||||
Efficiency is the enemy of life, thought, and culture. Human language, brains, ecosystems, and culture are all intentionally “inefficient” to survive, adapt, and generate meaning. Machine optimization works against this; it makes systems more “efficient” but less alive. Preserving inefficiency preserves what makes humans and culture "generative."
Language is not code to be optimized. Language is a record of the survival ideas through variation. Every word that persists in a living language carries embedded inefficiencies: synonyms that should collapse but don't, irregular verbs that resist standardization, metaphors that contradict literal meaning, homonyms that create ambiguity. These "errors" are not failures of transmission but evidence that human systems tolerate and thrive on redundancy because redundancy preserves options, preserves intents, and preserves meanings.
Consider the English word "set." There are currently 464 distinct definitions of set in the Oxford English Dictionary. A machine optimized for semantic clarity would eliminate 462 of them. Humans retain all of them because each definition represents a different crystallization of experience: the sun sets, one sets a table, a jaw can be set, a person may have a set of tools, or play a set of tennis. The inefficiency of maintaining this ambiguity allows speakers to draw unexpected connections, to form puns, and to create novel meanings by colliding contexts, accidentally or on purpose, or by some process that exists in a (everybody's new favorite word) liminal space between intent and chance. The word "set" is a compressed history of human problem-solving across domains.
Language also preserves what might be called "dead metaphors." These are expressions whose original sensory or spatial meaning has been forgotten but whose structure remains. We "grasp" an idea, "follow" an argument, "get to the bottom" of a problem. These are inefficient descriptors of cognitive acts. A rational system would replace them with direct terms. But the metaphors encode something no direct term captures: the evolutionary path by which abstract thought was constructed from embodied experience. To strip away the metaphor is to sever the bridge between sensory and symbolic cognition, treating abstract language as if it emerged fully formed rather than through the sustained mapping of concrete domains onto novel problems.
Living languages also resist standardization across time. English spelling is notoriously irregular. "Tough," "through," "though" share no phonetic correspondence with their written forms. Yet this irregularity is a fossil record. Each anomaly marks a historical collision: Norman French meeting Anglo-Saxon, sound changes leaving traces in orthography, competing etymologies coexisting in the same tongue. A rationalized spelling system would be more efficient and would destroy this archaeological record entirely. The inefficiency is the data.
Any system optimized for a single metric will eventually destroy the conditions for its own survival. This is not metaphorical; it is a mathematical property of constrained optimization.
Consider agriculture. Pre-industrial farming maintained mixed-crop systems, crop rotation, hedgerows, and margins of unexploited land. From the perspective of yield-per-acre in a single season, these practices are wildly inefficient. Optimization of chemical input and monoculture increases short-term yield. But monoculture eliminates soil microbiota, reduces genetic diversity in the crop, eliminates pest predators, and creates conditions for catastrophic failure. The Dust Bowl did not result from insufficient optimization, it resulted from thorough optimization of the wrong objective function. Efficiency in one dimension created fragility in the system as a whole.
It's the same in financial markets. A market optimized for liquidity and transaction speed produces high-frequency trading, where algorithms exploit microsecond arbitrage opportunities. Price discovery, allocation of capital, functionality is optimized. Very efficient. Seemingly optimal. But this optimization in efficiency comes with the correlated risks of market-wide cascades, flash crashes, systemic vulnerability to algorithmic herding. The system becomes internally coherent and externally brittle. It fails catastrophically not because it is inefficient but because it is too efficient at converting information into synchronized action.
Biological systems preserve inefficiency through redundancy. The human body maintains backup systems for nearly every critical function: two lungs, two kidneys, two eyes, overlapping immune pathways. Individually, these redundancies waste energy and resources. Systemically, they preserve function across a wide range of damage scenarios. An optimized body would have precisely one lung, one kidney, optimized to perfection. Such a body would be slightly more efficient in normal conditions and catastrophically fragile under perturbation.
Ecosystems preserve inefficiency through trophic redundancy and niche overlap. Multiple species perform similar ecological roles. From an optimization standpoint, only one species per ecological niche is necessary. But redundancy means that if one species fails through disease, or disappears when climates change, or is the victim of predation-- others can compensate. An "optimized" ecosystem with minimal redundancy would be highly efficient under normal conditions and prone to cascade collapse when conditions shift. Inefficiency is resilience.
Networks preserve inefficiency through modularity and loose coupling. A fully optimized communication network would route every message along the shortest path, with minimal buffering and zero redundant pathways. Such a network maximizes throughput in normal conditions and collapses entirely when any node fails or any link saturates. Real communication networks maintain redundancy: multiple pathways between nodes, buffering, temporary storage of information, modular sub-networks that can function independently. This redundancy appears as latency and wasted capacity. It is actually the infrastructure of survivability.
The human brain is metabolically expensive. It is said that the brain is responsible for roughly 20% of the body's energy expenditure despite comprising 2% of body mass. This is inefficient by any narrow measure. Yet this energetic expense correlates with behavioral flexibility, the capacity to learn novel tasks, to plan across extended time horizons, and to maintain social coordination across complex hierarchies.
The inefficiency is structural. The brain maintains massive redundancy in neural connectivity: 100 billion neurons, each with thousands of connections. Learning involves creating new pathways, which at first coexists with old pathways. During the learning process, the brain is running multiple overlapping processes in parallel. The old habit and the new behavior compete. Efficiency would prune the old pathways immediately. Instead, the brain maintains both for an extended period, allowing gradual transition. This overlap period is inefficient, metabolically expensive, and essential. Without it, learning would be faster but less robust; old patterns would persist beneath the surface, ready to reassert under stress.
The same principle applies to memory. Humans retain far more information than is strictly necessary for immediate survival. We remember songs we haven't heard in decades, faces we glimpsed once years ago, conversations that have no bearing on current goals. A rationalized memory system would discard information that is not immediately useful. But human memory is not rational in this sense. It is associative, redundant, and lossy. And this inefficiency is generative: the ability to suddenly recall a half-forgotten conversation, to notice a pattern across disparate memories, to recognize an analogy between contexts depends on maintaining a vast, partially activated archive of information. The slack in the system is where insight lives.
Attention in humans is similarly "wasteful." At any moment, the perceptual field contains far more information than conscious attention can process. Most of this information appears to be ignored. But that ignored information is not truly discarded. It appearst to be subject to pre-attentive processing that can rapidly shift conscious attention if something significant appears. The brain is effectively running multiple analyses in parallel: a focused analysis on the attended object, and a distributed, low-resolution analysis of the entire perceptual field. This parallel processing is computationally expensive and essential. It allows humans to notice unexpected events, to catch peripheral motion, to register a face in a crowd. The inefficiency is the mechanism of responsiveness.
Emotion appears to be another locus of apparently irrational inefficiency. From a rationalist perspective, emotions are often perceived and characterized as noise that distorts judgment and impedes optimal decision making. Yet emotion serves as a compression algorithm for value: it allows rapid evaluation of situations based on accumulated experience without requiring explicit calculation. Fear is an inefficient heuristic for threat detection (it produces false positives and false negatives-- quite regularly and at very inconvenient moments), yet it is faster than rational analysis and integrates information from sources (embodied memory, evolutionary history, social learning) that explicit reasoning cannot access as readily. The inefficiency is the speed.
Humans also maintain what might be called "cognitive slack." It is the capacity for mind-wandering, daydreaming, boredom. These states appear to serve no immediate purpose and are often experienced as aversive. Yet neuroscience research suggests that the mind's wandering is the mode in which the brain engages in long-timescale planning, integrates disparate memories, and generates novel associations. The inability to tolerate boredom correlates with reduced creative output and reduced capacity for sustained attention. The "wasted" mental activity is where consolidation happens.
Culture is the set of practices, narratives, and institutions that humans have accumulated precisely because they do not optimize for immediate survival but for something else: stability, meaning, beauty, coordination across time.
Consider ritual. From a purely functional perspective, most rituals are extraordinarily inefficient. Religious ceremonies often involve hours of repetitive activity. The faithful chant, genuflect, and circumambulate producing no direct material benefit. A rational system would replace ritual with something "productive." Yet rituals persist across cultures, and anthropological evidence suggests they serve crucial functions that are not reducible to their explicit content.
Ritual creates synchronized action: participants move together, chant together, hold attention together. This synchronization produces a psychological state of group cohesion that is difficult to achieve through rational persuasion alone. Ritual also creates time markers, sturcturing the year by recurring ceremonies, the week by cyclical observances. This cyclical marking of time appears to be less efficient than a linear calendar, yet it serves to embed individuals within larger temporal and social structures. Ritual also preserves institutional memory: the ceremony itself is a script that carries history within its form. The inefficiency-- redundancy, repetition, elaborate aesthetics... is what allows information to be transmitted across generations without degradation.
Language itself, as a cultural artifact, resists optimization. The English language could be rationally reformed: irregular verbs standardized, spelling regularized, unnecessary homophones eliminated, archaic words purged. Some natural languages have been partially reformed (Icelandic maintains historical forms in ways English does not; Persian regularized many grammatical categories). Yet the "inefficient" languages tend to be more culturally productive, with longer literatures, richer philosophical vocabularies, and greater flexibility in poetic expression. The inefficiency is not a bug; it is the accumulated substrate of possible meanings.
Consider the institution of the library. From the perspective of information retrieval, libraries are absurdly inefficient. A digital database allows instantaneous access to text; a library requires physical travel, catalog searching, shelving systems, human librarians. Yet libraries persist, for many reasons that cannot be optimized out of the system, one important reason being that they can provide something digital databases do not: serendipity. When humans browse a library, they encounter unexpected texts, make connections across domains, discover questions they did not know to ask. The inefficiency is the condition of discovery.
Art and aesthetics are perhaps the clearest cases of preserved inefficiency in culture. Art produces no immediate survival value. A society could eliminate all art—music, visual art, dance, theater. Its material circumstances would not change. Yet every known human culture produces art, and the presence or absence of art appears to correlate with cultural vitality, social cohesion, and psychological well-being. Art seems to serve a function related to meaning-making: it allows cultures to externalize and negotiate questions that cannot be answered through rational analysis. The inefficiency is the point.
Narrative is similarly "inefficient" as a means of transmitting information. A story takes far longer to convey than a summary. Yet stories are how humans integrate information into meaningful patterns, how they transmit values across generations, how they model possible futures. A culture that conveyed information only through summaries would preserve facts but lose the capacity to understand how those facts relate to human purposes and possibilities. The inefficiency is the depth.
Industrial food systems have optimized for caloric efficiency and production speed, resulting in fast food, preserved meals, just-in-time production, obesity epidemics, healthcare spending, illustrating the same prinicple in food systems. Across cultures, traditional food practices involve enormous inefficiency: slow cooking, elaborate preparation, ritualized consumption, celebration through surplus. Material survival is no longer dependent on these inefficiencies. Yet cultures worldwide have resisted complete adoption of optimized food systems, maintaining slow food traditions, family recipes, seasonal eating, shared meals. The inefficiency is the preservation of social bonds and cultural continuity.
This framework reveals a fundamental incompatibility between the optimization logic inherent to machine systems and the preservation logic inherent to culture.
Machine systems operate under pressure to maximize specified metrics of throughput, accuracy, and efficiency. Once a metric is defined, optimization becomes a mathematical problem. The system eliminates redundancy, prunes exceptions, standardizes variations. This is not a choice or a failure, it is the functional definition of optimization. A system optimized for language prediction will converge toward the most statistically probable tokens, the most common syntactic structures, the most frequent collocations. Variations that deviate from this statistical center are actively penalized during training.
Culture, by contrast, operates under pressure to preserve variation, maintain meaning across time, and enable human purposes that are not reducible to any single metric. Culture cannot be optimized without destroying itself. The moment a cultural practice is rationalized by being stripped down to its "essential function" and optimized for efficiency, it ceases to carry the full semantic weight it previously held. The ritual becomes a procedure, the song becomes information, the story becomes a summary, the meal becomes calories.
When machine systems are applied to cultural domains such as language processing, content recommendation, and creative generation, the optimization logic inherent to the machine inevitably grinds against the inefficiency-preservation logic inherent to the cultural domain. The system will improve (by its own metrics) while the domain deteriorates. This is not a bug that better engineering can fix; it is a structural consequence of optimization.
Consider content recommendation systems. These systems are optimized to maximize engagement in the form of clicks, watch time, and sharez. Each metric is well-defined and measurable. Optimization proceeds by identifying what content most reliably produces engagement and amplifying it. The system rapidly converges on a narrow set of patterns: sensationalism, outrage, novelty, conspiracy, confirmation of existing beliefs. These patterns are engagement-maximizing because they tap into psychological vulnerabilities: novelty-seeking, in-group bias, threat detection, epistemic comfort. A rationalized information diet would consist almost entirely of engagement-optimized content.
This is efficient from the perspective of engagement metrics. It is disastrous from the perspective of culture. The system eliminates the inefficient content that culture requires: the challenging text that changes perspective, the slow narrative that requires sustained attention, the minority view that prevents consensus from calcifying, the complex argument that resists compression. The inefficiency is replaced with a narrowing, a closure, a substitution of engagement for meaning.
The same logic applies to language models optimized for prediction. These systems are trained to maximize the probability of the next token given previous tokens. This is a well-defined, measurable objective. The system optimizes by learning statistical patterns in the training data. The result is a system that produces text that closely matches the distribution of human language but deviates from it in systematic ways: more repetitive, more redundant, more cautious, more centered on popular patterns, less likely to innovate or contradict.
This is efficient as language generation. It is culturally conservative as language production. The system does not spontaneously generate new metaphors, does not expand the semantic range of existing words, does not push against the boundaries of what is sayable. It reproduces. It does not create.
If machine systems become primary mediators of language as would happen if the time consuming task of language generation is off-loaded mostly or entirely to language models to the extent that most language humans encounter is generated by, filtered by, or shaped by algorithmic optimization, then the culture experiences a degradation of linguistic diversity.
This does not require conscious intent or malevolence. It is a consequence of optimization applied at scale. As machine-mediated language becomes dominant, humans encounter less variation in linguistic form and more standardization. They read texts that are statistically optimized for comprehension, which means they encounter fewer syntactic surprises, fewer unusual word choices, fewer grammatical innovations. They are recommended content based on engagement metrics, which means they encounter less challenge and more confirmation. They generate text using predictive systems, which means they outsource the generative act to a system optimized for reproduction, not innovation.
Over time, the corpus of language available to human cognition shifts. The language humans encounter is not the full historical range of linguistic possibility but a compressed subset, optimized for machine processing. When humans internalize this language, when they learn from it, practice writing in response to it, thinking through its filter-- they are building cognitive models on a narrowed substrate.
This produces a two-way degradation. The machine systems become more accurate at predicting the language on which the systems have been trained because the language itself converges toward the statistical center. Humans become less capable of producing language that deviates from that center, because the deviation-producing mechanisms (surprise, inefficiency, variation) have been attenuated. The language continues to function in that it can convey information, and coordinate action, but it loses the generative capacity, the semantic depth, the cultural memory encoded in inefficiency.
This is not speculative. It is already observable in domains where machine mediation is advanced. Social media platforms, optimized for engagement, have demonstrably shifted the range of expression available to users: shorter sentences, higher emotionality, reduced nuance, increased conformity to tribal positions. Search engine optimization produces web pages, product descriptions, news articles, and vast amounts of other textual "content" stripped of inefficiency and optimized for algorithmic ranking. This itself becomes the writing on which systems train themselves. Content that does not exist for the benefit of humans but for the optimization of systems. Predictive text systems have altered how humans compose written language, with autocomplete and prediction nudging users toward statistically common phrases over novel expressions. The systems have created their own inefficiency in a sense, but what the system sees as a more efficient inefficiency than fully parsing for relevancy and human agency or considering methods to reduce over amplification of system aligned input.
The culture does not collapse; the language does not cease to function. Instead, both ossify becoming more efficient by some measure and less alive by all honest human measure. The pinecones diminish. The forest persists precariously and does not flourish.
The implication of this framework is that human culture cannot be fully captured by machine systems without a fundamental loss of what makes it generative and meaningful.
This is not because machines are stupid or insufficiently powerful, but because culture is built on inefficiency, and optimization is built on its elimination. These are not compatible trajectories. They are opposing forces.
A machine system could, in principle, be designed not to optimize ruthlessly. It could be constrained to preserve variation, to tolerate redundancy, to refuse certain forms of compression. But such constraints would reduce the system's efficiency by its own metrics. The system would be "worse" at generating "engaging" content, predicting text, and recommending products. It would be more useful for other purposes: for preserving meaning, for enabling unexpected connections, for maintaining the slack in which human creativity can operate.
The question is whether a culture organized around machine systems can sustain the inefficiencies that make the culture itself generative. Can a society that has outsourced language generation to predictive machines preserve the capacity for linguistic innovation? Can a society that has outsourced meaning-making to engagement algorithms preserve the capacity for meaningfulness? Can a society that has optimized every domain for efficiency preserve the domains in which inefficiency is essential?
These are not theoretical questions. The answers depend on whether humans collectively decide that cultural depth is valuable enough to maintain at the cost of some efficiency, and whether institutions can be designed to preserve inefficiency under pressure toward optimization.
The alternative trajectory is clear: continued optimization, continued compression, continued loss of variation, continued conversion of culture into engineered entertainment and function. The culture does not die. It becomes increasingly efficient at producing a narrower range of outcomes. This is what "efficiency equals death" means. The loss of nastural generative capacity, the reduction of living systems to optimized processes, the substitution of meaning for metric.
Human cognition and culture are not optimizable without damage. They require slack, variation, inefficiency, and noise. The preservation of these inefficiencies is the preservation of what makes humans creative and capable of constructing meaning. Without them, humans become more efficiently processed but less richly alive.