Transcription
Andromeda, Gem, and Lattice were the biggest Facebook algorithm changes that I have ever seen. Today, I sat down for the last 25 minutes with a Meta executive who was responsible for these massive algorithm changes. So, in this video and interview, I was able to ask my most personal questions on how the algorithm has changed, how we should be thinking about it from Meta's perspective, not just what's in their white papers, not just what they tell us publicly, but a true behind-the-scenes understanding of the internal people who have actually built the algorithm themselves. Matt Steiner, welcome in. How are you?
>> I'm doing well, thank you. Great to be here. I appreciate you having me on the show.
>> There's a lot of questions that I want to ask you today. You have a pretty major role on the Meta side of things and specifically in the realm of known things like Andromeda and the algorithm in general, which has a million questions. I could have asked so much to you today, but what I really wanted to focus on was a group of questions that I think are going to be most impactful to selfishly some of the brands that I work with and also ideally a lot of people that are watching this right now. There's a lot of confusion on the algorithm. There's a lot of confusion on what goes on. Can you just tell the world what you've done on your side on the Meta end, what you do, what you're responsible for, cuz I think people being able to hear the answers from the source is is very rare and in your case going to be pretty unique.
>> Great. Yeah, thanks. Happy to. So, my name is Matt Steiner and I support monetization infrastructure, ranking, and foundational AI at Meta. This is effectively the set of systems that are retrieving which ads can be ranked for a person that shows up and then ranking those ads, trying to pick the best ad that will both maximize the value for advertising partners as well as deliver the best experience for the people who are seeing the ad, really trying to match people to their interests and products and services that they will love.
>> Which is a good segue to say this is what people have known to be this is Andromeda, right? In a way. So, I want to actually ask you that question. What is Andromeda? Why was it built? And how is it impacting advertisers so far from what you've seen?
>> Yeah, it's a great question. So, when you think about ranking for advertising, the first step is to select the ads that are candidates to be shown for the person who's appeared on a meta property and there's an ad slot to fill. That is where Andromeda comes in. Andromeda is our retrieval system that's been co-designed, the hardware, the software, and the machine learning model to optimize performance for retrieval and specifically personalized retrieval. With the rise of generative AI tooling, advertisers have created a lot more creatives, whether that's creative backgrounds or creative copy and mixing them all together in large number of variations. That kind of Cambrian explosion of creatives made that retrieval problem much more difficult for our advertising system. There were just so many more ads to consider when determining which ones to rank. So, we built Andromeda, which is a much more capable hardware system powered by Nvidia GPUs on a custom hardware skew, plus a machine learning model that is tailored to work really well on that hardware system, and the hardware system was selected to work really well with the model. So, we can maximize the performance in personalizing which ads to show for you. Cuz if you think about ad retrieval, a bunch of advertisers are going to try to target me with an ad, but not every advertiser that tries to target me with an ad has a product that I'm interested in. My interests are a subset of the advertisers who've targeted me with their products. And so, Andromeda is really the sophisticated hardware and machine learning model that selects of the advertisers who've targeted products at me, which ones am I most likely to be interested in. And then those progress further to the ranking stage, which is where you hear about our investments in the generative ads ranking model Gem and the adaptive ranking model which we just started discussing a couple weeks ago.
>> Yeah, cool. I want to hear a bit more about Gem as well, but I have a question on on the Andromeda side or maybe even like a summation of things as keeping it as simple as possible, right? Like people started making more ads, a lot more ads, right? People went down the route of essentially saying, "Hey, we have something that works. Let's go make 20 versions of it." Not old school. I remember the days you'd have one ad and you'd just smash it. You'd spend as much money as possible against it. And I think that that shifted naturally. There's reasons for that, right? We see less fatigue on on ads when we do that and we see more scale when we actually have more variations of ads. I think there's a misconception. I know you're not totally tuned in on the creative side of but I do think there is a misconception that now in this Andromeda era that building iterations and and building versions of good ads doesn't work. And what I hear from what you've just said, which is perfectly described of you know, what I've read and and understood is that it's not about you can't build variations, it's just that they're going to literally be retrieved and perceived by the Meta algorithm in a different way.
>> That's a good description. I think the way the Andromeda system would look at this is if you create 25 versions of an ad, maybe only three of those would be appealing to me potentially and a bunch of them may not be appealing to me because of the types of content that I look at, because of the backgrounds that it appears on, whether the color matches, the specific context that the ad appears in. So, a particular subset of those is going to be relevant to me. Not all of the creatives that you've created variations for may be relevant to me. And Andromeda is going to retrieve the ones that are most likely to perform for me. And some of the ones that will perform for me may not perform for you and some of the ones that perform for you may not perform for me. Um so, as a result, Andromeda is selecting the most likely candidates. It's good to have a reasonable candidate set of ads creatives for Andromeda to consider across your various audiences. at the same time. Creating a an extremely large number of them probably isn't going to help you more than having a reasonable diversity of ads.
>> That is the part that I'm so happy you just mentioned, right? Like creative diversity doesn't mean that you have like all these like crazy different ads all over the place. Some that are, you know, built for for one avatar and some built for a totally different avatar that has nothing to do with your brand. And I think people get that so confused these days. And what you just described is like, "Hey, you could build more. You should probably build more." I won't put words in your mouth on that, but from my perspective, you should probably continue to build more diverse ads or clusters of diverse ads. And even if you build 20 and three get run, three get delivery, that's almost to be expected here because a lot of the work that would have been done by spending is now potentially saved and done by the loads of computer rooms
>> [laughter]
>> that Meta probably has somewhere that's doing a lot of that processing up front, right?
>> Yeah, yeah, that's exactly right. The machine learning models, they're very good at figuring out what works and what doesn't work and optimizing for what is working, right? So, they'll attempt delivery on all the variations of the creative, then they'll figure out which ones are working best, and they'll start moving the advertising dollars to the ones that perform best for you. And that doesn't necessarily mean that the other creatives were bad. It's just that they were lower performant, and as a result, you probably don't want your dollars spent on those lower performing creatives. The machine is doing the optimization, and it's doing the optimization based on real-world performance data. And that's what really the machine learning models are best at, and that's why we have all of these large and powerful computers that were very fairly expensive sitting in those server rooms.
>> I'm sure. How does Gem and and Lattice, which are two other products or or whatever we call them, services, products that that Meta has developed over really 2025 and into 26, how are they playing in this Andromeda era?
>> That's a great question. So, uh Lattice is effectively our model consolidation technology. Previously, you would have a variation of a machine learning model that was extremely specific. This model is predicting clicks on Facebook. This model is predicting clicks on Instagram. This model is predicting conversions on Facebook. This model is predicting conversions on Instagram. It turns out that if you take all of those machine learning models and you collapse them into one model all of the training data that used to go into those four models into the one model, the performance of the one model is better than the performance of each of the individual models because that larger model is able to learn from more data, more of your interests, more of your interactions with businesses, products, services, et cetera. And that just improves the performance for the advertiser as well as the experience for the person who is seeing an ad. You're going to see a more relevant ad because we consolidated these models and the training data used on these models to drive better outcomes for advertisers. So, that's Lattice. Gem is a quite a bit further expression of that where we're creating this large generative recommendation model that is really our foundation model that all of our other models kind of learn from in that kind of second stage ranking environment. And then we try to take that super large foundation model, the generative ads recommendation model, and we build a LLM scale and complexity inference model out of that foundation model and that's what we started calling the adaptive ranking model where some number of people interact with brands and content a lot, have a lot of interest, they interact a lot on the platform, they purchase more than other folks, they have longer sequences of interactions, we can learn more from about them and make better recommendations for them from their more more interactions on the platform. Other folks, they don't interact as much as brands, they don't purchase as much. They have shorter interaction sequences, interaction histories. So we don't need to use as much compute in recommending an ad for them because there's just not as much input data. And so adaptive ranking model allows us to dramatically scale up on the kind of power user segment and produce substantial improvements in performance for those users who are engaging a lot with brands and content etc.
>> Woah, okay. Super good explanation. I've not heard that explained. So I need things sometimes in the simple version and and Matt that was the simple version of something that's extremely complex and I really do appreciate it cuz that to me is like simply put that is tying together the mass adjustments that Andromeda as this this era that we're in. It's kind of like I don't want to say like the sidekicks but like they don't they all work together. They they have to work together. Which is really cool. I want to ask you a question around what my audience has observed, what we've observed on a lot of the brands that we work with for context, right? We work with around 100 brands right now. We're spending millions of dollars every month across the Meta suite. And then we also have a lot of folks that watch in different modes that essentially like absorb a lot of what's going on and they see they're feeling it and not necessarily spending hundreds of thousands of dollars a month, maybe spending a couple thousand dollars, 5,000 or whatever. So something I've heard more is that brands are reporting increased delivery volatility. So just overall volatility of ads. Big shifts in in where the spend goes. Sometimes where they feel like campaigns are destabilizing especially without obvious triggers. Like they're not seeing like hey I changed something on site that day. They can't figure out what changed in that period of time. So my question is this known? Is this the trade-off that we just have to take for faster optimization cycles where volatility is actually the feature potentially and not the bug.
>> Yeah, this is really great question, and we have heard the same question from a number of our advertising partners, as well. And it's really important one that goes to how the auction actually works. So, because we are running a very large number of auctions in real time every second, and you have a very large number of bidders who are bidding every second, the more advertisers on the platform, and the more frequently they change their bid prices or bid volume, the more changes that everyone sees in the auction dynamics that play out in the marketplace. It's like the stock market. There are a bunch of different tickers, a bunch of people, retail investors, institutional investors, they're bidding up the prices of the stocks or down the prices of the stocks in real time as they think those stocks are more or less valuable to those investors. The same effect is playing out in the advertising auction. There are a bunch of people, and maybe some information comes in that this customer is much more likely to be interested in the product or service you're offering, maybe you increase your bid because they would purchase from you if they saw your ad. Or maybe some new information comes in, and they just bought a refrigerator, they're unlikely to buy a new refrigerator soon, and that customer is no longer as valuable to you, and you change your bid. So, as better information propagates throughout the ecosystem, more advertisers are changing their bids more frequently based on what they learn about customers and their desires. So, those bid changes at volume create the volatility that people experience on their campaigns cuz it's a dynamic marketplace that is optimized for delivering the best price for each advertiser. And so, as you change your bid, different people and conversions will appear to you, not just because of the changes that you've made, but because of the changes that everybody else is making in response. Uh so, volatility and uh associated price optimization are the opposite end of the spectrum of low volatility and higher prices. So, you kind of have to decide which matters more to you, lowest prices or um low volatility, cuz they're on opposite ends of the spectrum.
>> Uh which is like essentially the more you're spending and the more conversions you're driving, the less overall volatility you will have because in a very simple term, the algorithm will be more confident in the decisions that it's putting in front of you. And then one or two or 100 people or 1,000 people that may have been considered to purchase that drop out of your funnel because there's a signal like they bought that refrigerator, that's not impacting your total funnel. When you're spending less or when your average order values or cost per acquisitions are significantly higher, you then therefore have to just accept that you have to deal with volatility because the algorithm isn't just going to endlessly keep spending on people who are no longer actually a part of your your true purchase funnel anymore.
>> Yep. Yep. Yep. Um as we are now again in this like Andromeda era that everybody talks about and we see people get more comfortable with how advertising is supposed to work and basically how to treat the algorithm now. I would love to understand a bit more on Meta's long-term vision of the advertiser's role. Is there, you know, an end state where we just hey, here's my 10 creatives, here's my 100 creatives, go rip it, Meta. Like what do we do? I'd love to understand just like POV on that. And I know there's no perfect vision, but an idea would be nice.
>> Yeah, I mean it's a great question. I think if you look at the trends of the industry, the machine learning models and the computers, they're just getting better at what computers do. They compute things. They compute things quickly. They can make uh dynamic decisions based on really good data really quickly. And I think we're entering an area era where humans are never going to outperform the machines at that type of work again. Humans are never going to out optimize the machine. They're never going to move the budget as quickly, as responsively to price changes up or down uh kind of based on the conditions of the auction market on Meta or really anywhere else, right? But what are humans really great at? Humans are great at coming up with novel ideas to test new messages for your brand, new messages for a particular audience, and then looking at the performance data and reasoning about why did this message work with this audience, and why did this message not work with this audience. The computers are not going to be able to reason about that with human-level intuition. So, humans are really great with coming up with the ideas, and they're really great at connecting the performance ideas with hypotheses about why a particular message worked for an audience, resonated with that audience, drove conversions with the audience, and why a particular message may not have. And then, once you have some hypotheses from looking at the real-world performance data, you can create variations on your message, on your audience, and try again. Right? So, really this is a long-term human and machine partnership where the humans are doing the ideating about the initial messaging, and then reasoning about the performance data with human intuition and insights, and then starting the process over again. While the machine handles all of the computational optimization that they're so good at.
>> I'm going to start using computational optimization in my my daily language. Can we expect massive changes like this, um, or do we generally feel like, "Hey, this this big new era that we're in in the algorithm, this is here to stay." I know some brands feel like they don't have enough time to actually adjust before the next big thing actually rolls around.
>> This is a really interesting question, and and I think if you look at the broader trends in the computing industry, algorithmic advances are moving faster than ever before.
>> Yep.
>> The pace of AI innovation is absolutely astoundingly rapid. New papers are coming out literally every day about breakthroughs in specific components of artificial intelligence and machine learning models, and how that drives performance in real-world applications. We, of course, are focused on delivering the best results for our customers and the best, uh, kind of experiences for people that see the ads. And so, we're trying to incorporate as many of these technological breakthroughs as we can, as quickly as we can, because that produces better results for our customers. So, I think it's safe to say that the pace of innovation will continue to increase, will continue to drive as much innovation as possible into real-world performance benefits for advertising partners and for people that see ads. And that's going to come with an associated need for advertisers to understand what the opportunities are with these new algorithmic advances and system changes and how to really take advantage of those to get the best return on your advertising spend and the results for your business. I wish I had some better news that things will slow down, but I don't think they will. I think they'll continue to accelerate and hopefully that produces massively better results for your businesses, but it will come with a cost of needing to keep up with all of the changes that are happening in the world.
>> Yeah, um probably good news for me because I'll try to keep updated on everything and keep folks updated on on everything. Um but it is challenging, right? There's no doubt about that. And I think just like a reminder to everyone watching this is that from my perspective, I I won't speak for you, Matt, but from my perspective, all advertising platforms, Meta, and anywhere else you're advertising, the objective is quite simple, I think. I think Meta's job is to get you to want to spend as much money as possible. And that want is the keyword. You want to spend as much money as possible when you're getting the business result as closely as possible to your target or exceeding, right? So, any change, like I would be stunned if Meta ever made an algorithmic change or a major sweeping change just because it feels good or to try to move things forward. Like everything is rigorously tested. Everything is meant to literally put more dollars or confirm more objectives for whatever the advertisers are trying to do, right? So,
>> Yeah, that's exactly right. Every change is AB tested and we only launch things that move our our metrics in a positive direction.
>> Mhm.
>> Cool. Matt, thank you very much for a rapid-fire of five or six questions here. I do appreciate what you do. I'm loving the Andromeda era right now. I'll admit it was a little tough to get used to a few things on some smaller brands that we work with, but um you know, for those that that we've battled through, it's it's nice now. We we have what I think and what many advertisers out there have experienced now are more comfortable in in what's working and how to approach things. So, yeah, keep rocking on. Thanks for helping us make more money.
>> Great. Thanks thanks for your time. Great chatting with you, Sam, and have a great day.
>> Thanks, Matt.