Using Proxy Metrics for Startups and Life
Four articles, one diagnosis
Four unrelated essays crossed my desk this week: different authors, different industries, no shared citations. But they keep circling the same diagnosis: modern systems are getting better at their metrics while getting worse at their aims. A neuroscientist on science, a Popperian on startups, an economist on technical labor, and a young essayist on personhood. Different domains, same drift: proxies become targets, and optimization eats the thing it was meant to serve.
The lineup:
Tim Requarth in Persuasion - “The Real Reason Science Is Broken.” AI turbocharges individual scientists but narrows collective science. More papers, fewer breakthroughs. The bottleneck isn’t technological but institutional: broken incentives that reward production over progress.
Andras Ludanyi on Medium - “Startups as Epistemic Extremes.” Startups are structurally forced into epistemic recklessness. Refutation is terminal. Large organizations could absorb failure but instead suppress it politically. Neither system lets learning survive.
Noah Smith in Noahpinion - “The Fall of the Nerds.” Software stocks crater as “vibe coding” renders craft expertise obsolete overnight. The entire economic age built on human technical capital may be ending.
Freya India in After Babel - “You Have To Be Human.” When everything is automated and optimized, being genuinely human—stumbling, risking, holding convictions—becomes the rarest and most valuable thing left.
I should confess: I noticed the pattern because I live inside it.
I spend an embarrassing share of my day trying to optimize my thinking and input: headlines, argument structure, justifying procrastination with stats, the little dopamine drip of analytics.
I am a person who has spent years learning to think technically—to treat every question as a system to be tuned and every ambiguity as a problem to be resolved. I have built my working life around the assumption that the right framework, properly applied, would eventually yield the right result.
These four articles, taken together, suggest that assumption is part of the problem.
I. The Optimization Trap
Every one of these pieces describes a system that replaced its purpose with a metric, then optimized the metric until it consumed the thing it was supposed to measure. This is the classic problem described by Theodore Porter, whose portability thesis described how institutions substitute quantification for judgment: nuance versus legibility (the man’s name is really Porter!?)
Start with science. Requarth reports on a Nature study analyzing 41 million research papers:
“Scientists who adopt AI tools publish three times more papers and receive nearly five times more citations. Their careers accelerate. But the collective range of scientific topics under investigation shrinks by nearly 5 percent, and researchers’ engagement with one another’s work drops by 22 percent. The tools that turbocharge individual scientists appear to be narrowing science as a whole.”
The metrics go up but the discovery goes down. And the reason is structural, not technological:
“Researchers are responsible for raising their own funds through grants with success rates of around 10 percent. That creates enormous pressure to keep producing — to always have the next application in the pipeline, the next paper ready to publish. Institutions reinforce this by judging researchers on what’s easy to measure: publication counts, grant dollars, citation metrics. These are markers of production, not progress.”
Requarth lands the sharpest metaphor of the bunch: AI in science is “like adding lanes to a highway when the slowdown is actually caused by a tollbooth.” The technology isn’t breaking science. It’s accelerating scientists through an already-broken reward system.
Now look at the startup world. Ludanyi, writing from a Popperian epistemology of science, sees the same structure in venture-backed companies:
“A startup is best understood as a single conjecture embodied in an organization. Its central hypothesis is not the product design or the technology, but the existence of a viable market that will accept and sustain it. This conjecture is exposed directly to market selection, and refutation is terminal. When the hypothesis fails, the organization ceases to exist.”
The startup doesn’t optimize for learning. It can’t. It optimizes for survival narrative: the appearance of inevitability for investors and markets. And the result is that genuine discovery gets sacrificed to performative certainty:
“Markets are used not to test ideas, but to justify their continued existence. Speed, focus, and narrative coherence become survival strategies, not epistemic choices.”
“Many startup failures attributed to poor ideas are better understood as failures of epistemic timing. The conjecture may have been refutable cheaply in a local context, but the organizational form made such refutation impossible without collapse.”
Smith extends this into the labor economy itself. The “nerds”—the entire class of technical professionals whose human capital drove American prosperity for decades—built their fortunes on the assumption that expertise was irreplaceable. It wasn’t. It was routine:
“Software engineering was probably less of a ‘creative class’ job than we had allowed ourselves to believe, and more of a ‘routine cognitive’ task — the kind that’s especially vulnerable to automation.”
“A class of workers can only earn fantastic wealth for so long before inventors start looking for ways to replace their skills with machines that can be scaled up until the cost of those skills falls.”
And then there is Freya India, who traces the same pattern not through institutions but through persons. Her generation didn’t just live inside optimized systems; they became optimized systems:
“We became automated ourselves. We learned to speak like robots, think like robots, act like robots. We hid our opinions, for fear of upsetting anyone. We affected the agreeable robot voice, to accommodate everyone...So we played it safe, convinced we could not step out of line, could not have our own thoughts, could not feel too much.”
“Now here we are, many of us functioning like autocomplete, capable of thinking and saying only the most acceptable and predictable things.”
The overlap is hard to ignore. Science optimized citation counts until AI could produce citations faster than scientists could react. Startups optimized survival narratives until the narrative consumed the hypothesis. The tech economy optimized human capital until the capital was no longer human. A generation optimized social acceptability until they could no longer distinguish themselves from machines.
In every case, the proxy ate the purpose.
II. And the Point of Failure?
If the diagnosis is that our systems are optimized for the wrong thing, the deeper question is: why?
Ludanyi gives the most structural answer. The problem across all four domains is that failure has been made either lethal or invisible—but never informative:
“Independent startups: Refutation is clear but lethal. Learning is clean but lost. Embedded ventures: Refutation is survivable but often suppressed. Learning is possible but politically dangerous. Both fail...in different ways. The former kills the learner; the latter kills learning.”
This is the key for every article in the stack. In science, failure is punished by tenure clocks. In startups, failure is terminal. In the tech economy, human capital built over decades is erased overnight, and the knowledge that took a lifetime to accumulate vanishes with the career it supported. In India’s psychological landscape, being wrong—publicly, visibly, humanly—was treated as career-ending long before AI showed up.
Ludanyi puts it bluntly:
“The tragedy of the current system is that it forces learning to occur at the point of death, and then destroys the learner.”
That sentence should haunt anyone who fancies themselves an entrepreneur or leader. Once you see the pattern in institutions, it’s hard not to notice it in the institutions we are: families, churches, friendships, any place where being wrong becomes socially unaffordable.
Requarth sees the same architecture in the NIH grant system:
“Scientists spend roughly 45 percent of their time on administrative requirements rather than doing science; grant applications have ballooned from four pages in the 1950s to over one hundred today; and worst of all, the average age at which a scientist receives their first major independent grant is now 45.”
The system doesn’t just punish failure. It punishes risk. And the result is a culture of premature certainty — in science, in venture, in careers, in souls:
“Under these pressures, scientists inevitably pursue safe, incremental projects that will reliably yield papers, even if they never significantly advance understanding.”
Smith’s economic history adds a chilling coda. The entire “Revenge of the Nerds” technical expertise contained the seed of its own destruction:
“A generation of smart apes trained themselves to make their brains act like computers; it’s hardly surprising that computers eventually rose up and reclaimed their core competence.”
The most highly rewarded form of human intelligence turned out to be the most easily automated, precisely because it had been optimized into something machine-like. The system selected for the reproducible, the scalable, the legible. And now the machines are better at legibility than we are.
III. What Would It Take?
So what kind of system would make failure survivable and informative? What would it look like to build institutions—in science, in venture, in education, in the formation of actual human beings—where being wrong doesn’t end you and being right doesn’t require performing certainty?
Ludanyi sketches the design specs:
“A genuinely Popperian organization would need to: Internalize variation to reduce the cost of refutation. Isolate projects to prevent political contagion. Reward termination as evidence of learning. Separate epistemic success from narrative success. Preserve knowledge independently of project survival.”
That list is extraordinary, and not just for epistemology. “Reward termination as evidence of learning.” “Separate epistemic success from narrative success.” These are not merely organizational design principles. They describe a posture toward reality that most of us, and most of our institutions, have systematically abandoned.
Some traditions call it antifragility; others call it a learning culture. Christians have an older word for it: grace—the capacity to absorb failure without being annihilated, and to let it teach rather than merely scar.
Whatever you call it, the four articles on my desk this week converge on its absence. The reason AI is exposing so much hollowness across so many domains simultaneously is that these domains had already been hollowed out. The optimization was the hollowing. And the technology that does the optimized thing better than we do simply makes the emptiness impossible to ignore.
Freya India, writing from the most personal register of the four, lands on something that sounds like a lifestyle manifesto but is actually closer to institutional theology:
“When so few seem interested in being a person, isn’t that the best time to be one?”
“I want to start sentences that can’t be autocompleted because even I don’t know where they’re going.”
“I want to get things wrong and apologize and sometimes I want my opinions to be contradictory and incomplete because I am feeling my way through this world and I am not a machine with all the answers.”
That’s not just personal courage. That’s a design principle. It’s what Ludanyi’s “corrigible enterprise” looks like when it’s not an organization but a person. A person who has decided that being refutable is more valuable than being optimized.
I think the honest conclusion is this: The AI moment isn’t primarily a technology story. Maybe it’s a stress test. It made the dehumanizing, hollow substitution obvious:
production posing as knowledge
narrative posing as learning
routine posing as expertise
politeness posing as character
So the question isn’t simply whether AI will replace us. It’s whether we can rebuild environments—labs, companies, careers, inner lives—where being wrong is survivable, where risk is rewarded, where knowledge outlives the project, and where a human being can speak without sounding like autocomplete. That might be the scarcest asset now.
If these convergences interest you, subscribe. Endwise runs at the intersection of theology, technology, and whatever is actually real.