Today in Digital Marketing - BONUS: AI and the Future of Meta Performance Marketing
Episode Date: June 22, 2023Matt Steiner, Meta's VP of Engineering, talks about generative art, chatbots and how Meta views the future of advertising in a machine-learned world.-Each day this week, we are putting an extra ep...isode in your feed — earlier this month Meta recently held its Performance Marketing Summit. Since Meta is such a big part of many marketers' spend, we are replaying some of the most relevant presentations from the summit.This is not a paid placement — Meta hasn't paid for this, and didn't ask us to do this. Also, of course, these are Meta reps at a Meta conference, so it's pretty heavily promotional, sometimes comically so. That said, there are some pretty important things discussed in these sessions — like their take on AI modeling, how they see the future of creative, and a bunch more.If this isn't your jam, you can just delete these — our regular daily newscasts will continue to come your way.Our Sponsors:* Check out Kinsta: https://kinsta.comPrivacy & Opt-Out: https://redcircle.com/privacy
Transcript
Discussion (0)
Each day this week, we are putting an extra episode in your feed.
Earlier this month, Meta held its Performance Marketing Summit.
Since Meta is such a big part of many marketers' spend,
we are replaying some of the most relevant presentations from the summit.
I should note, this is not a paid placement.
Meta hasn't paid for this. They didn't ask us to do this.
Also, of course, these are Meta reps at a Meta conference,
so it's pretty heavily promotional,
sometimes comically so. That said, there's some pretty important things discussed in these
sessions, like their take on AI modeling, how they see the future of creative, and a bunch more.
If this isn't your jam, you can just delete these. Our regular daily newscasts will continue to come
your way. Do you have business insurance? If not, how would you pay to recover from a cyber
attack, fire damage, theft, or a lawsuit? No business or profession is risk-free. Without
insurance, your assets are at risk from major financial losses, data breaches, and natural
disasters. Get customized coverage today starting at $19 per month at zensurance.com. Be protected.
Be Zen.
Today's episode is from a session they did called
AI and the Future of Performance Marketing.
Matt Steiner, the VP of Engineering,
talks about generative art, chatbots,
and how Meta views the future of advertising
in a machine-learned world.
I've been at Meta since 2014
and have been working on performance in AI since 2016.
So I've been a part of several evolutions of our ads, product, and technology stack.
Today I'm going to share with you some of our investments in AI and how we're building
the future of AI-fueled advertising at Meta.
Let's talk technical. The business that the world that
businesses operate in today is substantially different from the world
of a few years ago. As Simon mentioned at the start of the day, artificial
intelligence is rocking the technology industry and the implications are
rippling through the business world. And as Kristen mentioned, AI isn't new at Meta.
It's powered all of our apps and services
since the earliest days of Facebook News Feed in 2006.
Our ads ranking system has used machine learning and AI
for over a decade to show the right ad to the right people
at the right time.
Each year, we invest in our AI's
foundational capabilities to drive performance and efficiency for
advertisers. More recently we've used AI to suggest audiences, measure, improve
measurement and make campaigns easier to set up. And these are still the early
days of AI. To support AI innovation, Meta is investing tens of billions of dollars annually in our
AI technology stack.
From data centers to networking to cooling to chip design to machine design and graphics
processing units, an increasing portion of our capital expenditures
have been dedicated to building out our AI capabilities.
These investments are already producing huge benefits for the people and businesses that
use our products, including well and beyond our core business. On top of that, AI also helps detect and remove harmful content and fraudulent advertisers
as well.
Today I'm going to share with you some of Meta's current focus areas for investment
in AI and how they aim to drive better results for businesses and better ad experiences for
people.
We'll start with foundational infrastructure and AI technology.
Then we'll discuss AI modeling.
And finally, AI-powered ad experiences.
Across all of these areas, we've been working diligently to be thoughtful about the responsible use of AI.
I'll wrap up by taking a look at the future of AI
in our products with some exciting new developments.
Let's start with the foundational infrastructure
and AI technology.
Meta has been investing across the entire vertical technology
stack to improve the performance of its AI systems.
We've upgraded our data centers with larger networks with more cooling to support higher-powered
GPUs, graphics processing units, which are commonly used in AI today.
Using GPUs allows us to bring much more compute to bear to train AI models efficiently.
We've shifted away from CPUs, central processing units, the chips that typically power your
laptop or servers in a data center, towards more specialized compute and chips like GPUs
that power things like high performance gaming computers and, of course, mine cryptocurrency.
We've also modified our software infrastructure to allow us to train, transport and evaluate
larger and more complex AI models.
Our updated networks and software infrastructure allow us to coordinate large numbers of GPUs to train truly massive machine learning models on massive
data sets to improve the performance of these AI models.
This enables us to improve the predictions of which ads are relevant for which people
and at which times.
We've also worked to improve our AI tooling to drive more reliability and efficiency for
our AI models and allowing our AI engineers to focus more on foundational technology development.
Our focus on foundational infrastructure and foundational AI technology is enabling Meta's
ads team to drive better results for businesses and better
experiences for people.
Let's take a look at the second aspect, AI models.
The changes across the advertising ecosystem have led to reduced access to granular data.
So our engineers have had to create models that can help us fill in the gaps where
data is only partial or missing completely. These modeling improvements have led to measurable
improvements in performance, which many of you have noticed throughout the year.
One innovation that we are particularly enthusiastic about is the use of large transformer models that
can support many types of input and many types of prediction outputs across a variety of
objectives.
Transformer models are foundational AI tech that powers large language models like ChatGPT.
Previously, we used separate AI models
to predict the probability of a click
and the probability of a sale, and we
would compare the predictions across both models.
This presented challenges because the models weren't
seeing the same input data, so the predictions
weren't easily comparable.
Now with our transformer models, we train one large model
to predict both clicks and
sales, resulting in more consistent predictions that are more easily comparable, which allows
us to improve the performance for both types of objectives.
This also allows us to reduce the model maintenance burden, allowing our engineers to spend more
time focusing on innovation
and improving performance and efficiency for our advertising partners.
To explain how combining these models work with a truly oversimplified analogy, I want
you to imagine two students learning to play the piano.
If one of them already knows how to play a different instrument, like the violin, our violin-playing piano student will be able to more quickly learn the piano
than the second student who has never played an instrument before.
This is because while the piano and violin are different instruments,
they share a lot of related musical concepts between them.
Notes, scales, timing, reading music, and so on.
Now there are many differences between the ways that people learn and AI models learn,
but the basic concept called transfer learning describes how the models are able to reuse
patterns they learned in one use case for a different use case, improving their performance
on both.
Transfer learning and transformer models are just two of the AI modeling innovations that Meta's ads teams are using to drive better results for businesses
and better ad experiences for people.
Do you have business insurance?
If not, how would you pay to recover from a cyber attack, fire damage, theft, or a lawsuit?
No business or profession is risk-free.
Without insurance, your assets are at risk from major financial losses, data breaches, and natural disasters.
Get customized coverage today starting at $19 per month at zensurance.com.
Be protected. Be Zen. Third, let's talk about using AI modeling
in AI-powered ad experiences. You already heard a lot about Advantage Plus, and it is a huge
opportunity for advertisers. Advantage Plus shopping campaigns automate campaign setup
and can test up to 150 creatives at once.
Advantage Plus Shopping can also optimize delivery based on the audiences most likely
to make a purchase as well as predict where they would make a purchase, whether that's
in an app, in a shop, or on a website.
Also as you heard from Kristen earlier today, we recently announced some new features to Advantage Plus shopping and a new product to make it even easier to enable automation
as well as understand what's working and find new audiences outside of your traditional
parameters that will love your products.
We're expanding the ways that we use AI to enhance ads creatives with advantage plus creative
enhancements which will recommend variations and improvements to your
traditional and existing creative assets to drive better performance with the ads
audience some of the examples include of course adjusting the brightness or the
contrast or the color or aspect ratio, adding labels
from a page or altering text between the headline, description, et cetera.
Advantage Plus is bringing a more complete, automated and performant customer journey
across all of Meta's services powered by Meta's powerful AI.
We're continuing to improve the experiences like Advantage Plus
to predict which formats or surfaces
will drive conversions or other outcomes you're bidding for.
These are just a few of the AI-powered tools
and experiences that Meta's ads teams are developing
to drive better performance for businesses and better ad experiences for
people.
Now, I'd like to discuss how Meta is committed to the responsible use of AI.
As we develop new features and capabilities, we keep in mind issues of privacy, fairness,
and transparency and develop this technology in an open and collaborative way in consultation
with external AI experts and
regulators.
We want to build AI responsibly, always starting from a place of what is right for people.
Our responsible AI efforts are driven by a cross-disciplinary team of people whose mission
it is to ensure that Meta's AI benefits people and society as a whole.
You may have seen we recently developed an industry leading variance reduction system
to meet the needs of our advertising partners in highly regulated industries like housing,
employment and credit.
The variance reduction system uses reinforcement learning, a type of machine learning that
learns from trial and error, to optimize towards a specific outcome.
Using the variance reduction system, the audience for housing, employment, and credit ads more
closely reflects the population who are eligible to see the ad.
Meta's ongoing work in responsible AI is driven by our belief that AI should enable
everyone to have equitable access to information, services, and opportunities.
Our investment in responsible AI will enable Meta's ads teams to drive better results
for businesses and better experiences for people. Meta's ads teams evolve campaign set up with powerful AI.
Our next step is leveraging the new hotness, generative AI, to build, test and iterate
on ads creatives.
Our ads and AI teams have created an AI sandbox to explore the application of generative AI
to ads creatives.
The AI Sandbox supports early experimentation with our AI-powered features.
Our goal is to roll these out more broadly in the near future.
Let's talk about the features.
The first of which in the AI Sandbox is creative text variation, then creative background variation,
and third, creative outcropping, image outcropping.
Let's talk about text variations first.
An advertiser could quickly develop a large number of copy variations with each variation
performing potentially better with a different audience or on a different surface.
For example, an affordability message may work better for one audience, where a fashion-forward
message may work better for a different audience.
This can offer greater personalization and improve results.
We're also hearing from advertisers that better matching tone or creating more variations
of tone to match a brand's voice will be a useful application of this tool.
Second, we're introducing creative background generation.
An advertiser can generate backgrounds to better highlight their product, services,
or their existing creative.
We've heard from advertisers that the background generation tool can be a thoughtful partner
for creating new assets, especially for businesses
without a ton of resources. Creative background generation allows marketers to try out a variety
of backgrounds and iterate more quickly. As with text, there likely isn't a single best-performing
visual tone for all people on all surfaces. The third tool is image outcropping.
The image in a creative can be extended to fill the available space, for example, for
different aspect ratios on different surfaces, like reels or stories.
Image outcropping can help advertisers save time and cost while achieving better performing
ads.
Now, many of you will want to rush to call your meta partner about this technology today,
maybe even knock down their doors.
But aligned with our responsible AI philosophy,
we're being careful about the rollout.
In July, we will begin to offer access to advertisers
for the AI sandboxbox, and you should
also expect to see a number of these appearing in product later in the year.
Generative AI tooling is one more way that Meta's ads teams are using AI to drive better
results for businesses and better ad experiences for people. AI is helping Meta to improve ads' relevance, testing, efficiency, and experiences for advertisers.
These capabilities allow advertisers to drive better results, save time and effort, and
learn what works best so you can make better decisions with your advertising budgets.
We plan to continue our deep investment in AI to achieve better performance,
inspire new ideas, and help you succeed on our platform.
Thank you for partnering with us in this new and exciting era of AI-powered marketing.
We're really excited about what's ahead and look forward to working with you to making it a reality. Thanks.