Different targets, different paces, and rewards I can’t fully trust.
When I was a kid, rewards were easy to read. Get a hundred on a test, and the people around you stop telling you what to do. You get your freedom back. Finish your homework, and the rest of the day is yours.
The tasks were simple, so the reward worked. A math quiz has right answers. One number told me most of what I needed to know.
Then the tasks got bigger. Picking a major. Choosing a job. Deciding who to spend time with. Building something over years. The rewards still came in the same shape: one number, one yes or no, one “great work.” But they told me less and less.
It took me a long time to notice. The more complex the task, the more a single reward can mislead.
One reward, long task
I spend a lot of time around machine learning, and this problem shows up there constantly. On a short task, one reward at the end is fine. On a long task with many steps, it gets shaky. It says whether the whole thing worked. It doesn’t say which steps helped and which ones hurt. That’s the credit assignment problem.
And if the reward is even a little off, a model will happily learn to chase the reward instead of the goal. That’s reward hacking. I wrote about this for AI agents in an earlier post. At some point I realized my own life had the same shape.
A promotion, or not getting one, is one signal for years of work. It folds in skill, a manager’s taste, what the team needed that quarter, who happened to notice, and plain luck. I used to read praise as proof I was on the right path, and silence as proof I wasn’t good enough. Now I think the signal carries some information. It just isn’t the answer key.
Some rewards also arrive late. A choice I make today might only show what it meant years from now. By then, the short-term reward has long since told me something else.
And I only ever get to see one path. I picked A, so I know what happened on A. I’ll never see the same me, at the same moment, take B.
So one outcome can’t really tell me whether a decision was good. Good decisions end badly sometimes. Bad ones get lucky. These days, when I look back at a choice, I try to remember what I knew at the time, not just how it turned out.
There’s a quieter problem too. Goodhart’s law is usually put this way: when a measure becomes a target, it ceases to be a good measure. Grades, titles, and pay started out as rough signs of something real. I never thought I cared about them that much. But it’s easy to get pulled into the rat race anyway. Once everyone around you is chasing the same numbers, you end up chasing them too, and they stop pointing at anything real.
Maybe life has more than one target
If one reward can’t describe a long task, maybe part of the problem is that I’ve been treating life as if it had one target.
Maybe it’s closer to multi-task learning. Work is one target. So is health. So are the people I love, the things I’m curious about, the things I make, some quiet, and some freedom.
Some of these help each other. A good friendship makes a hard week easier. A long walk helps me think. Others compete for the same hours and the same attention. And a few don’t feel like targets at all. Health, basic dignity, a couple of principles. Those feel more like floors than things to trade.
Seeing it this way changed how I read rewards. A good review at work is a signal about one target. It says almost nothing about the rest.
Different targets, different settings
This is the part that surprised me most. If these are different targets, there’s no reason they should all be learned the same way.
In ML, every problem comes with its own settings. The learning rate. How often you update. How far ahead you look. How much you explore. Get those wrong, and even a good model learns badly.
Life’s targets seem to run on different clocks too. Work skills can move fast, with feedback in weeks. Health is slow, and big swings rarely stick. Relationships are slower still, built more by showing up than by any single moment. And curiosity barely has feedback at all, which may be part of why it’s fun.
Some settings are shared, though. In multi-task learning, many tasks share one backbone, and each gets its own small head on top. Life seems to have a backbone too. Sleep, attention, how I treat people, how I handle a setback… When those are in good shape, everything feels a little easier. When they slip, everything gets harder all at once.
And sometimes targets pull in opposite directions. ML calls this gradient conflict: a step that helps one task hurts another. Late nights before a deadline help work and hurt health. I used to see that as failing at balance. Now it just looks like what having more than one target feels like.
And the targets keep moving
Even with good settings, the problem doesn’t sit still. What I wanted at twenty isn’t what I want now. New targets showed up: new interests, new people, new responsibilities. Some old ones split, merged, or quietly faded. In ML terms, the problem is non-stationary.
That makes life look a lot like continual learning. Learning from a stream that keeps changing, while trying not to lose what’s already there. Sometimes only the how changes. Sometimes how much something matters changes. And once in a while the targets themselves change.
Continual learning has two classic ways to fail. The first is catastrophic forgetting. You learn the new thing, and it overwrites the old one. I’ve seen this happen with a new job. It takes over everything, and a year later old friends, old hobbies, and old skills have just slipped away. No one decided to drop them. There was just never time to go back. In ML, one common fix is replay, revisiting a bit of the old data while learning the new. I like that idea more than I expected to.
The second way to fail is the opposite. Holding on so tightly that nothing new gets in. That’s the stability–plasticity dilemma. Too stable, and old beliefs lock in. Too plastic, and nothing builds up.
So I’ve started to think forgetting isn’t only a bug. Some things seem worth letting go of. Targets that don’t fit anymore. Old scorecards, like the habit of waiting for a grade. A story about who I was supposed to be. Spending years on something doesn’t mean I have to keep it forever.
What stays is what carries over: skills, people, ways of thinking. That’s why changing direction rarely feels like starting from zero. ML calls this positive transfer. I wrote more about how it might work for agents in another post.
And learning doesn’t only happen while I’m pushing. Bread is a good reminder. You knead the dough, and then you leave it alone. More kneading won’t make it rise.
Something like this seems true for people too. Sleep plays a real part in how new memories settle and get reorganized. Some of the learning happens after I stop. And even on days when rest doesn’t make me better at anything, I’m trying to let it be enough.
Maybe it’s not meant to be solved
Looking at life as multi-task learning made a lot of things clearer for me. Why one reward can mislead. Why each part of life needs its own pace. Why forgetting can be part of learning.
But clear isn’t the same as solved. I can tune every setting and optimize every part I can reach, for years. The targets still move. The feedback stays noisy. There’s no final score at the end.
Maybe that’s the part I like most. If there were one answer, everyone would end up living the same life. Same targets, same settings, same score.
But there isn’t one. Some mysteries stay mysteries. Some problems never get solved, only lived with. And what works for one person barely works for the next. So people end up in all kinds of places, chasing all kinds of things, for reasons that don’t always make sense from the outside.
That’s a lot of randomness. It’s also most of what makes life interesting.
Maybe a problem with no clean solution isn’t something to fix. Maybe it’s the reason no two lives look alike.