person who doesn't
know the tools is going to havethe job.
There are other people say,well, ai is probably going to
replace a lot of these jobscompletely, because the
technology that you're sharingright now, but you had to
(19:17):
program that you had to put allof these together to, in essence
, make sure that all of thesedifferent models are doing
exactly what you want for eachspecific client and for each.
In essence, make sure that allof these different models are
doing exactly what you want foreach specific client and for
each.
You know each goal of eachclient as well.
Speaker 3 (19:31):
Yes, I coded all of that.
It's probably that component ofmy software itself is probably
10,000 lines of code.
I say I coded it, but I haveall like Claude coding it for me
Well take the credit.
Yeah, yeah, take the credit.
But the thing that you have tobe careful of and I think a lot
of people will talk about this,especially as you get into the
more agentic world, where youhave agents talking to other
(19:52):
agents and doing stuff and theagent is calling other agents is
they all still hallucinate alittle bit.
So if they don't know something, they'll make something up.
So I have one of theinstructions.
When I write code is I have atremendous amount of output.
I did this and here's why.
So almost the reason why we didthings, if I can get reasoning
logic out of it, this is whywe're coding this and this is
(20:14):
the response.
So, as a human, I can review itand say that's not that far
fetched.
That's cool.
Fortunately, I'm in advertising, so hallucinations make it more
creative.
But if you were doing somethingsuper complex that required
precision and a deep knowledge,I don't know that AI is the
place for that, but yeah, I'vecoded all of this up sitting on
(20:38):
a server at DigitalOcean,believe it or not.
Yeah.
Speaker 2 (20:42):
So, yeah, agents are here.
Right, I listen to a lot of AIpodcasts and a lot of the
advertisers are like the agentsare coming.
Well, they're clearly here.
It's not a reason utilizingthem to the full effect.
And especially people like me,who I'm not a technical user
right, I'm learning from thebusiness perspective.
I do know a lot of people inthe ai world.
I know a lot of founders ofsome of the smaller tools, so
(21:05):
I'm getting to learn them.
But what else do you thinkpeople who are in marketing
advertising, specifically on thedigital side, what are the most
important skills that they needto have?
Because it sounds like it'sbeyond just knowing how to use
ChatGPT or Cloud or Notebook, lmor Gemini or Copilot.
Speaker 3 (21:25):
It doesn't have to be .
You could use those.
So if you really break it downand take away, like the actual
Laura training that I'm doing,I'm just doing really
sophisticated prompt chaining,right.
So that's even the context.
Like I'm taking big sentenceslike instead of saying I got 15
drops on me today, it's wetoutside, I can just say it's
(21:47):
raining today and that token isvery small and I can put that in
and still have a tokenallowance in the GPT prompt if I
want to for context, right.
So I think a hundred percent,like OpenAI, ChatGPT is moving
to an enterprise model wherethey want to store all your
information.
Like just understandingfundamentally how the AIs work
(22:07):
in the models but then gettingreally good at prompting them.
You can't go wrong.
Any engine you cannot go wrongor any model going and prompting
I think is tremendouslypowerful and Clod is amazing
because it just has so likemillions of tokens you can put
in.
So just dump a graph in and say, tell me about this advertising
and generate a graph.
(22:27):
Right.
Like, definitely go, do that Ahundred percent.
You could get through your dayto day and make yourself worth
10 of you just using these freeand open source tools, Right?
So I would say Claude Lama andopen, like Chad GPT, play in
them equally.
Right Lama is freaking awesome.
I guess Glock too, but inTwitter, but I don't use that
that.
(22:47):
I guess Grok too, but inTwitter, but I don't use that
that much, but I want it.
I have nothing against it.
So definitely keep promptingright, Because how you ask it
doesn't matter if you built yourown custom solution or whatever
.
How you ask these questions isgoing to be critical because
it's going to take the tokens ofthat question and run it
against its token database ifyou will and say how should I
(23:10):
answer this?
So that's critical.
And then just a fundamentalunderstanding of if you go to
chat GPT, you're probably goingto get the voice of chat GPT,
that model, how it trained, likethe grandfatherly voice of I
know everything.
How do you make it sounddifferent?
You can do that with promptingor you can use LORAs for both
the images to make the imagescome out different when you
(23:31):
generate them, as well as thetext.
So just understanding that andhow they get context and
temporal data.
And honestly, you can go askChatGPT like, hey, can you
explain to me LORA and RAG andtemporal data and how that would
all be used to do this, andjust read the answer and know it
, and then keep using those todo your job in media, because
(23:51):
there's still the interfacesthat you have today that are
free for the most part areincredibly powerful.
Incredibly powerful for it, soI'd say, keep going.
Speaker 2 (24:03):
Okay, nice.
Yeah, I will say that most ofmy students know maybe not
everybody listens to thispodcast, but I'm also in school
right now and for some of myclasses I will put in.
Can you explain this concept tome, exactly what you're saying?
Because, to be perfectly blunt,claude explains it a lot more
clearly to me than the videos.
(24:23):
The textbook sometimes theprofessor, sometimes the
professor, you know it breaks itdown so that I can really.
Okay, now I understand how touse Tableau appropriately, or
how to you know, because I justfinished a data and business
analytics class and had to workin Solver and Power Query and
Tableau and like all thesedifferent things to create
(24:45):
graphics and images.
But so I didn't ask you to dothe work for me, but I asked it
just step by step.
Tell me you know I'm gettingthe wrong answer.
What should I do differently?
Can you look at what I put infor my codes?
And so it's been a reallyamazing asset.
I sometimes call it my bestfriend.
Speaker 3 (25:01):
I would challenge you in two areas there.
One, before you ask it toexplain something, give it some
context and maybe things you'refamiliar with Say, I'm used to
Excel, Right, and I know how todo this in Excel, and also I
need to understand this becauseI'm going to teach it to other
people and that willdramatically change its response
(25:24):
out and give you scenarios onhow to help you do that.
And then the second thing wouldbe if it can do it, let it do
it.
Let it write the code.
I haven't written a SQL queryin over a year.
I dumped the database table andI say I need to know this about
this and it will write a SQLquery and I'll paste it in and I
don't even think about itanymore.
Nice, Like it's perfect.
So as a developer who writescode and was like and I have an
(25:47):
advantage in this ecosystem ofadvertising because I can write
code to do things faster, I'mdone writing code.
I'm literally done with it.
I just have it write it, Icompile it.
If it works, I'm like great.
Could it be better?
I don't know, but it works.
So we're done.
It's pretty wild.
Speaker 2 (26:03):
So, justin, what does success look like to you today?
Because I'm sure it's changedalso since you started your
agency.
Speaker 3 (26:10):
Yeah, I don't.
I really don't know.
This is probably the first timewhere I really don't feel like
I have an advantage, becausethis is becoming accessible to
people so rapidly.
So everything that I've beenworking on, in a year's time
everyone will have access to.
So this is almost like the wildwild west the first to deploy
(26:30):
it and be the agency of record,if you will, to help people get
integrated.
They're going to win, but in ayear from now, me to come in and
say, no, use me.
Instead, there will be noadvantage.
That I have right, because AIis the same as AI.
If they already have an AIconsulting agency, I won't win
their business unless they justabsolutely suck, and I doubt
(26:52):
they will.
You know what I mean.
So it's almost like be thefirst one there, but then after
that, once everybody has an AIagent, there's no room for
people to build AI agentsbecause they have AI agents that
will probably build the new AIagent.
So I'm decently terrified thatI won't have a home in the world
in the future and I don't likewhere I would make money.
(27:14):
It would still be inadvertising, but again, like
bots will just see.
I just had a conversation todaywith someone and they were like
we have 3 000 domains in theirprime real estate, domains like
razorbladecom.
That wasn't one of them, but itcould be and I was like that
doesn't matter to me anymore.
Like, yeah, you're going to gettype in traffic, but ideally
what you would do with that isyou would put up razorbladecom,
(27:38):
you would put a model behind it,train it on everything it needs
to know about razor blades andjust have a prompt and say you
could go to chat GPT and it'lltell you what it knows about
razor blades.
But if you come torazorbladecom, this thing will
tell you, up to date and withprecision, everything you wanted
to know about razor blades.
Right, and it's just a promptedwebsite.
(27:58):
It's not even what you justcome ask it.
I just don't know.
I have no idea where I'm goingto be in six months, no idea.
Speaker 2 (28:08):
Let's talk further about this.
The last prompt that I gave tomy students for the first
quarter of the program for DMM510, which is all about
advertising you know dataanalytics.
You know how to measure ethics,privacy, implications.
We got into mad tech, but Ialso like to also incorporate AI
(28:31):
tools into every class.
And the last prompt I said,instead of doing the
participation prompt, read thispress release from a guy who I
was on his podcast.
He's in, I think, sweden.
You know about Web 4.
And here's what he's trying todo and here's what he thinks Web
4 is going to look like.
And it was very puppies andsunshine and butterflies and
rainbows oh my goodness rightlike it's going to solve all the
(28:54):
world's problems, and thisisn't this.
This will happen out of it forme.
I just get excited to thinkabout and I'm sure I can do this
now and you're gonna have toteach me how you know having an
agent of my own, like mypersonal assistant agent, like
go through these emails.
You know, having an agent of myown, like my personal assistant
agent, like go through theseemails.
You know, put them intocategories by X, Y, Z, put them
in this folder.
(29:14):
These emails answer this waybecause I get 20 million guest
requests for my podcasts on adaily basis and you know the
people follow up if I don'trespond right away and I don't
usually have time to respondright away but to yeah, okay,
book this appointment for mebecause I don't have time to
look through their calendar,find time that syncs up with all
(29:36):
of that kind of stuff, and thenalso being able to talk to
other people's agents, as youmentioned.
Speaker 3 (29:42):
Yeah, that's here right now and all that needs to
happen there.
I think you'll start to hearpeople more talking about an API
driven web versus like a dub,dub, dub web.
And that's because if, like,you could go in and tell your AI
agent, your specificpersonalized agent, this is what
I ate today and now I'm feelingthis, like I feel sick.
(30:04):
If you tell it everythingyou're doing, the pattern
recognition will be tremendousand it can tell you.
On a Tuesday, you typically eata lemon today, but every time
you do it you don't feel wellafter.
Maybe don't do that today, andhere's the reason why.
Right, but you have to feedthat data.
But if everything's an API yourcalendar, all of your emails,
your health apps, all of that itcan just consume it and detect
(30:29):
these patterns without youhaving to prompt it.
So you will be the blocker forhaving a great AI agent because
it just won't have enough ofyour data to make really good
decisions for you.
So, yeah, I would say definitely, build your own AI agent and
figure out how to get all of thedata that matters to you into
it and it can have temporalknowledge of, like let's just
(30:51):
review yesterday, what wasyesterday?
Like.
Or you can say let's go backand review all the times I ate a
lemon Is there a pattern thereof me hating that or something
like that?
Right, and it'll just like.
You can just talk to it and saylet's talk about lemons.
I eat them.
Why are they great for me?
I no.
Every time you eat it, you feelhorrible and here's probably
(31:11):
five reasons why.
And here's some substitutes onwhat you should do, game changer
.
But it needs that data, thatcontext about you.
Speaker 2 (31:19):
Yeah, well, and for the lay person right, a person
who's new to AI, maybe they'rein the field of marketing
digital media already, maybethey're not, because I do.
We have a lot of students whocome to the program who already
are working in the field.
We have a lot of students,equally, who you know have done
other careers or adjacent things, and they know that they need
(31:41):
this knowledge.
So what are some tools that youwould recommend for them to
start exploring so they canunderstand agents, understand
how to create their own.
Speaker 3 (31:51):
Oh, my goodness, Honestly, I don't think I would
leave ChatGPT or Claude.
I wouldn't leave to an externalresource because all of these
external resources I've built AItools in the context of how
they want to monetize AI right.
But, you could almost say thatClaude and Gemini and ChatGPT
(32:12):
are like their own, like they'regoing to be honest with you and
tell you the evolutions ofwhat's happening.
So if I were brand like I don'tknow anything about digital
marketing and I want to be adigital marketer, right, I would
first what do I need to knowabout digital marketing?
I would ask that, put that in aGoogle doc right, Start a
folder.
And then I would go researchall of these topics.
So what does the Googleplatform say about media buying
(32:33):
and what are its best practices?
Load it.
Let's summarize that.
What does Google say about it?
Now, let's come up with somephrases and keywords that we can
search in Google and let's gosee what Reddit and Quora, what
people say specifically LikeI've tried that Put.
Say specifically, like I'vetried that, put that in.
And by the time you're done, youkind of have a really good view
(32:54):
of like, well, if I buy aclick-based, maximized click
campaign, this is what Googlesays it does.
This is the experience peopleare having with it.
Like, maybe, if I were to trysomething for a client, this is
what I would try.
And then you ask the cloud orthe GPT, like, hey, I'm thinking
about trying this for thisclient 's a website.
(33:15):
Here's the thing.
What do you think?
Oh yeah, you should totally dothat, or don't do that.
Or here's some other ideas.
You could be a media buyerinstantly and be as
knowledgeable as me.
I'm not even gonna lie like my,the competition for media
buying if you had the confidenceto trust ai enough to say the
most professional media buyer isat my keyboard's end.
I just have to ask it the rightquestions.
You would win.
(33:35):
You would win all day.
Speaker 2 (33:38):
Okay.
Well, going back to the otherside, you still have team
members, so it's not all AIdriven.
You still have copywriters,right?
You still have web designers.
You still have other people.
Speaker 3 (33:47):
Yeah, so I mean that's the other part of it is
like all of this AI was trainedon things that creative people
built, inspiring people built.
Like it's not making up copy,so you still need people to
generate original things.
And the thing about AI movingforward if we talk about like
(34:10):
okay, you can't, like searchengines will go away when people
go to find a solution forsomething, they're going to go
to an AI.
That AI is going to be lookingfor general, like inspiring
responses that people havecreated.
Like I don't know the answer tothis, or I do know the answer,
but I want a very updated orcreative answer.
This person wrote it.
I'm going to feed them in andsay this is what they wrote and
(34:32):
I will describe it.
So we've gone from likeeveryone using AI to write copy
to now AI is going to be.
If it were a human desperate forfresh content, like I'm tired
of reading my own work on theweb, someone give me something
fresh and exciting.
So I think there'll be atransition into that of creators
really having a space.
(34:53):
Now that might be creators thatspecifically copywriters that
write specifically for AI.
It might look different thanwriting for a particular brand,
but, yeah, they're necessary andAI, like I said, has its.
No matter how much you promptit, it will always say the thing
you don't want it to say.
Right, imagine a world or put arocket ship emoji, like it's
(35:16):
always going to just do that andsometimes you just can't tell
it not to.
It just won't listen.
Speaker 2 (35:22):
Yeah, Wow, we've covered so much and I feel like
it's daunting but also exciting.
Speaker 3 (35:29):
Yes, so much and I feel like it's daunting but also
exciting.
Yes, and I would.
I would challenge everyone wholistened to how I explain things
.
Go find two more people thatexplain it as well, because I am
also every day.
This changes.
So, as much as I follow thisstuff technically and read like
all the models and what they'redoing and what they're good at,
like this is hard to like have aday job leveraging AI to do
(35:51):
good things and then alsounderstand it enough to tell
people.
So take the way I said it, butdo your own research and say is
that really how that works?
Hopefully we're close.
Yeah, but, chad, you think I'mpretty smart?
Speaker 2 (36:05):
Well, that's good.
Speaker 3 (36:06):
Yeah.
Speaker 2 (36:07):
Fantastic.
You've given us a lot to thinkabout, a lot to explore.
I really appreciate everythingthat you've shared and prompted
us to do.
Speaker 1 (36:18):
Yeah.
Speaker 2 (36:19):
Yeah, any last words that you want to leave the
audience with?
Speaker 3 (36:24):
I would say, just like, the world needs explorers,
and now we have better tools soeveryone can go explore almost
equally.
Let's just see where it goes.
Keep exploring.
Speaker 2 (36:34):
Great.
Thank you, Justin.
I will have Adelaide Media inthe show notes for anybody who
wants to learn more about Justin.
I do know that we have onestudent in the DMM program who's
interning for you right now.
Speaker 3 (36:49):
Yes.
Speaker 2 (36:50):
So there may be some other opportunities open for
those of you intent in theprogram.
Who actually listen to thepodcast?
Speaker 3 (36:58):
Yes, yes.
Speaker 2 (36:59):
Yeah.
Speaker 3 (36:59):
Fantastic, I will be AI heavy, but we will go through
all the old school media buyingtechniques to get history so
you can have that advantage.
Speaker 2 (37:07):
Fantastic, and we're going to probably do it.
We'll have to do a live demo ora recorded demo at some point
as well.
So, counting on that.
Absolutely Fantastic.
Thank you to everybody wholistens to Mediascape Insights
from Digital Changemakers.
Please leave us a rating orreview, or both on your favorite
platform or as many as you'dlike to, because it really does
(37:30):
help us get more discovered,find more listeners and get more
engagement for our show.
With that, I'll be back againnext week with another amazing
guest.
Speaker 1 (37:40):
To learn more about the Master of Science and
Digital Media Management program, visit us on the web at
dmmuscedu.
Popular Podcasts
Bookmarked by Reese's Book Club
Welcome to Bookmarked by Reese’s Book Club — the podcast where great stories, bold women, and irresistible conversations collide! Hosted by award-winning journalist Danielle Robay, each week new episodes balance thoughtful literary insight with the fervor of buzzy book trends, pop culture and more. Bookmarked brings together celebrities, tastemakers, influencers and authors from Reese's Book Club and beyond to share stories that transcend the page. Pull up a chair. You’re not just listening — you’re part of the conversation.
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com