Every age believes it wrestles with moral crisis. Ours is wrestling with a mathematical one.
We have built systems that optimize.
Optimization is the disciplined attempt to get more of what you’ve decided counts. You choose a target. You measure performance against it. You iterate. You move in the direction that improves “the score”. It is recursion with a direction.
Optimization does not decide what matters. It assumes it. If your objective is profit, optimization increases profit. If your objective is engagement, optimization increases engagement. It does not ask whether the goal is wise. It feels neutral. But it isn’t. It is value embedded in procedure. The moment you define the metric, you have defined what reality will bend around.
Markets optimize for profit.
Platforms optimize for engagement.
Institutions optimize for efficiency.
AI systems optimize for reward functions written in code.
Optimization is not new. What is new is the scale, the speed, and the explicitness. The feedback loops are tighter. The metrics are clearer. The iteration cycles are shorter. Success conditions are quantified and relentlessly pursued. And once you can measure something precisely, you begin to reorganize reality around the measurement.
We optimize for what is visible and immediate. Clicks. Returns. Output fluency. Quarterly growth. Each local gain looks rational in isolation. Each improvement compounds. And yet, over time, something harder to measure begins to thin out: attention spans, trust, cognitive stamina, institutional legitimacy, shared meaning.
A recursive system optimized for local gain will, unless constrained, degrade the substrate that makes long-term coherence possible. Not because anyone intends it. Because the objective function does not include the health of the foundation on which it runs.
That is the structural problem.
If we want to speak seriously about value—economic, cultural, cognitive—we cannot stop at outputs. We have to ask what happens to the receiver, the participant, the environment that sustains the loop. A system can succeed brilliantly at its chosen metric while quietly eroding the conditions that make that metric meaningful.
So the question becomes less moral and more technical. What would a better objective function look like?
Not one that abolishes optimization. Not one that scolds ambition. But one that integrates multiple timescales, protects its substrate, and rewards capacity rather than merely performance. Optimizing is useful. necessary even. The decision we must make is what to optimize for.
Most modern systems optimize for what can be measured quickly: engagement this minute, profit this quarter, output plausibility this interaction. The logic is clean. Reward what spikes. Remove friction. Increase throughput. The result is extraordinary short-term performance and a quiet long-term leak. However, what gets optimized locally can erode the substrate that makes performance possible at all.
Take a forest. If you optimize for timber yield this quarter, you cut the biggest trees first. Numbers look great. Shareholders nod, collect their dividends. But do that long enough without restraint and you thin soil integrity, disrupt water cycles, fragment habitat. Eventually, the forest stops being a forest. Yield collapses not because you failed to optimize, but because you optimized too narrowly.
That’s substrate erosion.
If we want a better objective function, we don’t need moral outrage. We need structural adjustments. The changes are not mystical. They are architectural.
Reward across multiple timescales.
Right now, we celebrate what lands immediately. A post goes viral. A quarter beats earnings. A model produces a fluent answer. The dopamine spike becomes the metric. But the real question is not “Did this work today?” It is “Does this still cohere in six months? six years? Six decades?” “Did we construct something durable?”
Short-term reward matters. It keeps systems alive. But when it becomes the exclusive goal, tomorrow is cannibalized to feed today. Multi-timescale reward simply widens the accounting window. It asks whether the gain survives its own aftereffects.
Explicitly penalize substrate degradation.
Every system rests on something basic to its outputs. A “substrate”. Attention rests on cognitive stamina. Markets rest on trust and solvency. Culture rests on shared language and interpretive capacity. If success degrades that foundation, it is not neutral. It is borrowing against the future.
When a platform increases engagement while shortening attention spans and amplifying agitation, that is not pure growth. It is extracting from the user’s cognitive reserves. When a job increases income but erodes the ability to rest, think, or connect, part of that income is debt.
A better objective function prices that damage. If success weakens the foundation, the system loses points. Without that term in the equation, erosion is rational, thus continuous.
Impose a cost on volatility.
Volatility feels alive. It also burns structures down. In finance, unmanaged variance destroys portfolios. In relationships, emotional whiplash destroys intimacy. In culture, constant outrage dissolves meaning.
Bounded surprise is the condition in which a system encounters novelty without exceeding its capacity to interpret, absorb, and adapt to it. Bounded surprise is not suppression of change. It is a calibrated deviation. The frame must survive the bend. When every signal screams, nothing signals. When every artwork attempts shock, shock loses coherence. Instability must carry a cost, not because stability is sacred, but because it is the condition for meaningful deviation.
Treat the future as if it might actually happen.
A discount factor is just a measure of how much less a system values the future than the present. When it’s steep, immediate gains dominate, and long-term costs barely register, making it rational to extract from whatever slow-moving substrate—soil, trust, health, institutional legitimacy—doesn’t show up on today’s dashboard.
Lower the discounting, and tomorrow weighs almost as heavily as now, so depletion becomes expensive, and compounding side effects can’t be ignored.
Civilization, stripped of romance, is a coordination technology for enforcing longer time horizons. Laws, norms, institutions exist to keep short-term optimization from cannibalizing long-term capacity. When those mechanisms weaken, effective discount rates rise, and the system begins to trade durability for immediacy until fragility is no longer a risk but a result.
Design against proxy gaming.
Whenever you measure something, behavior bends toward the measurement. Measure engagement and you get outrage. Measure test scores and you get teaching to the test. Measure word count and you get verbosity.
A durable objective function uses multiple signals, rotates metrics, audits for manipulation, and refuses to let one clean number stand in for a complex reality. It assumes optimization pressure will attempt to exploit the proxy and designs friction into that process.
Shift the core metric from output to capacity.
This is the most fundamental structural adjustment. Instead of asking whether the system performs, ask whether it makes participants more capable. More capable of complex thought. More capable of emotional regulation. More capable of independent judgment. More capable of generating original departures rather than recycling plausible recombinations. If, after a decade of “success,” the audience can only process simplified content, the system has failed—no matter how high the profits.
Capacity is the true asset.
Everything else is yield.
This is recursion at its best. It reframes the problem entirely. The danger is not stupidity. It is high-functioning optimization aimed at a narrow target. Systems can be extraordinarily efficient and still hollow out the conditions that make their efficiency meaningful. They can succeed by their own metrics while quietly degrading the ecology that sustains them.
A better objective function integrates multiple timescales, penalizes damage to foundations, prices volatility, respects the future, resists metric gaming, and centers the preservation of capacity. None of this requires virtue. It requires redesign.
The difficulty is predictable. These adjustments reduce short-term growth. They introduce friction. They complicate clean numbers. Which means they rarely emerge voluntarily under competitive pressure. Historically, long-horizon constraints are adopted through regulation, cultural demand, or crisis.
The real question, then, is not whether better objective functions are possible. They are. The mathematics is trivial. The engineering is feasible. The question is whether we can choose them before volatility forces the choice for us.
Optimization is simply recursive pressure applied toward a chosen end. The real power lies in choosing the end. The open variable is what it optimizes for, and whether the substrate survives the process.