How systems learn a quirk you don't even know you have — and why that's a gold mine.
One day, while ordering from Subway through a food-delivery app, I noticed something odd: I couldn't find my usual coupon.
I have a habit. For any order above ₹300, I check the price across different apps and pick the best one. But when I'm in the office I usually don't bother — even if no discount coupon is available, I still place the order, because it's about time. For the first few days I saw the coupon. After a few days, it stopped appearing. Strangely, that same week, when I was working from home, the coupon showed up again.
Were they really acting on my behavior pattern? Maybe I was wrong; maybe I was mistaken. But it kept me thinking.
A behavior pattern is different from a user-segmented flow — for example, iOS users being shown a higher price. The key difference is that this is a quirk or pattern the user themselves isn't aware of, and a system identified it.
Take the famous market-basket analysis: at Walmart, shoppers tend to buy beer and baby diapers together on Fridays, and placing them side by side lifts sales. But that's a system spotting a behavior the customer already knows they do. The interesting question is the other one — what about a quirk the user probably isn't even aware of?
To explore this, I built the most famous game of all — Rock, Paper, Scissors — and made the camera read your real hand, not a button (a button only trains the system on your clicking habits).
Show a real Rock, Paper or Scissors to your camera. Play a few games and watch the AI's win rate climb as it learns your quirk.
The coupon I saw (and didn't see) is just the entry point. The same idea — read a quirk you can't perceive, then ship a personalized "delta" in the flow — scales in three tiers.
Offers tuned to how you behave, not who you are — your hesitation, scroll speed, what you abandon in the cart. Coupons withheld from people predicted to buy anyway; nudges timed to a detected mood. (My Subway coupon lives here.)
The flow builds itself around you — AI-generated screens unique to each person, and a reinforcement-learning loop (exactly like this game) that discovers the nudge sequence that converts you. Friction made as hard as you'll personally tolerate.
The board flips. When your own AI agent shops for you, sellers can't read you directly — so they target the agent instead, and predict life events (a move, a new job) to pre-position the offer before you've decided.
The system keeps several running models of you and updates all of them on every throw. Each one predicts your next move from a different angle:
For every prediction it considers three layers of second-guessing — counter you, counter your counter, counter that — so it still wins when you try to outsmart it. It scores every model-and-layer combination on how well it has predicted you recently (across a fast and a slow memory) and plays whichever is winning. A pure-random strategy sits in the pool as a floor: the moment you crack its pattern, it falls back to unbeatable randomness. Everything is stored on the device and carries across games, so the read of you keeps sharpening.
That gap is the entire point. Predictability is exploitability. The more legible your behavior, the more the system beats you — and in a real store, that "win" is your wallet. The defense in the game is the same as in the app: be aware, and don't be readable.