It is a ML World, and it will be nothing without a playable or two 

The way UA and creatives worked for a good part of the past decade is dead. You could smell it dying when ATT hit. 

Facebook's god-tier user graph - that beautiful, terrifying machine fed by in-app events from half the digital world - started to rot. That machine that had insane intelligence on pretty much all of us, what we bought, what we did across all our apps and games, started to decay.

The targeting that did 80% of your job for you was gone. 

Suddenly everyone started to scramble, and the industry narrative became "creative is king now."

But that's only half the story.

Creative matters more than ever. But not the way most teams think. Not the way agencies sell it. Not the way your last three hires learned it.

Here's what actually changed: You're feeding an algorithm that's searching for patterns you'll never see, making connections you'd never make, finding audiences in places you'd never look.

Welcome to the new game.

When Engines Beat Grandmasters

Spring 1997. Deep Blue beats Kasparov in Chess. Everyone loses their minds.

But that's not the interesting part.

The interesting part came later, when engines became unbeatable and grandmasters had to learn how to train with them. They discovered something uncomfortable: The moves that actually win look completely f*cking wrong.

Engines sacrifice pieces in ways that violate every principle humans spent 200 years teaching. They embrace complexity when classical theory screams "simplify when winning." They make moves that look like mistakes until you're getting destroyed and you still don't understand why.

One chess master nailed it: "If a 1960s grandmaster played against AlphaZero, they'd think it was playing like an average club player... until they got crushed without understanding why."

For two centuries, chess training was built on human pattern recognition. Control the center. Develop your pieces. Castle early. Don't hang your queen.

Engines don't care about your patterns.

They calculate positions humans never considered worth calculating. Turns out a lot of what we thought was "good chess" was just "chess that makes sense to brains that can hold seven pieces of information at once."

The same thing is happening in UA right now. And most teams are still playing by the old rules.

What is Actually Replacing Facebook Era Targeting

Listen to how AppLovin's CEO Adam Foroughi describes AXON:

"Today, in our advertising system, the advertiser can do almost no targeting. They can pick their country, they can put in their economic goals, they can put in a budget and off they go. The one manual lever is creative."

Read that again. No targeting. Just creatives and ML.

Pre-ATT Facebook had the most effective targeting machine ever built. Rich user graph. In-app events from every app that mattered. Product-audience fit so precisely you could target "women in Denver who just broke up and like true crime podcasts + just made an in-app purchase in a match-3 game." With super specific audience building creative took a backseat. 

That world is over.

What rose from the ashes - AXON, Unity's Vector, the whole new breed - works nothing like that. 

These are different beasts entirely. And if you're still trying to win the old way, making a few great creatives based on your gut about a well-defined audience, congratulations: you're the 1960s grandmaster wondering why you're getting crushed by moves that look wrong.

How Modern ML Driven Actually Works (And Why It Changes Everything)

Nobody knows exactly what happens inside AXON's neural network. Black box. That's the deal with ML. You can't open it up and see the reasoning. A core characteristic of ML is lack of interpretability.  

But we can observe it from the outside, and observe the dynamics of how it works, and how the biggest UA teams in the world are winning. And once you see it, really see it, you'll understand why the old playbook is, well, old.

The Search Problem

Every time AXON looks at a bid request, it can be viewed as solving a search problem.

Hundreds of millions of users it could show ads to. Your set of creatives sitting there. 

It needs to predict the probability that the user will deeply engage with a creative, let’s say a playable. 

It needs to pick the creative most likely to drive engagement. It also needs to predict based on that bid request, as well as engagement within the playable, or other creative, the estimated LTV of that user. Finally, to decide what the bid should be, if any.

Its job in a nutshell: predict which combination produces high-LTV players and then bid aggressively on them, scaling your spend profitably. 

Step 1: Mostly Random Sampling

Given a set of various creatives and playable ads - the engine starts to sample impressions and start to gain learnings about which bid request and user characteristics correlate with the likelihood to engage (meaning enjoy in most cases) with the various playables and creatives. 

In the case of playables, each concept represents different game and play experiences, thus allowing the machine to make more sophisticated learnings based on these patterns. 

Step 2: Finding audiences that enjoy different playable concepts 

Later on, the machine will become better at predicting which bid requests / users will engage with which creative and is able to match creatives with users much better. 

We may think about it as identifying different groups of users / audiences that share the same characteristics that are more likely to engage with idle mechanics and certain themes, and other audiences that are more likely to engage with puzzle mechanics - as an example. 

When it comes to playables we can estimate that the much deeper signals that rise from a user playing and engaging with a playable game (as opposed to passively watching a video, whether they want to or not), allows for much better learning of the ML UA engine, meaning the AppLovin AXONs and Unity Vectors of the world. 

Step 3: Finding which playable concept play signals lead players to install and pay 

As more users engage with the different playable and the machine can observe the creative performance, meaning its IPM, various conversion rates, and later the ROAS it drove, it can learn which styles of play represented by the playables - had a high affinity to the game itself - and drove players to play, engage, retain, and monetize well. 

When this happens, and a model has been established for predicting high LTV for potential players through these playables and creatives, your network can bid aggressively and scale spend - creating the sought after winning creative. 

The machine that match audience → playable concept → high LTV in your full game 

So an interesting way to think about it all, is that you have access to a machine that will search for: 

• Players that will likely to engage and enjoy your playable concepts and the gameplay they represent 
• Playable concepts and game experiences that have a high affinity for your actual game, meaning that players who enjoy the playable will likely enjoy the full game, become retained users and will have a high LTV. 

In this perspective your job can and should be to produce many different types of playable concepts, game loops, themes, and aesthetics, each basically testing the hypotheses of:

1. There is a wide and large enough audience of potential players for this type of playable experience.
2. There is a strong affinity between enjoying the playable game concept and enjoying the full game. 

You are basically creating, and testing through playables, many FTUEs, or “intro games” for your actual full mobile game. 

You are also de-facto, performing a casual-ization or even hyper casual-ization of different slices of your real gameplay to find better “entry points” and smoother onboarding to your game straight from the ad. 

Why Not Supplying the Network Enough Diverse Creatives Is a Death Sentence

So in order to allow these ML machines to actually explore effectively the immense search space of potential players to bid on, you need to face the reality of not knowing which creative will work and why. 

As we said, ML is not interpretable, we can not know for sure why a certain creative won as opposed to another one. What exactly was the learned insight that led to identifying that a certain set of user characteristics led to them engaging with a playable, and why that playable concept had such a strong affinity to your game. 

In real life, why so many users that enjoy and engage with the famous Whiteout Survival playable concept are not discouraged when seeing the actual 4x game for the first time, but instead they retain, and monetize at impressive levels (which must be true if these playables scale so much that you probably know exactly what I’m talking about).

Given that you don’t know what will work, the main way to figure that out is to let the machine do its thing - maximize its learnings by supplying it with many options to test fast until it finds a winner. 

Providing it with too few creatives seriously minimizes its possible action space, meaning which creatives it can try for various audience segments in order to learn. 

A Concrete Example So This Isn't Just Theory

Say you have a 4X strategy game.

You make playables exploring different loops:

• Idle tower defense (tap to upgrade, defend against waves)
• Card collection (build deck, battle with cards)
• Hypercasualized base building (gather → build → grow)
• Puzzle mechanics (match-3 meets resource management)
• Endless runner combat (dodge, shoot, collect)
• Merge loops (combine units to make stronger ones)

Each one is genuinely fun. Not a reskin. Different core loop. Different game feel. 

AXON starts testing these across different user types.

After a few thousand impressions, patterns emerge:

The idle tower defense playable crushes with a specific user segment AXON identifies. Starts bidding more aggressively on users matching that pattern. That creative scale. Becomes 30% of your spend.

Meanwhile, the card collection playable works incredibly well with a completely different user type. That scales too. Another 25% of spend.

The puzzle playable that your team thought was the creative winner? Performs okay but players installing from it do not monetize and churn quickly so its share of spend never increases. 

The creative that looked "best" to everyone in the room is your fifth-best performer.

The merge loop playable that one designer made almost as an afterthought? Turns out there's a massive audience of users who love merge mechanics and also love your type of strategy game. AXON found them. That playable becomes 40% of your spend.

AXON, for some reason unknown to us, also found that a lot of this user segment exists within coloring games - a connection that would sound insane if you tried to explain it to your CMO - and it becomes your biggest channel for the next month or two. 

This is why the creative lottery is over.

You can't predict winners anymore. The patterns are too complex. The audience segments are too high-dimensional. The connections between what people enjoy and what they'll pay for exist in spaces human brains can't navigate.

You can only create enough good options that the algorithm can discover them.

What This Actually Means for UA Teams

Three things changed fundamentally:

1. Volume and Diversity Beat Creative Genius

One "brilliant" concept based on your creative team’s intuition about your audience? Not enough. Not even close.

Thirty genuinely different concepts that explore different game loops, different mechanics, different ways players could have fun with your game? Now you're playing the right game.

Fundamentally different game loop experiences that each provide AXON a different way to learn about user preferences.

The old model was: Make a few brilliant things to the audience you think you know and place them perfectly.

The new model is: Make many good things and let the algorithm discover which ones work for which audiences.

2. Taste Matters Differently (And Maybe More)

You still need taste.

But not only taste in "what will this audience respond to?" That's mostly the algorithm's job now. Especially as you can’t clearly define the audience ahead of time. 

You need taste in "is this playable actually fun to play?"

Because if the game loop isn't satisfying - if it's not genuinely entertaining - it won't generate the right signals even if the concept is sound.

A boring tower defense playable won't teach AXON that certain users love tower defense. It'll teach AXON that users don't engage with this specific bad execution of tower defense.

You need taste in game design itself. Core loops. Meta Loops. Progression. Moment-to-moment fun.

It's a different skill. Closer to being a game or level designer than being solely an art director. Closer to understanding juice and game feel than understanding audience creative art preferences.

3. Your UA & Creative Org Needs to Feel Also Like A Lab and Not Just A Creative Agency

The goal isn't to craft one perfect creative vision and execute it flawlessly.

The goal is to systematically explore a massive space of core game loop variations, genres, aesthetics, and themes, feed them to the algorithm, observe what works, and make more of that.

While continuing to explore wildly different concepts.

It's less Mad Men, more a research lab.

Less creative genius, more systematic discovery.

Less "I know what will work," more "let's give the system enough good options to find out."

The teams winning now feel different. They ship faster. They test weirder concepts. They're comfortable not knowing which thing will work. They trust the process over their intuition.

They've made peace with the fact that the algorithm is better at finding audiences than they are.

The Uncomfortable Truth About Taste vs. Systems

Here's the part that makes creative people uncomfortable:

In the old world, having great taste in what audiences wanted was the most valuable skill. The creative director who could look at a concept and say "this will crush without audience" was worth their weight in gold.

In the new world, that skill is worth less. Not worthless - you still need to judge whether a playable is fun - but less.

What's worth more now: The ability to systematically produce high volumes of diverse, quality playables.

That's a different kind of taste. Taste in systems and processes. Taste in what makes a good creative production pipeline. Taste in how to structure a team so they can explore broadly without sacrificing quality.

The valuable people are the ones who create conditions where winners can be discovered.

The Broader Implication Nobody's Talking About

This probably applies to more than just UA.

Any time you're feeding an ML system that's trying to optimize for something complex, the same logic holds:

Your job isn't to predict the right answer. It's to give the system enough good options and actions to take that it can discover the right answer.

That's a completely different problem than what most people are trained to solve.

Most of us were trained in a world where taste and intuition were the most valuable skills. Where "knowing what people want" was the pinnacle of expertise. Where the person who could look at options and pick the winner was the most valuable person in the room.

In an ML-driven world, the most valuable person is the one who can create systems that generate enough good options for the algorithm to explore.

Human intuition is still useful - for making each playable good, for understanding what makes game loops satisfying, for knowing which concepts are worth exploring, for recognizing when something is genuinely fun versus technically correct but soulless.

But it's not useful for predicting which specific creative will scale. Not useful for guessing which audience will respond. Not useful for placing your bets on a few concepts you're sure will work.

So What Do You Do?

If you're still trying to win by making three brilliant creatives based on your intuition about your audience, you're playing yesterday's game.

And yesterday's game is dead.

The game now is: Build systems that let you create enough diverse, quality playables that the algorithm has options to explore.

Feed the beast. Let it hunt. Scale what it finds.

Your taste matters in making each individual playably genuinely fun. Your strategic thinking matters in deciding which concepts are worth exploring. Your craft matters in engineering a system for winning.

But your ability to predict what will work?

That's worth a lot less than it used to be.

The sooner you make peace with that, the sooner you start winning.