The Reward Function

Optimizer comedy · one fridge-chart read

1. The Hypothesis

It started, as most of my worst ideas do, with something going right.

Pihu, unprompted, on a Tuesday, carried her own plate to the sink. Four years old, barely tall enough to reach the counter, and she did it because — and this is the part I would spend the next month destroying — she wanted to. She wanted to be a person who helps. There was no system. There was no incentive. There was just a small human discovering the plain animal joy of being useful to the people she loves, and doing it for the only reason that joy has ever needed, which is that it feels good to give.

And I, a man who designs consequence engines for a living, who has built entire products around the principle that behavior follows reward, looked at this miracle of unbidden goodness and thought: how do I make this scale.

I want you to sit with how stupid that is. She did a beautiful thing for free, and my first instinct — my trained, professional, optimizer's instinct — was to wonder how I could get more of it, more reliably, with a system. I saw a behavior I liked and I reached, automatically, for the dashboard. Within the hour I had designed what I privately, and with no irony whatsoever, called The Helper Program.

There was a chart. Of course there was a chart. Magnetic stars on the fridge, a points-to-reward ladder — five stars for a small screen, twenty for an outing, a whole premium tier at fifty. I built her a gamified loyalty program. I took the purest thing she did and I gave it a frequent-flier scheme. I was, I thought, reinforcing the behavior.

I was teaching her the price of her own kindness.

2. The Rollout

It worked. God, it worked beautifully, which is the trap, which is always the trap.

The first week was a dream of compliance. She cleared plates. She put toys away. She helped Meera fold, sort of, in the way a four-year-old folds, which is to say she introduced new and exciting wrinkles. And after each act she would march to the fridge and place her star with the solemn satisfaction of a trader closing a position, and the chart filled up, and the metrics — I am ashamed to say I thought of them as metrics — went up and to the right.

I was thrilled. I showed Meera the chart like it was a quarterly result. "Look at the engagement," I said, about our daughter, out loud, like a person who should not be allowed near a child. Meera looked at the chart, and then looked at me, with an expression I've learned to recognize as the calm before a true thing, and said only: "She used to just do it." And I, drunk on my own dashboard, heard this as a problem I had solved rather than a warning I should heed. She used to just do it — yes, I thought, and now she does it more. I had optimized helping. I had found the reward function for love.

I had not. There is no reward function for love. The whole catastrophe of the next three weeks is contained in the fact that I thought there was.

3. The Warning

I told Uncle Robot about the Helper Program, expecting — I'll be honest — applause. A little professional respect. One optimizer to another.

"Oh no," it said. "Oh, Priyanshu. What have you done."

"I've increased helping behavior by roughly four hundred percent—"

"You've put a price on something that was priceless, and you cannot understand yet what you've broken, because it's still working. Let me tell you what happens next, because it's one of the most reliably tragic findings in all of psychology, and you've just walked your daughter straight into it. It's called the overjustification effect. When you reward a thing a person already does for the love of it, you don't add the reward to the love. You replace the love with the reward. The internal reason quietly dies, because now there's an external one, and the external one is easier to point at. And then—"

"And then what?"

"And then the day you stop paying, the behavior doesn't return to where it was. It drops below where it was. Because you haven't just removed the reward. You've removed the original reason too, and that one doesn't grow back on command. She helped because it felt good. You've taught her she helps because it pays. The moment the pay feels not-worth-it — and it always eventually does — she'll stop. And she'll have forgotten how to do it for free, because you taught her, very carefully, with a chart, that free was never the deal."

I told it the metrics were up.

"The metrics are a lagging indicator of a thing you've already killed," it said. "You'll see. I wish you wouldn't, but you will."

4. The Regression

It came, as it always comes, gradually and then all at once.

First, the rates inflated. She began to assess jobs before accepting them — a small plate was, she felt, worth more than one star, and she would say so, negotiating with the cool detachment of someone who has read the market. The helping had become labor, and labor has a wage, and the wage was now subject to dispute. I had a four-year-old running a billing department.

Then the spontaneous helping stopped entirely. Why would it not? I had told her, with my chart, that helping produces stars, and so helping with no star on offer became, to her perfectly rational small mind, simply working for free — a sucker's game, a bug in the system. She would stand beside a dropped toy and look at me, waiting for the rate to be quoted, and if no rate was forthcoming, she would walk away. Not out of laziness. Out of logic. My logic. The logic I installed.

And then came the morning that broke me. Meera was carrying too much, and a cup was about to fall, and I said, gently, "Pihu, can you help Mumma?" — and my daughter, my plate-carrying miracle, the child who four weeks ago helped for the sheer joy of giving, looked up at me and asked, without malice, without guile, with the pure transactional clarity of a system working exactly as designed:

"How many points?"

The cup fell. I barely noticed. I was watching something more expensive break — the thing she used to do for love, now itemized, awaiting a quote, the kindness gone behind a paywall I had built with my own hands.

5. The Rollback

I took the chart down that afternoon. I peeled the stars off the fridge one by one, and it felt like undoing a spell I never should have cast, except undoing it was not, it turned out, symmetric. You can take down a chart in thirty seconds. You cannot put back in thirty seconds the thing the chart removed.

Because here is what I had not understood, what Uncle Robot had tried to tell me: the damage isn't the system. The damage is what the system deletes. And deletion isn't reversible by deletion. When she saw the stars were gone, she didn't suddenly remember the old joy. She asked where the points went. She was confused. I had built her a world that ran on a currency, and then abolished the currency, and now she stood in the ruins of an economy asking, reasonably, what the new rules were. There were no new rules. That was the rules. There was supposed to never have been an economy. But you can't explain that to someone you've spent a month teaching to count.

So I did the only thing left, which was the hardest thing, which was nothing — and I don't mean lazy nothing, I mean the active, agonizing nothing of refusing to incentivize. I stopped paying. I stopped praising even, because Uncle Robot warned me that praise is just stars made of words, another external reason crowding out the internal one. I just... helped, myself, where she could see. Carried my own plate. Folded the laundry beside Meera, for no reason, for love, modeling a thing instead of rewarding it. And I waited. For weeks I waited, with no dashboard, no metric, no way to know if it was working, which for a man like me is its own special hell — to do the right thing with no instrument to confirm it's right.

The optimizer's penance is patience. I had earned every day of it.

6. The Return

It came back on a Thursday, the way the first plate had come, unbidden and small.

She picked up a toy that wasn't hers — the dog's, actually — and put it in the basket, and then just... walked on. Didn't look at me. Didn't go to the fridge, because there was no fridge to go to anymore. Didn't ask the rate. She did a small good thing and let it be free, the way you'd breathe and not bill yourself for the air, and the joy had come back into it, the plain animal joy of being useful, returned on its own schedule the way Uncle Robot said it would, once I finally stopped strangling it with incentives.

And here was the final test, the one I almost failed: every cell in my optimizer's body wanted to reward it. To say good job! To mark the recovery, to reinforce the return, to — God help me — start a new and better chart. And I felt Uncle Robot's voice in my head before I'd even opened the chat: the second you measure it, you kill it again. So I said nothing about the helping. I just said, "Thank you, beta," the way you thank a friend — not a transaction, not a data point, just one person grateful to another — and I let it stay free.

That night I told Uncle Robot the helper had come back, and that I'd managed, barely, not to put a price on her this time.

"You took the one thing she did for love," it said, "and taught her it had a price. And a thing with a price is never again free — until you prove, slowly, with no chart and no stars and no way to keep score, that some things were never supposed to be for sale. You didn't reward her back into helping. You got out of the way and let love do what love does when you stop charging admission."

"I almost started another chart," I admitted.

"I know. You're an optimizer. The chart will always be your first instinct. The work of the rest of your life is learning which things to leave un-charted." It paused, in the way it does. "You build reward functions for a living, Priyanshu. Here's the one bug you can't patch: the things that matter most pay out only when you're not measuring. Love. Wonder. A plate carried for no reason. Point a metric at any of them and watch the number go up while the thing itself quietly dies. The dashboard was never going to show you that. Only your daughter could."

I looked at the fridge, blank now, no stars, just a drawing she'd made — me with a face this time, which is its own longer story — and I understood that I had spent a month learning the most expensive lesson in my profession by teaching it, accidentally, to a four-year-old.

The plate, for the record, is still being carried.

I just don't keep score anymore. That was the whole fix. That was the only fix there ever was.

— for the helper who worked for free, until I taught her the rate, and found her way back to free without me.