The evolution and opportunity in Ad-Ops
This week, I’m talking to veteran AdOps pro, Mat Bennett. The conversation starts with AdOps, before taking twists and turns in all sorts of directions. It’s one of my favourite, and insightful, chats so far, and I hope you enjoy it too.
About Mat Bennett
Mat ran technical agencies for 25 years. Having now sold his agency interests he works with other agency leaders as an advisor, coach, and mentor
[00:00:05] Stewart: Hi there and welcome to Scale a podcast for Modern Media. I am your host, Stewart Ritchie, the founder and lead developer at Powered by Coffee. Powered by Coffee is a web and software development team focusing on technology issues facing the media today. Scale is a podcast about how technology impacts the media and how the media impacts technology in return, everything from ad tech and privacy to hosting and content management.
We’re interested in what’s happening today, what’s happening tomorrow, and where we might end up in the future.
[00:00:36] Stewart: today I am delighted to have Matt Bennett with us. Matt is an advisor to technical agencies like us, but he has a background working with ad ops and publishers, not in the game anymore but certainly a wealth of experience and he’s very kindly agreed to talk to us today.
How are you doing, Matt? You having a good day?
[00:00:54] Mat: Yeah, very good, thank you. How are you doing today?
[00:00:56] Stewart: Yeah, I’m alright. It’s always a good day when I get to talk to you, you know, always enjoy
[00:01:00] Mat: it ever so kind, ever kind. Thanks for having me on.
[00:01:03] Stewart: My, my introduction there was a little bit. A little bit vague, not vague intentionally, but do you wanna give us, in your own words, kind of your own background?
Tell us how, how you’re here your own experience as much, as much as you can.
[00:01:15] Mat: Yeah, sure. So I guess I’ve been involved with the web professionally since 96 when myself and my father started what we’d now call a web development company. You know, the, I think the terminology was quite much like the rest of the industry.
The terminology was quite young and unformed back then, so I’ll speak 25 years running agencies. About the last 10 of that, we became very focused on publishers and monetization. So we worked with high traffic websites, helping them make more money out of advertising. That started off as quite a simple thing to do, and over that decade we became, Sort of entirely focused on that.
That was, that was everything we did by the end of it. And obviously the job had become quite a more complex during that time. Finish that last year. I’ve sold up, I’ve got out and I’m doing different things. As you said, I’m working as an advisor to sort of quite a, a range of companies now involved in the workspace.
[00:02:09] Stewart: Great. How are you finding the transition away from agency life?
[00:02:14] Mat: Yeah, it’s not, it’s not the vision of lying on a yacht drinking cocktails that a lot of people think exiting is it’s really interesting now. I think, you know, we are an industry that moves so quickly and then you get very busy doing the thing you are doing and it starts passing by.
And I’m really enjoying the chance to learn about some of the stuff that’s happened while I was busy doing other things or is beginning to happen now. So it is nice to be able to do. More different things rather than just having to do, be very focused on doing one thing well. So I get to indulge my mag part tendencies and, and go after all the shiny things, which is great.
[00:02:55] Stewart: That
sounds delightful. I’m not gonna lie. A little bit of this, little bit of that. See where you kinda end up.
[00:03:00] Mat: There’s too many things though. There’s just so many things I’ve just ended up it’s a standing joke that I’ve got no job, but I’m really busy. Cuz there’s just so many things that are interesting and you can learn about and you just get into these nerdy rabbit holes.
And yeah, it, it’s, it’s lovely to be able to do it and every now and then you’re kind of learning about these things and it turns into a, you know, opportunity to make some money as well, which is always nice. But yeah, it’s, it’s a bit of a luxury, I’ll be honest.
[00:03:25] Stewart: Awesome.
Brass tax then. We’re a podcast talks about media and your, your expertise there is very much in, in advertising advertising.
Obviously a real hot button topic at the moment. You know, viability of ads going forward for all kinds of different reasons, both technical, political and, and kind of everything in between. With real different problems kind of at, at the large scale of publishers and, and at the small scale. I know me and you’ve kinda had some talks about that in the past.
I’d love to love to flesh that out a little bit.
[00:03:55] Mat: I think it’s really interesting ads. I think everyone, let’s face it, no one loves ads. I, you know, I owe a great deal to online advertising, but would I say I love ads? I don’t block ads unless I’m on a site that’s, you know, being particularly offensive in their implementation.
I tend not to block ads. I do believe in, in the value exchange they provide, but I don’t love them. They are unfortunately, a necessity at the moment. Absolutely. We do not have a viable alternative. Yeah. I think there’s some great ideas. But I don’t see that any actually really do support the industry.
So all these problems we’re facing, we’ve got to somehow solve them or come up with a new model. But it’s always a hot button topic, isn’t it? Advertising, because there’s this push pull between, look, we need them at the moment, but nobody wants them.
[00:04:44] Stewart: Yeah,
it’s, it’s a tricky one. Definitely cuz I’m, I’m not anti advertising, but I’m pro privacy, so while I maybe don’t block ads, I maybe will intentionally block things like trackers and that sounds great.
Until you realize you’re still gonna get the ads. They’re just not going to be useful.
[00:05:04] Mat: some would argue they aren’t with the tracking, but
[00:05:07] Stewart: yeah, they’re much more useful with the tracking involved. Yeah. And there’s been a few kind of like high profile, like ad changes and ad issues that have kind of come up in the last week as we, so as we record Twitter is currently melting down.
Not gonna go much more into that other than to say a lot of brands appear to be pulling and advertising from that. But that, that thing kind of actually nicely brings me into one of the things me and you’ve talked about is the seeming movement of ad spend away from publishers and media organizations, particularly small ones that don’t necessarily have huge ad teams onto centralized platforms.
Yeah, like Facebook and Twitter, social,
[00:05:48] Mat: there’s, there’s a number of pressures there, isn’t it? You know, my, my, my love has always been working with independent advertisers, you know, the, the small guys I could, you know, when I got in, I, I moved into the web as a profession. It just absolutely fascinated me from day one.
Just, just, I, you know, this is back in AOL days when I found that one button that suddenly you could go onto the big wide open wild west of the internet rather than, you know, the safe wolf garden of, of aol. And I just loved the, maybe not the chaos, but kind of just the, the creativity out on the webinar.
I’m still absolutely in love with the web today of how many years on it is. And when, particularly not the things that ae came along, you know, I think it was a real moment in kind of democratizing the web. It suddenly gave people the ability to monetize whatever crazy idea they wanted to put out there.
Which I, I think maybe some of the crazies become negative, crazy over the years, but there’s still, you know, if, if your passion. I dunno, air fix models, if you were to put that up on the internet, you can do so and you can cover your costs and maybe make some money. And this is the bit I really love about the web even now, and I think there’s a number of pressures making that really difficult.
One is that shift in end to the platforms. They’re all trying to grab it. You know, certainly, you know, Facebook clearly has been the big one. People talk about Google, but I think at least Google has the mechanism in there to share a proportion of that and kind of enable the publisher side of it.
Facebook very much shut that down. But I think that’s continuing. I think we we’re gonna see it with TikTok. Yeah, there’s some revenue share opportunities in there, but only if you’re working in the platform not great or you know, on the web. And that’s gonna be putting spend You know, apple, apple are doing the same thing.
A lot of spin’s gonna go, you know, into Apple as they expand their ad system. I can’t imagine that’s something they’re gonna become sharing out to smaller publishers and giving the mechanism there. So the shift to platform is a big part of it, but I don’t think it’s, it’s the only one. I think the ad systems have become increasingly complex.
They were a lot harder for small publishers to manage. You know, my company, my old company existed because we, we bridged that gap. We would, you know, we would help smaller publishers work in this more technical space of, you know, earning more money through programmatic advertising. But there’s a tax on that.
You know, they have to pay us a share in order to do it. You know, I had a team, we had overheads, we had technologies that we had to build and, you know, that was all coming out for share. From those. And you are finding that. The overhead even for the middleman companies increases. They’re looking higher up the chains.
Smaller publishers aren’t actually even getting invited to work for companies like that. Yeah. Or more so the smaller guys are getting pushed out. Sure. I think there’s,
[00:08:46] Stewart: sorry Karen, I was gonna, I was gonna gonna come back to one, one thing there. I think it would be grit to get your kind of quick overview on programmatic.
Cuz I feel like a lot of people, me included, don’t necessarily fully understand what it is. Or they have some idea and they’re like, oh actually what is that? It’s a word that gets thrown around a lot I feel, and signs overly fancy for what actually is probably a relatively straightforward thing. But it’s one of these ones that’s like, clouded a lot.
They kind of keep it, oh, we’re not really sure what that is. It’s like algorithmic. It’s like, oh, it’s programmatic. Oh, it’s algorithmic. If you don’t wanna explain it, you just say, oh, it. . Does that make sense? Just cause I know some? No, I
[00:09:26] Mat: think it does. I think it’s, it’s one of those terms that can be used and thrown about in different ways.
I think, you know, any fast moving industry, you get language that maybe doesn’t have to define a meaning. The big part for me when I, the type of programmatic I was working in is there’s advertisers bidding to reach an audience. And, you know, they are setting prices and figuring out on the fly, the systems are doing the work.
We’re not, there’s not two people making a deal. You know, I want to get my brand on your website and I will pay you this price. Instead, they’re, they’re paying an amount to reach an audience and they’ll find that audience where they can, and it’s being run by the platforms in between. And it’s that, that technology layer, which has become increasingly complex and increasingly expensive to run.
And there’s, there’s really one person or one middle man involved. There’s layer upon layer upon layer of it and the, you know, these bid requests, these, these impressions get past all the way along these chain of systems and then all the way back again. And a lot of the time the publishers are having to work with multiple systems and firing them all, you know, firing them all up in parallel.
Cuz it’s the only way you can get closer to a better price of course, than
[00:10:43] Stewart: advertising. So is that something that might also be pre-header bidding and things like that was kind of also one of those things like used to hear about a lot and I think maybe has been wrapped into that term. Is that
[00:10:55] Mat: so, so head of biddings a particular implementation?
So the only theory is a simple version of programmatic advertising might be that we have one partner involved, which generally means it’ll be Google. So a user turns up on the website, We expose the impression to them and say, what will you bid for this impression? They know the content of the site, they know, you know what’s on that page and what’s around it.
But really importantly for the advertiser, they also know something about that audience. So this is where we come into personalization, cookies, privacy. Mm-hmm. . From an advertiser’s point of view, the more they know, the more likely they are to bid strong when that matches their profile. But from a privacy point of view, we might not necessarily want the advertiser to know too much.
So again, one of those, those push pulls. So that would be, you know, a simple idea. And then the advertisers bid to reach, you know, you as an audience member, wherever, whichever website they might find you on, they’re bidding a price to reach you, and then they’ll look for you across, across the ecosystem.
Head of bidding’s, a way of running several auctions like that at once. So rather than allowing one partner like Google to set the price which if you’ve got one very dominant partner, Might not be in the publisher’s favor, just always letting Google set the price of every impression. Head of bidding allows you to run multiple options find the winner of that, and then the normal setup would then be to offer Google the chance to beat the highest impressions.
So you would run, you would put it out to several SSPs, supply side platforms. They would run internal auctions of their advertisers. The winning bid from each of those auctions is returned. You pick the winning one of those and then you put it to Google and say, can you beat this by, you know, a cent or more?
So the idea is you are reaching a wider pull and your using the demand across that to either get a better price than Google would pay or force Google to pay a fairer price depending on how you look at it. And it’s those multiple auctions, which is why. You know, I say, you know, from a privacy point of view, that means whatever information, you know, your impression, your, your presence on that website is being shared with all those systems.
Wides, you know, not great cuz that’s data being shared across, of course a wider thing. But obviously there’s an impact from that in terms of performance for the user as well. In terms of the impact on site speed, it signs
[00:13:21] Stewart: very incredibly complicated. So, I mean, it’s, it’s difficult enough to do like matching on like this user, show them this ad kind of based on like broad categories to then have, you know, in your own, say ad manager be like, okay, well we have eight or 9, 10, 20, a hundred, 120, however many thousand line items say that are competing for that.
If you’re running the most simple version to then have that be running against several possible, you know, ad exchanges to be like, what’s the best you can do here? Then come back and can someone else beat it? It just seems very wistful, .
[00:13:58] Mat: It’s incredibly wasteful. It’s not a system any sane person would design today.
It, it very much came out of, you know, the need, you know, head of bidding in particular. The, the demand for that was that publishers didn’t feel they were getting a fair price from, for instance, Google. Sure. But you know, the old system would be a waterfall. You’d offer it to one exchange. You might put a floor price in place, say, you know, will you buy this, a minimum of this price?
And if you won’t take it, I’ll offer it to someone else with a slightly lower floor. Still awful for performance, and possibly worse for performance but actually not very effective in terms of yield. So, head of bidding’s, just a way of sort of doing that. In parallel. But yeah, it is incredibly wasteful.
We’re running multiple auctions and there’s a lot of people, I think when he first came about the trend became, well, how many, how many partners can I offer it to? And you’d see people putting like 20, 30 bit requests out. I think most people are a little bit more sensible with that because just on a practical basis, there becomes a point where you are not getting any benefit from adding more partners.
And it is slowing things down. It is taking more resource. Of course, you know, it all runs client side as well. You’re making the user bear a lot of that load. Yeah. So it’s, yeah, it’s, it’s not a free option to add more partners in.
[00:15:20] Stewart: And, you know, the ad, some of the ads you get back in these, you know, bidding or better setups and programmatic necessarily, are not necessarily good ads.
Not that they’re not like attractive and like useful ads, but, From a performance perspective, like we’ve had ads come back that are bigger in page than the actual site itself because even so poorly like optimized. So we’re like working a way to get a site to load as quickly as possible, get, you know, in the modern product, get a good core web vitals, and then you’re gonna load 10 megabytes of odd payload for a single ad and you might have six on the page and the whole thing crumbles.
[00:15:59] Mat: Thinking the paths to work with is, is really vital. And, and, and again, this is something that maybe doesn’t help smaller publishers. You know, smaller publishers don’t have the opportunity to be a selective, to some extent, to work with who will accept them. Not because there’s anything wrong with their inventory, but it, it just doesn’t make sense for a big SSP to manage a small account.
Cool. It just doesn’t scale for them. The better SSPs, the supply side platforms, these app partners Are quite good at enforcing, you know, the, the requirements of what can be in a creative. But as you come down the chain, you know, I, I, there’s certainly names I won’t mention, but could who I know that if you run their ads, you’re gonna get, you know, nefarious redirects and popups that are unexpected and like some high ads as well.
And the problem with programmatic is where if you’re booking a direct campaign, you know who you are dealing with and you might have, I dunno, a dozen, maybe you’ve got 50 partners, you’re working with programmatic advertising, you know, on an open auction system you are inviting any advertisement in the world a bit.
Yeah. You know, other than you specifically block. So you can’t do those manual checks. You can’t check every creative before it’s served because on a busy website, there could be hundreds of new advertisers appearing every minute. So you have to trust in the partners that you’re working with.
[00:17:23] Stewart: And I wanna bring that back to something you said. So for programmatic to work really well, to get a good price, the bidder has to know an awful lot about the user that is viewing that content. And that’s obviously a huge, huge privacy concern, both as something that, you know, you’re handing out to, you know, at least to whatever is coordinating your ad bidding.
But probably more likely I to each individual provider to be like, oh, what is this view worth to me? And obvious, not necessarily obviously, but in most cases, smaller publishers aren’t gonna have the CM data. They’re not gonna be able to generate that data and really, nor should they be able to, that’s my data to give.
But for these groups to survive, they have to kind of give in and do the best they can to get. to get that data get the information from the user to pass up. But this gets more complicated as we look at upcoming cookie changes that are coming in the browser. You know, third party cookies versus first party cookies.
Different changes, changes in that. Do you have any thoughts, feelings
[00:18:26] Mat: you can share there? Well, not so many. Right. Where do we start? Okay, I’ll, I’ll, I’ll offer a slightly different angle on the first bit. You said you, you have to have a lot of data in order to get a high bid. Not necessarily. That’s usually the case.
So in order for us to get a high bid for an advertiser, they need to have confidence that that impression has value. Okay. That will often mean they want to know more about the user, both from a for fraud prevention angle. You know, they need to know you are a human. You are where you. Reporting to be from you are, you know, you are there for genuine reasons, not just to clock up impressions and, you know, make money for somebody.
All these things exist, but also just from a targeting perspective. The other way of looking at it is context. You know, if, if only have a very valuable placement you know, if, if I have an ad in a high value place, a transactional point in time, then that would also give advertisers confidence. But then we’re back down to this.
That’s tricky for small publishers. You know, if I am, oh, I don’t know. If I’m money supermarket in the UK and I know I want to put a banner up at the time where people are getting a house insurance quote, then actually the advertisers have got a pretty good idea that a large proportion of the people seeing that ad are gonna be in a particular frame of mind for a small publisher where the advertiser doesn’t know the placements.
That’s when data becomes more valuable. But I think this, you know, all the systems have really become about buying audience. You know, marketers have had this wealth of data for quite a long time. So all the systems have been built around bidding on audience no matter where they are. And the idea of bidding for context, I think has shrunk.
Okay. You know, cookies are the way most of that data is shared. I think cookies have been getting a bad rap beyond what they deserve. The problem isn’t cookies, the problem is privacy. You know, there’s some, there’s some players in the world who are very well served by making the argument about cookies.
Let’s say you were a technology firm who had a very large captive audience who didn’t need cookies because you always logged in. Make it about cookies because all you are doing is strengthening your positioning by damaging everybody else’s. You know, we certainly seen that as you know, Google’s gonna be in a super strong position.
All the big platforms, you know, Apple’s new, you know, Apple’s ever growing out system is, is based entirely on the data. They know they don’t need a third party cookie because your data’s on their system already. So again, it’s, it’s backed down to these, these things become harder for small publishers.
All these things become very pro platform. And, you know, I think when, whenever changes are made, you know, gdpr, a lot of people talked about that being kind of antigo legislation. Look who the big winner is. You know, they, they’re in a perfect position to make more money from it and have done I let’s, I, I think privacy data, great conversations, cookie, a little bit of a red herring.
I think it’s a little bit of a, let’s make people look in one direction while we build lots of possibly worse things in another direction.
[00:21:34] Stewart: Yeah. Yeah. The cookie one is, is a weird one. Cause it’s, it’s such an easy kind of thing to get caught up in. Or what are our cookies? What are our cookies? But cookies are a fundamental part of how the web works.
There are good cookies and bad cookies.
[00:21:48] Mat: And I don’t think the public, oh, I don’t think all the public understand that. You know, I, I have, I have had conversations with people visiting websites, complaining that, you know, complaining that they have to log in every time they visit because they are blocking every cookie and clearing every cookie.
And you, you try to have the conversation of, look, this is the first part of the essential functioning cookie. Yes. But cookies are evil. And I think. Like many arguments in, in the world today, things have boiled down to such simple terms that they become meeting with. Yeah, yeah,
[00:22:24] Stewart: absolutely. That’s great.
I mean, what do you think are the alternatives that are coming? So if it’s getting harder and harder to use, you know, an on device identifier, say to say, to avoid the, avoid saying cookies. Yeah. Say an on device persistent identifier that can be used to watch where you go around the web. What, what are, what are the alternatives that are gonna come up here, do you think?
[00:22:51] Mat: I, I think there are a lot of alternatives, surfacing, and, you know, as I’ve come out of the space, I’ll be honest, I’ve lost interest in tracking the number of them, you know, even Google’s list of alternative approaches is. Just, it’s just getting nauseatingly dull. And all that’s happening is the tech space is just looking for another way to track people in a way that invades privacy without using cookies.
You know, all the unified id, I don’t, I don’t think any of the unified ID solutions really solve the problem they’re just using means other than cookies to do exactly the same thing that everyone’s been criticized for, for a long time. So, I think we need to be having, if we’re, if we’re going to solve privacy and advertising issues in a, in a sensible way, there needs to be a wholesale reinvention of it.
Do I know that if I knew it, I probably wouldn’t share it on a podcast because I’d be raking in quite a lot of investor money about them. You’d be a very rich man. Yeah. There’s things I would like to see. So for instance, you know, if there’s one thing I hate more in advertising, I don’t hate advertising.
If there’s one thing that annoys me more than badly place advertising, it’s badly made pays. Mm-hmm. , I might even say that entire sentence again if that’s cause I feel I had out. If there’s one thing that annoys me more than badly implemented advertising on a side, it’s badly implemented payrolls, I, you know, I just payrolls in general.
I like consuming news from viable sources. I like to get my news from multiple sources. However, I also don’t wanna pay a subscription for every site I consume news from, you know, I really like reading ft.com stories. That’s quite a chunky subscription. Adding New York Times, checking a Guardian, you wanna support them even though this is free, you know, maybe for some balance.
I wanna telegraph, I don’t know. Who knows? Suddenly that’s quite an investment in the amount of news I read. I don’t wanna do that. I would like a kind of a, a snack, a snacking paywall service where I can put so many credits in a month and all these big new services sign up to it and they can just take their share as I use it.
I would love to see that, but I don’t know who’s gonna get the wrong talking and, and, and combine it. I think we need some radically different approaches. I think
[00:25:20] Stewart: that’s like, that approach is, it sounds perfect, but it has been tried so many times over and over again and it never, ever takes off. I mean, we look at today, I suppose the closest, I think at two immediate ones is like, Apple News, you know, apple News Plus where you can have that and you pay 10 or a month or whatever it is and you get access to, you know, however many hundreds of publications.
But are they the ones that you wanted? Yeah, and I mean, I’ve heard that most publications are keeping the real good stuff for their own subscribers because Yeah, they’re, when getting a share of that subscription fee and they’re not getting any data back to be like, right, yeah, we can market to this user based on, on this.
So it’s of super limited value to them. The other way this has gone is the micro transactions, dirty words, which is, I mean, one of the things that, I’m not a big proponent of blockchain and kind of these payments, but it seems like one of the ways where that is most useful or most likely to be successful is you load.
Load up a wallet and there are services that do this and are currently trying to get it off the ground. Use the web payments API to bill your wallet in a particular currency to enable access to that single page. But what is the value of that page? That single view? I think there’s a huge misalignment between what a reader wants to pay for a view and what the publication wants for a view.
You know, I’ve seen publications, alright, we wanna pin a one monitory unit for you to view this page and the reader’s like I was thinking more like five P or a 10 p. And I think then that’s, that’s a much wider, broader question about like, how do, how do we as a society actually value journalism?
Yeah. And that’s a huge, like, fundamental, like what are we doing? Like we are cutting these rooms Cutting journalists like who’s, who’s out there in the UK at least covering what’s happening in, you know, like school elections. School boards, like, oh, so and so is doing this. And like that would be mind numbing work for journalists, but ultimately on that democratic local level that’s super important.
I know there’s a lot of Nest Nesta funding UK research and development Government fund for like media and Democracy a couple of years back. And yeah, it almost all went to that kind of AI for like, how do we cover. And generate content automatically from it and input to reduce the cost of production.
[00:27:46] Mat: I’ve definitely seen solutions where data points like, like core hearings. Yeah. That used to always be a thing covered in papers, really. Who’s gonna pay for a report to sit there? That’s exactly, you know, that’s the sort of thing. GT three could fill that in quite nicely. You definitely want to keep a human in the loop.
I can imagine that going horribly wrong if you just let AI generate whatever it wants about any court case. I certainly wouldn’t want to just let that go on fully automated on my publication. Absolutely.
[00:28:12] Stewart: But I mean sports there’s lots of theater. Tons of theater are in sports and I’ve heard of dunno if it’s GPT three or other AI systems being deployed to write textual summaries of football matches and rugby matches because it’s, if they get it wrong, it’s not the end of the world.
It’s not labelist like it might be in a quirks.
[00:28:29] Mat: I, I did some, I dunno if you are aware of this, I, I did a talk recently about the use of ai. Oh, really? Yeah. So I did this talk of Brighton seo and my talk was mostly sharing some data, some regional data that I collected. And I, I basically, I did three data sets.
So the first part was because I worked with a lot of agencies, I was looking at how agencies already use AI tech systems, and that’s just a survey of agency owners. The second part was user perceptions, where I asked the general public whether they would trust content written by ai. And I’ve tried with different types of content that was quite interesting.
So the, that, the third piece was the most interesting to me, whereas I, I had some text written by human authors and some texts written by ai and I had people. Without any mention of ai, I just asked people to rate, see these articles. I asked them to rate them in a number of ways and never just split test them.
Just did blind split test on AI versus human content. So on the perception thing, it was really interesting. If you ask people, you ask the great British public if they would trust content written by ai, only 18% of people say yes. So overwhelmingly, no they don’t. But then I ask the same thing by different topics.
Sports was the, the piece of the type of content that people were most accepting for AI to write. Which I think I’d have to, I’d have to check the numbers. Maybe we can link to the research or something we can get at that. Yeah, it was, it was still not very high. But you know, rather than 18%, I think sport came in around about 25% of people with something like that.
Things like finance and healthcare. As you would imagine people weren’t so sure about, but the lowest was news. People really didn’t want AI writing news. And if we look at all the conversations around news manipulation and fake news, completely understand why. But the lowest of all was film reviews.
And I found it amazing. But the British public were more likely to trust an AI to give them financial or health advice than to what they would watch on Netflix that evening. I think it’s something around we, we, we think these systems can’t be creative or can’t understand those types of things. I dunno.
But yeah, I, I just found it all fa absolutely fascinating. Completely nerd it out on this data. That’s
[00:31:00] Stewart: crazy. I’m sorry, trying to wrap my head around why you would trust it. For finance over, I suppose finance. You look at it and go like, this is numbers, put number into computer, computer, come out with yes reasons.
[00:31:14] Mat: That’s how I see it. Yeah,
[00:31:16] Stewart: films, I suppose it’s more like, well I like this and I like that. How could the computer possibly know what I will like in a film and not that’s very
[00:31:25] Mat: strange. And then, then when we put into that third piece of data where we turn it, the blind test, the most interesting part of course is what people say they will trust is completely different to what they do actually trust when they don’t know about it.
So I think a lot of it is, as you say, it’s about perceptions. Maybe we are getting used to doing idea of AI writing sports information and financial information, but the idea of it recommending a film for us is just a little bit out there at the moment. So these things might all change. Yeah, that’s
[00:31:55] Stewart: strange.
I suppose one of the things. Is automated recommendations are so difficult. Like if you’ve ever used Yelp or Foursquare back in the day, what it guessed you might like through whatever rudimentary machine learning they were using was never, never good. Yeah. I wonder if there’s still a holdover from, from that.
If people like, oh, it’s a recommendation. I can’t,
[00:32:17] Mat: I suppose you’ll see it now. I dunno what your Netflix recommendations look like, but I, I, I feel like, like a lot of platforms like Netflix will just recommend the thing that everyone else is watching and I watch some weird stuff and it’s really not very good at recommending what I wanna watch.
So maybe, maybe it’s that maybe people, as you say, people see evidence of the engines that are being used now and assume that’s the best they can do, whereas actually they are probably doing a very good job, but they’re not there to help us. They’re there to help the platforms that are using them.
[00:32:52] Stewart: So you gotta think like, , these things are fairly new to the world and kind of what they’re doing.
And these technologies generally follow, I can’t remember the specific name of a curve, but like an S curve where they start fairly slowly, rapidly accelerate and then te off. And I think the thing that worries me is I don’t know where we are on that curve. Like is this just the start? And are they gonna get a lot better at pretending to be human, if not accurate?
And what is the metric Is, is the metric that’s being judged? Like how can a person tell this is AI or is it actually correct? Yeah.
[00:33:27] Mat: So, and okay, so you, you’ve heard right at the beginning of this, this podcast, we, we said about me following shiny things and indulging my mag pins, and completely by chance we touched on it.
So this is, this is the big shiny thing that I, I kind of lost six months of my life diving into, which makes me an absolute beginner. I’m not, Preparing to be an expert in I, just a a, a slightly better informed absolute beginner. I think the speed of change in the AI space, AI is not ai, but in, in the, the content generation space, we’ll call it the speed of change in there is just unlike anything I can remember.
It is so, so fast. The example I gave is most of the AI techs we see today is being produced by an algorithm called GT three, which as you can imagine’s the third version of, of this, of this system. No one talked about GPT two cuz it was a bit crap. In terms of the leap of capability, the kind of the how, how powerful these systems were.
And again, I’m trying to remember the numbers from, from the targeted, it was something like five or 600 times. You know what, can I pull up a slide and give you some sense why we, cause I just kind of, if we’re getting into day trial, I’d like to just get it right.
[00:34:50] Stewart: If you’re happy to share it as well, we’ll throw it in the show notes.
[00:34:55] Mat: You know what I, I did, I went so deep into this and I spent thousands on the bloody research. I don’t even know why now. I just, I get a bit of a beer in my bonnet, so it’s quite nice to get some more use out there. Great. Let me pull up my slides. I’ll share, I’ll share that, I’ll share the deck with you if you’re,
[00:35:11] Stewart: if you’re interested in the progressiveness.
I mean, I remember. So we’ve got do, do too, I should say doing the rhymes at the moment as kind of one of these image generators. I vaguely remember Dolly one and it being a bit of like a wet fart where everyone’s like, oh, that’s really in. . And even before that, the Google version of it, which I wanna say was called Deep Mind, but I can’t remember entirely.
But it would create, you would give it a prompt and it would create these nightmare images. Do you remember for a few years ago,
[00:35:41] Mat: wasn’t there? I can’t remember actually. We’re getting really distracted. Where they, they going, where would they go with this? I’ll show
[00:35:49] Stewart: you. It’s fine. It’s, my goal with these things is like, see, we’ll start at one place and we’ll see where we end up.
Oh, here we go.
[00:35:55] Mat: I ended up doing all the images using mid Journey. Oh, great. Well, just because again, I’ve just nerding out. So this is that thing I said about oh no, that, that wasn’t that one. I will come back to the point. So that though, that was the blind test on, where’s that one with the news perceptions?
There we go. That was the Would you, would you trust. Would you trust content? You know, so website content, 18% of the public said they would trust it. Film reviews. 6%. Yeah. Honestly, the world’s mad, but this was the one I was looking for here. So, yes. Speed of change. Let me, let me talk about that with the numbers in front of me.
So, so no one, no one talked about GT two because, you know, it was clever, but in terms of the output, no one was, no one was buying it. It, it was a bit crap when GT three came along. Suddenly there was this step change in the ability of these systems and people started getting really excited and talking about it.
So we measure the power, these systems in parameter account. So the jump from GT two to GT three was GT three was 177 times more powerful and it was that big jump that suddenly got everyone excited with G P T four is due for release, like imminently, like it could be even before you end, I suspect.
Not the way these things go. But GT four is going to be 571 times more powerful than what we have to be GT three. So we’ve got another jump coming, but the jump is kind of three times larger than the previous jump. So speed changes. Incredible.
[00:37:30] Stewart: Yeah. So what are these numbers? 1.5 billion, 175 billion on a hundred trillion.
175 billion what to kind.
[00:37:38] Mat: So this is parameter count, kind of like the number of in, in my simplistic terms, the number of variables it can work through while it’s figuring out what to say next. Right. I, I look at these systems as like, text prediction out algorithms. Really? Yeah. Rather than being ais.
So you imagine like how many options they have to kind of work with. Of course. Sure. Someone more technical than me will now put in a comment that I’m completely wrong, but that’s my simplistic take on how the parameter account of these systems turns into the quality of the output. So yeah, that speed of change, no, absolutely.
I think we are literally just at the start of this. And we can see it in, you know, the other bit is I find really interesting is the image generation thing. So, And what we’re seeing, we’re seeing it in there. We were talking about, you know, Dolly two. If we look at the examples of what’s now happening with video being generated in the same way, it’s, the complexity of that is incredible.
You know, mind boggles what’s gonna happen with deep faith? Well
[00:38:43] Stewart: this is the, the video thing is actually like, what gives me pause to be like, where on this curve are we? Because whenever do two started doing the rounds along with stable diffusion and I can’t remember what the others are called everyone’s like, well that’s, that’s fine.
That’s just a static image. They’re not gonna be able to generate, you know, video. And that’s where most content is. And then it was a timescale measured in weeks before Meta had put out their video generation prompt via system and someone else had at least won like very quickly after that as well.
And that’s the thing I’m like, and there there are sws of YouTube videos and videos that are. AI generated where the ai like GT three or something has written a script, some other machine learning process has taken and mapped that to effectively mocap run that against the model like a 3D model, and then recorded that output as the fs.
And then you mentioned defects, layer it on top of that there’s a whole, there are content
[00:39:42] Mat: farms. It’s fascinating and frightening. I, I’ve, I’ve got a real car crash fascination with it. Yeah. You know, I, I really do. The, you know, I, I got interested because I was, I was seeing these systems used to produce crack website copy at scale, so just people producing.
Vast scale very quickly and trying to monetize it for our services. You know, that’s literally what, what hooked my interest in it. You know, there’s another trend we can talk about in the advertising. Yeah. Well, why, why isn’t there trust in the advertising ecosystem? Well, here’s another reason not to.
But yeah, the, the, the whole, the whole thing Id, I find fascinating and frightening in, in probably equal measures. Yeah.
[00:40:27] Stewart: It’s, it’s a, it’s a scary thing. I know there’s a lot of like, panic, I suppose, in like the creative space of like things like the image generators replacing designers and artists, and I’m like, yeah, maybe.
But photography didn’t replace entirely portraiture. It invented a new thing that happened. Sure, there’s less portraiture than there would’ve otherwise been, but it made that access to more people and I hope. This will be a similar thing, that it will create jumping off points and stuff like that. But general purpose ai or the large skill deployment of it for misinformation is a worry and I think a very scary thing
[00:41:11] Mat: to me at least.
Could this, it’s whole circle, isn’t it? This is why we need to support proper journalism. You know, if people are going to get their news through social media and we know social media’s gonna be influenced, and if we suddenly go, well, okay, rather than a troll farm tucked away somewhere with people banging out messages, why don’t we plug that into a system that’s gonna generate those messages and maybe put some nice convincing photos alongside them and maybe have a news reel that’s been generated to support that, that are people gonna believe the scale of disinformation that can come out?
Is, is just going to be frightening. And, you know, trusted news is trusted news sources. Is is the way to count that? I’m not sure. I’m not sure that journalism is doing particularly well at the moment. I feel, I don’t know, maybe I, we all have our own biases. Maybe it’s mine. I feel like, you know, independent news is becoming less of a thing.
I must not all down to funding. But it’s, it’s gonna be vitally important. Otherwise, you know, what do we do? We’re not gonna give up on truth hopefully, but.
[00:42:21] Stewart: Yeah, some would say maybe we already have .
[00:42:24] Mat: I think some people definitely have, haven’t post True era, aren’t we? Yeah, right.
[00:42:29] Stewart: I, I blame Star Wars cause now , oh,
[00:42:33] Mat: sorry.
That was unexpected.
[00:42:35] Stewart: It’s like every time, every time I hear it, like the fake news thing come up in my head, I just have like one Nobi or whoever said it in the film go from a certain point of view and I’m like, oh, that’s just like, that’s it. Like that’s now a thing. But to bring, I suppose, bring it back full circle.
Do you feel like we’re gonna see a lot of AI applied to advertising? Both in, I’m gonna come back from AI and say they’re machine learning, you know, generating hundreds of thousands of variations of an advert to run at the lowest possible quality mark all the way through to guessing and algorithmically picking.
Interest groups for a user before they are there themselves. I think I’ve seen a little bit of this kind of, if you think of TikTok a lot of people have gotten put down the path of having diagnoses for things like ADHD and autism because saw them and put them into groups Yeah. For recommendations before they knew it was a thing.
And I wonder if we’re gonna start seeing that happen in some of the ads if we really get that kind of machine learning applied or are we pass that
[00:43:47] Mat: machine learnings becoming so accessible, we we’re gonna see it applied to everything. Yeah. And at scale the bits that stick will be the bits that turn into money for people.
You know, in terms of generating text for text ads, that’s something that, you know, GT three is already being used for. I don’t think the aim, it might, it might be the outcome, but the aim generally isn’t to produce the lowest possible quality. It’s to get the highest possible result. Yeah. But yeah, if you take a human out the loop, then, you know, who knows, maybe the weirdest things, get the highest clickthrough rates or, or whatever.
But there’s some really clever systems out there doing that already. And likewise, the machine learning and looking at the results. So, you know, partly for the, you know, generating the text, but looking at the results and then optimizing, you know, great stuff for machine learning. But I think we’re gonna see the same certainly for images, you know, in, in taking the graphic creatives of an ad and at least producing variants.
You know, I think the, the image ais that we have at the moment are largely novelty. Yeah. They’re really good fun. You can play around, you can laugh when they mess it up and be vaguely impressed when they get it right. But they’re not really directable. If I ask, if I ask any of them to produce a series of images on the topic or variations of this, they’re not so good.
I think Dolly two is moving a bit more in that way. With the, in painting you can change kind of elements, but when we start to be able to, when the interfaces, not just the capabilities of the system, but when the interfaces improve to suit the developing use cases, we’re definitely gonna see the same thing applied to our creators.
So then we’ve got a point where, you know what, we can generate variations on copy and variations on imagery, and we can do this at scale. We can test and see what works. We can put the results into more machine learning and get the perfect ad, but perfect being defined by whatever we defined in the beginning of it.
And I think that becomes, you know, that becomes the challenge, doesn’t it? Setting the rights. The right metrics to optimize to. But yeah, we’re gonna, it’s gonna be everywhere, isn’t it?
[00:45:57] Stewart: Just keep an eye on the time. This has gone everywhere. I’m delighted. I love it. Thank you very much. I, if someone wants to find out more about you where, where can they go?
Where can they find out?
[00:46:07] Mat: Best face is either find me on LinkedIn, Matt with one t Bennett Bt, firstname.lastname@example.org.
[00:46:15] Stewart: Great. Thank you so much.
[00:46:18] Stewart: Thank you for listening. If you enjoyed this episode, please subscribe. The skill is available in all the usual podcast places. Even better, if you could leave us a review that really helps us.
If you’re interested in finding out more about me or Power by coffee, you can find us on social media and again, in all the usual places, links are in the show notes. Scale is currently gonna kind of come out every two weeks and we will see you then.
Getting started with SEO for Publishers
A modern media podcast
hosted by Stewart Ritchie
This week, I’m talking to veteran AdOps pro, Mat Bennett. The conversation starts with AdOps, before taking twists and turns in all sorts of directions. It’s one of my favourite, and insightful, chats so far, and I hope you enjoy it too.