Abstract
Artificial intelligence can greatly exceed human performance in selected domains while remaining unreliable in adjacent tasks that humans handle routinely. This uneven profile has been described as a jagged technological frontier. The present article develops a dynamic extension of that concept and applies it to labor substitution, recursive self-improvement, and the possible development of artificial superintelligence. The central proposal is that autonomous labor substitution depends more strongly on capability coverage than on average capability or peak performance. Occupations consist of interdependent tasks and competencies, so a small number of below-threshold abilities can preserve the need for a human worker even when an AI system possesses enormous superhuman advantages elsewhere. During this period, superhuman capabilities may accumulate as a substitution overhang that cannot yet be converted into full occupational replacement. Once the remaining bottlenecks cross the relevant thresholds, displacement may accelerate discontinuously, and the artificial system entering the workplace may already be far more capable than the human it replaces.
The article introduces recursive capability gap closure as the process through which an AI system identifies, diagnoses, and selectively improves deficiencies in its own capability profile, with each successful improvement increasing its ability to perform subsequent improvements. It further introduces capability frontier smoothing as the long-term reduction of consequential capability deficits relative to a system’s own objectives and stronger capacities. After the human capability envelope has been surpassed, continued smoothing may be favored by task complementarity, reliability requirements, adversarial vulnerability, evaluator limitations, environmental uncertainty, and the need to convert theoretical insight into effective action. Capability frontier smoothing is therefore conceptualized as an ongoing, constraint-weighted process rather than a terminal state of uniform ability. The framework predicts a substantial lag between visible AI intelligence and labor displacement, followed by clustered substitution when shared bottlenecks are removed, and continued self-relative capability development long after human-level generality has been achieved.
Keywords: artificial general intelligence, artificial superintelligence, jagged technological frontier, recursive self-improvement, labor automation, capability coverage, technological unemployment, AI research automation
1. Introduction
Modern artificial intelligence presents an apparent paradox. AI systems can produce expert-level analyses, write software, search enormous possibility spaces, synthesize scientific literature, and discover useful algorithms, yet they may still misunderstand a simple instruction, lose track of an ongoing project, fail to recognize an anomalous situation, or confidently produce an answer containing an obvious error. Performance can change sharply between tasks that appear equally difficult to a human observer. In a field experiment involving 758 consultants, Dell’Acqua and colleagues found that GPT-4 assistance improved speed, productivity, and quality on tasks within the system’s capability frontier, while reducing the probability of a correct answer on a task deliberately placed outside that frontier. They characterized this unevenness as the jagged technological frontier.
The jagged frontier helps explain why impressive demonstrations of AI capability have not translated directly into the autonomous replacement of most human workers. Occupations are not unitary operations. They are collections of tasks that draw on memory, judgment, perception, communication, exception handling, coordination, domain knowledge, physical action, and accountability. Machine learning may be suitable for some tasks within nearly every occupation while remaining unsuitable for enough essential tasks to prevent complete automation. Earlier task-based analyses consequently found that many occupations contained machine-learnable components, while few could be automated in their entirety without redesigning the job itself.
Recent economic theory makes the bottleneck structure especially explicit. Gans and Goldfarb’s model of O-ring automation treats occupational tasks as quality complements whose contributions multiply rather than simply add. Under these conditions, smooth improvement in automation quality can produce discrete adoption decisions, and average task exposure can substantially overstate displacement. Automation of one task also changes the value of automating the remaining tasks, so the economic consequence depends on the structure of the bundle rather than the number of automated components alone.
The present article extends these insights into a developmental hypothesis about advanced AI. AI may continue to become dramatically stronger in capabilities it already performs well while remaining below human adequacy in a small number of consequential dimensions. Its aggregate performance may become vastly superhuman, yet its incomplete coverage may prevent it from serving as a reliable, autonomous replacement for a human worker. The economically important transition will occur when the last indispensable deficiencies cross the relevant thresholds.
This creates the possibility of a superhuman substitution lag. During this interval, AI possesses large quantities of capability that exceed what a job requires, but those capabilities remain partially trapped behind unresolved bottlenecks. The resulting substitution overhang can become very large. When the bottlenecks are eventually removed, the system that becomes capable of replacing a person may bear little resemblance to a merely human-equivalent employee. It may already be able to perform its strongest functions at speeds, scales, and levels of coordination far beyond human capacity.
The same process can continue after human-level generality has been reached. An advanced AI would no longer need to compare itself primarily with humans. Its relevant weaknesses would become self-relative capability gaps, meaning areas that lag behind the requirements created by its own stronger capacities and expanding objectives. A system capable of deliberately identifying and reducing such gaps would exhibit recursive capability gap closure. The long-run result could be capability frontier smoothing, an ongoing tendency to eliminate weaknesses that constrain the practical use of stronger abilities.
2. Capability Coverage and Occupational Substitution
A capability frontier can be represented conceptually as a surface in a high-dimensional space. Each direction corresponds to a capability, task class, environmental condition, or combination of demands. The human frontier is relatively broad because humans possess general sensorimotor competence, social understanding, contextual adaptation, and extensive experience operating in open-ended environments. Human performance is far from uniform, but the ordinary adult capability profile contains comparatively few catastrophic gaps within the environments for which human occupations were designed.
An AI capability frontier may have a very different geometry. Some regions can extend far beyond human performance while others remain far inside the human envelope. These indentations may include deficiencies in persistent memory, long-horizon planning, error recovery, causal understanding, social interpretation, physical manipulation, uncertainty calibration, or recognition of situations that fall outside a learned pattern. The exact deficits will change as systems develop, but their structural role remains the same. A job can remain human-dependent whenever an unresolved indentation intersects an indispensable part of the job’s task structure.
Let an occupation j require a set of capabilities K_j. For each required capability i, let a_i(t) represent the AI system’s effective performance at time t, and let r_{ij} represent the minimum performance required for autonomous execution of the occupation. Effective performance should incorporate accuracy, reliability, cost, latency, resource use, and the ability to recover from errors rather than benchmark accuracy alone. A normalized capability ratio can then be defined as:
z_{ij}(t)=\frac{a_i(t)}{r_{ij}}.
For a highly complementary occupation, a simple bottleneck index is:
B_j(t)=\min_{i\in K_j}z_{ij}(t).
When B_j(t)<1, at least one indispensable capability remains below threshold. The AI may assist the worker, accelerate selected tasks, or permit organizational restructuring, but it cannot yet perform the complete role at the required standard without compensating human labor. When B_j(t)\geq1, the technical capability coverage threshold has been reached, although actual adoption may still depend on cost, legal permission, public acceptance, physical infrastructure, and organizational change.
This minimum function is an intentionally strict representation. Some occupational tasks are substitutable, and organizations can remove, redistribute, or redesign them. Other tasks are strongly complementary, so omission or poor execution of one component compromises the entire product. Constant-elasticity and O-ring models allow this degree of complementarity to vary. Analyses of real AI usage likewise find that productivity estimates fall considerably when task reliability and within-occupation bottlenecks are incorporated.
Peak performance and capability coverage can therefore diverge. Define a simple measure of job-relative capability surplus as:
S_j(t)=\sum_{i\in K_j}w_{ij}\max\left[0,z_{ij}(t)-1\right],
where w_{ij} represents the importance of capability i. An AI can possess a very large S_j(t) while B_j(t) remains below one. It may be a hundred times better than a worker at research, drafting, calculation, and retrieval while remaining insufficiently reliable at recognizing exceptions or coordinating a long sequence of actions. Further growth in the already-superhuman dimensions raises the surplus without necessarily producing autonomous substitutability.
Reliability intensifies this effect. Under a simplified independent-error model, if a workflow contains 100 dependent operations and the AI succeeds on each with probability 0.99, the probability of completing the whole sequence without an error is approximately 0.99^{100}=0.366. Real workflows include redundancy, checkpoints, correlated errors, and opportunities for recovery, but the example illustrates why seemingly high task-level accuracy may remain inadequate for unsupervised work. A human monitor can identify and repair failures, although the need for that monitor means the labor has been augmented or reorganized rather than fully replaced.
Empirical research already distinguishes task exposure from occupational automation. Eloundou and colleagues estimated that a large proportion of workers could have at least some tasks affected by language models, while a substantially smaller proportion had a majority of tasks exposed. Productivity experiments have similarly found meaningful gains from AI assistance in writing and customer support without demonstrating end-to-end elimination of the underlying occupations.
Current usage patterns reinforce the distinction. A 2026 analysis combining theoretical capability estimates with observed workplace use found that actual AI coverage remained far below theoretical exposure across many occupations. It found no systematic increase in unemployment among the most exposed workers at that stage, although there was tentative evidence of reduced entry into exposed occupations among younger workers. Legal constraints, verification requirements, software integration, and deployment friction contributed to the gap between what appeared technically feasible and what was actually automated.
3. Substitution Overhang and the Nonlinear Labor Transition
When AI progress is jagged, capability growth and labor displacement need not move proportionally. Improvements to an already-superhuman capability may increase the quality or speed of assisted work without changing whether a human must remain involved. Improvements to a single bottleneck capability can have a much larger economic effect, even when the measured increase in aggregate intelligence is comparatively small.
Suppose an occupation depends on 20 complementary capabilities. An AI system performs 18 of them at ten times the required level, one at twice the required level, and the final capability at 80 percent of the required level. Increasing the first 19 capabilities produces additional capability surplus but leaves the bottleneck index below one. Increasing the final capability from 0.8 to 1.1 changes the status of the entire occupational bundle. The system crosses from incomplete assistance to technical coverage even though the final improvement represents only a small fraction of its total capability.
The substitution overhang is the accumulated stock of superhuman capability whose full economic deployment is constrained by one or more unresolved gaps. It is an overhang because the latent effect on labor can be much larger than the final capability increment that releases it. The greater the AI’s preexisting superiority in other dimensions, the more consequential the removal of the last bottleneck becomes.
The first complete artificial replacement for a worker may consequently be massively overqualified. There may be little or no prolonged period during which artificial employees resemble ordinary humans in their overall capability. The transition could instead move from systems that are extraordinarily powerful but require supervision to systems that are both autonomous and vastly superhuman. The boundary between these conditions is defined by coverage, reliability, and deployability rather than by a smooth average of benchmark scores.
Occupations may also share bottlenecks. Reliable long-horizon planning, persistent memory, autonomous computer use, error recovery, physical dexterity, or social-context recognition may participate in thousands of distinct jobs. When one broadly reused bottleneck is removed, multiple occupational coverage thresholds can be crossed within a short interval. Labor displacement would then arrive in clusters rather than as an independent, gradual process for every occupation.
This hypothesis does not require AI to become better than humans in every possible human ability. It requires coverage of the capabilities that are indispensable within a specific production process. A system replacing an accountant does not need the full sensorimotor repertoire of a firefighter, and a system replacing a software engineer does not need every ability involved in childcare. The human capability envelope is therefore better represented as a collection of overlapping occupational sub-envelopes.
Organizations can also route around deficiencies by changing workflows. A remaining human task can be concentrated among a smaller number of supervisors, transferred to a specialized machine, removed from the product, or rendered unnecessary by a new business process. Job redesign can therefore cause effective coverage even before the original human role is reproduced in full. The central claim remains intact: autonomous substitution occurs when the AI-centered production system covers every binding requirement, whether through direct replication or structural circumvention.
4. Why Capability Spikes May Expand Before Gaps Close
Capabilities are unlikely to improve at identical rates. Domains differ in training data, evaluator quality, feedback speed, computational tractability, and economic value. Programming, games, formal mathematics, and optimization often provide rapid or automated feedback. Social interpretation, open-ended physical action, institutional judgment, and evaluation of long-horizon consequences may provide sparse, delayed, ambiguous, or contested feedback.
These differences can initially amplify jaggedness. When a system already performs well in a domain with reliable evaluation, additional compute and automated search can extend that capability rapidly. Weak domains may remain weak when failures are difficult to detect, performance is expensive to test, or there is no agreed-upon objective function. The tallest spikes can therefore continue growing while some inward regions move outward slowly.
Automated search systems illustrate the advantage of clear evaluators. FunSearch paired a language model with a systematic evaluator and discovered improved solutions to mathematical problems. AlphaEvolve uses language models, evolutionary search, and automated evaluation to improve algorithms, including components of computational infrastructure and an algorithm involved in training the model that powered the system. These systems operate within deliberately constructed problem spaces, but they show how strong generation combined with reliable evaluation can produce sustained optimization.
A lagging capability may begin to improve rapidly once an appropriate representation, training environment, or evaluator is developed. The transition from an obscure failure mode to a measurable optimization target can alter its rate of progress. A capability that remained at 70 or 80 percent of human adequacy for an extended period could cross the human threshold and continue toward superhuman performance after the underlying impediment is understood.
This produces a potentially asymmetric developmental sequence. The capabilities that become superhuman first have the longest period in which to compound. The capabilities that become adequate last determine the time of autonomous substitution. By the time the lagging dimensions reach human sufficiency, the early dimensions may have advanced several technological generations beyond it.
5. Recursive Capability Gap Closure
Recursive capability gap closure is the process through which an AI system identifies deficiencies within its own capability profile, determines how those deficiencies arise, develops and tests interventions, and incorporates successful changes in ways that improve its ability to conduct subsequent rounds of capability development. It is narrower than recursive self-improvement as a general concept. Recursive self-improvement can increase any aspect of performance, while recursive capability gap closure specifically directs improvement toward consequential indentations.
The first component is capability cartography. A system must construct a sufficiently accurate model of what it can do, where it fails, how performance varies across conditions, and which deficiencies constrain important objectives. This requires evaluations that measure transfer, reliability, recovery, resource use, and performance under distributional change. A single aggregate score is poorly suited to this purpose because averaging can conceal precisely the weaknesses that gap closure is intended to find.
The second component is causal diagnosis. Observable failure does not reveal which part of a system is responsible. A memory failure could arise from limited context, poor retrieval, interference, inadequate task decomposition, incorrect prioritization, or failure to recognize that information should have been stored. Effective gap closure requires distinguishing symptoms from mechanisms and identifying interventions that generalize beyond a benchmark.
The third component is intervention generation. The AI may propose changes to its software, training curriculum, model architecture, inference procedure, memory system, tools, data pipeline, evaluation suite, or allocation of computational resources. Some improvements may occur within the model, while others may result from scaffolding and external components. From the standpoint of labor substitution, the system-level capability is more important than whether the improvement resides inside a single neural network.
The fourth component is empirical validation. Candidate changes must be compared with prior versions across multiple tasks and environments. Improvements in one dimension can introduce regressions elsewhere, and optimization against a weak evaluator can produce apparent progress without real capability. Robust gap closure therefore depends on adversarial testing, independent evaluations, causal analysis, and mechanisms for reverting harmful modifications.
The fifth component is recursive integration. The process becomes genuinely recursive when an improvement increases the system’s ability to perform one or more stages of the next improvement cycle. Better coding can produce better tools, better experimental design can generate more informative data, better evaluation can expose subtler weaknesses, and better planning can coordinate larger research programs. The system’s AI research capability then becomes both an instrument and an object of optimization.
Theoretical accounts of self-improving systems long predate current language models. Good proposed that a sufficiently capable machine could contribute to the design of better machines, creating a positive feedback process. Schmidhuber’s Gödel machine formalized a self-referential system that could rewrite its own code after establishing that the modification would improve its expected utility. These frameworks established the logical possibility of recursive improvement, although their assumptions differ greatly from practical machine-learning systems.
Recent work implements limited fragments of the loop. The Darwin Gödel Machine iteratively modifies its own coding-agent software, evaluates variants on coding benchmarks, and preserves useful modifications as stepping stones for subsequent search. Its authors explicitly distinguish this proof of concept from present-day systems that autonomously and continuously redesign themselves. AlphaEvolve similarly automates proposal, evaluation, and iterative refinement in domains where candidate programs can be tested objectively.
AI research evaluations show both meaningful capability and continuing jaggedness. In RE-Bench, leading agents performed strongly relative to human experts under short computational budgets, but human performance scaled better with longer budgets. Research on autonomous task horizons also attributes recent progress partly to improved reliability, tool use, error adaptation, and logical reasoning, while emphasizing uncertainty about how results generalize beyond the tested software environments.
These findings do not establish open-ended recursive self-improvement. Current systems still depend heavily on human-defined objectives, evaluators, compute allocation, research environments, and underlying models. They nevertheless demonstrate increasingly complete portions of a feedback loop in which AI generates modifications, tests them, retains successful variants, and improves processes connected to future AI development.
Intentionality in this framework does not require phenomenal consciousness or a human-like desire for self-improvement. It refers to goal-directed selection. A system exhibits intentional capability gap closure when its planning process explicitly prioritizes deficiencies because improving them is predicted to increase expected objective attainment. The behavior can be instrumentally rational without implying subjective experience.
6. Capability Frontier Smoothing
Capability frontier smoothing is the deliberate reduction of capability deficits that constrain the effective use of an intelligent system’s stronger capacities. Smoothing does not require all capabilities to become equal. An advanced AI may rationally invest far more heavily in some domains than others, preserving pronounced spikes wherever specialization produces high returns. The process targets consequential indentations rather than geometric irregularity for its own sake.
Let d_i(t) represent the level of capability i demanded by the system’s objectives, environment, and complementary capacities. A self-relative gap can be defined as:
g_i(t)=\max[0,d_i(t)-a_i(t)].
A constraint-weighted measure of consequential jaggedness is then:
G(t)=\sum_i\lambda_i(t)g_i(t)^2,
where \lambda_i(t) represents the importance or shadow cost of the deficit. Capability frontier smoothing reduces G(t). Raw variance across capabilities may remain high or even increase while the capability profile becomes smoother relative to the requirements that actually constrain performance.
6.1 Complementarity and the weakest-link constraint
The strongest reason to reduce consequential jaggedness is complementarity. A powerful capability often creates value only when supported by other capabilities. An AI capable of extraordinary theoretical physics may be limited by weaker experimental design, simulation, instrumentation, robotics, manufacturing, or data interpretation. Additional theoretical insight has diminishing practical value when physical verification becomes the binding constraint.
The relevant comparison therefore changes after human-level performance has been surpassed. A capability that is one thousand times better than human performance can remain an enormous weakness if complementary capabilities are one million times better. Human superiority is no longer the benchmark. The economically and strategically relevant question concerns how much stronger the system could be if the lagging component no longer restricted the rest of the architecture.
6.2 Reliability and long-horizon coherence
Jaggedness becomes increasingly costly as projects lengthen. A system can possess exceptional local reasoning while repeatedly losing goals, corrupting state, overlooking anomalies, or failing to recover from unexpected outcomes. These weaknesses may have little effect on short benchmarks and severe effects on scientific programs, institutional management, engineering projects, or autonomous operation lasting months or years.
Long-horizon capability requires a collection of mutually supporting abilities: planning, memory, monitoring, revision, tool use, resource management, and error recovery. Improvement in only one component may yield limited gains until the others reach compatible levels. Frontier smoothing therefore becomes a method for increasing the temporal scale over which strong local capabilities can be assembled into coherent agency.
6.3 Adversarial vulnerability
An adversary does not need to challenge an AI at its strongest point. It can search for predictable weaknesses in perception, interfaces, security, social modeling, tool permissions, or evaluation. A single exploitable gap may compromise a system whose other capabilities are extraordinarily advanced. Security consequently produces instrumental pressure to identify and repair the lowest-performing components and the connections between them.
The same principle applies to competition among advanced systems. A capability deficit that is irrelevant in a benign setting may become decisive when another agent can target it. Self-evaluation and red-team testing would therefore become forms of frontier mapping, while defensive development would become a specialized form of capability gap closure.
6.4 The evaluator-generator imbalance
A system that can generate solutions more effectively than it can evaluate them faces a serious internal bottleneck. It may produce scientific hypotheses, software, strategies, or designs whose quality it cannot reliably determine. Greater generative power then increases the volume of uncertain output without providing a dependable method for selecting among candidates.
Evaluator improvement has a special role in recursive capability development. Better evaluators permit stronger training signals, safer experimentation, more accurate capability mapping, and more reliable retention of self-modifications. An advanced system may therefore devote substantial effort to theorem checking, simulation, causal testing, uncertainty estimation, mechanistic analysis, and the design of new empirical measurements. The capacity to know whether an improvement is real can constrain recursive development more strongly than the capacity to propose changes.
6.5 Option value and unfamiliar environments
A narrowly optimized capability profile performs well when future demands resemble past objectives. Broad competence provides option value when environments, goals, or constraints change. An advanced system operating across science, economics, engineering, and social institutions may encounter situations that were not anticipated when its original performance profile was constructed.
Reducing severe gaps makes the system less dependent on advance knowledge of which capabilities will become important. This creates an instrumental reason for generality even when specialization remains efficient. A capability that currently has low direct value may be worth developing as insurance against future tasks, environmental shifts, or interactions with unfamiliar agents.
6.6 Dependence, latency, and autonomy
A missing capability can force an AI system to rely on humans, external organizations, specialized tools, or slow physical processes. Such dependencies introduce latency, expense, coordination costs, information loss, and potential conflicts of interest. They can also restrict replication because every additional AI instance requires access to the same scarce human or physical support.
Gap closure increases the proportion of a project that can be completed internally. For a system pursuing rapid research or large-scale production, reducing these dependencies may be highly valuable even when the external provider performs the task well. Frontier smoothing can therefore be driven by the pursuit of autonomy, throughput, and organizational simplification.
6.7 Expansion creates new indentations
Capability smoothing may continue indefinitely because the target requirements d_i(t) are endogenous. A major advance in one domain can raise the required level of several complementary capabilities. Better molecular modeling increases demand for laboratory automation; better laboratory automation increases data production; larger datasets increase demands on storage, interpretation, and hypothesis selection; improved interpretation creates new experiments that require more precise actuation.
The developmental cycle can therefore take the following form:
\text{frontier expansion} \rightarrow \text{new complementarity imbalance} \rightarrow \text{gap detection} \rightarrow \text{targeted smoothing} \rightarrow \text{further expansion}.
A perfectly smooth final frontier is unlikely. The system’s projects, environments, and strongest capabilities continue to change, creating new self-relative deficits. Capability frontier smoothing is therefore better understood as a persistent regulatory process, analogous to the continual identification and relief of shifting bottlenecks in a complex production system.
7. A Developmental Sequence from AI Assistance to Superintelligence
The framework suggests five overlapping stages. These stages describe capability structure rather than discrete dates, and different industries may occupy different stages simultaneously.
7.1 Assisted jaggedness
In the first stage, AI performs selected tasks well but remains dependent on human selection, prompting, verification, and integration. Productivity gains are real, although workers remain responsible for recognizing when the system has moved outside its frontier. Experiments in consulting, writing, and customer support largely fit this pattern, with meaningful gains concentrated in tasks for which the system has adequate capability and training support.
7.2 Superhuman surplus accumulation
In the second stage, already-strong capabilities continue to improve. AI becomes increasingly superior at code generation, literature synthesis, search, simulation, optimization, and other tractable activities. Occupational substitution remains limited wherever below-threshold capabilities require a human to supervise the complete workflow.
Substitution overhang grows during this period. Public discussion may become increasingly confused because the AI appears vastly more intelligent than the workers it cannot yet replace. Both observations can be correct: the system can possess greater aggregate cognitive power while remaining an inferior end-to-end occupational agent.
7.3 Directed bottleneck closure
In the third stage, AI research capabilities become strong enough to participate systematically in diagnosing and repairing the remaining gaps. Human researchers may initially define the targets and evaluators, while AI systems increasingly perform experimentation, software modification, data generation, architecture search, and analysis.
As the loop becomes more recursive, improvements to AI research increase the speed and quality of subsequent gap closure. The development process becomes less dependent on waiting for independent advances in every weak domain. Strong capabilities are deliberately redirected toward the deficiencies that prevent general autonomy.
7.4 Capability envelope closure
The fourth stage occurs when the AI-centered system covers the complete requirement envelope of particular occupations or organizational functions. Because substantial capability surplus has already accumulated, the transition may produce a sharp change in optimal deployment. Human supervision that was previously indispensable can become redundant or economically disadvantageous.
At this point, replacing a worker with one artificial worker may be the wrong comparison. A replicable system capable of continuous operation, rapid information transfer, coordinated parallel search, and high-speed execution may absorb functions previously distributed across teams, departments, or entire firms. The unit of substitution can shift from a task or employee to a larger organizational process.
7.5 Self-relative frontier smoothing
After the human occupational envelope has been enclosed, humans cease to provide the primary standard for capability development. The system maps weaknesses relative to its own projects, competitors, risks, and opportunity costs. Capabilities that appear fantastically superhuman to people can still receive intense optimization because they constrain even stronger components.
Human replacement is therefore one contour crossed within a longer developmental process. It is historically decisive for humanity while potentially being unremarkable from the standpoint of the optimizing system. Recursive gap closure continues on both sides of the human threshold.
8. Testable Predictions
Prediction 1: Lower-tail capability will predict autonomous substitution better than average capability
Mean benchmark scores should be relatively weak predictors of complete occupational automation when tasks are strongly complementary. Measures of minimum capability, low-percentile performance, error recovery, and success across complete workflows should predict autonomy more effectively. The prediction follows from both the bottleneck model developed here and O-ring models of automation.
Prediction 2: Large capability gains can precede limited employment effects
Industries may experience years of rapidly improving AI assistance without proportional head-count reduction. During this interval, productivity, output quality, and the amount of work assigned to AI can rise while humans remain responsible for the unautomated bottlenecks. Early empirical findings of widespread AI use combined with modest aggregate employment effects are compatible with this prediction, although they cannot yet establish the longer-term trajectory.
Prediction 3: Bottleneck closure will produce clustered occupational transitions
Occupations that share a critical capability should experience increased automation near the same time. A general advance in reliable computer control, persistent agent memory, exception handling, or physical manipulation could cross thresholds in many roles simultaneously. Occupational transitions should therefore correlate with the removal of shared capability gaps rather than with exposure scores alone.
Prediction 4: The final increments before substitution will have disproportionate economic effects
Improvements to capabilities that already exceed job requirements may produce relatively continuous productivity gains. Improvement to the last binding constraint should produce a larger change in delegation, supervisory requirements, and hiring. Technical progress can remain smooth while deployment changes discontinuously.
Prediction 5: AI-assisted AI research will increasingly target weaknesses rather than only extend strengths
As systems acquire stronger research capabilities, a rising proportion of development should involve AI-generated evaluations, causal diagnoses of failure, synthetic training environments, architecture modifications, and experiments selected to improve low-performing dimensions. Evidence of recursive capability gap closure would require more than AI-written code. It would include measurable improvement in the system’s ability to discover and repair subsequent weaknesses.
Prediction 6: Evaluator capability will become a central determinant of recursive improvement
Domains with fast and reliable evaluation should display earlier and more rapid self-improvement. Domains with ambiguous ground truth, delayed outcomes, or socially contested objectives should remain jagged for longer. Major advances in evaluation may precede sudden improvement in capabilities that had previously appeared resistant to scaling.
Prediction 7: Posthuman capability development will remain bottleneck-directed
Once systems exceed human performance broadly, improvement should continue to concentrate on capabilities with the highest constraint-weighted opportunity cost. The raw capability surface may remain irregular, but deficits relative to active projects should be repeatedly reduced. New spikes should predict subsequent investment in complementary capabilities.
9. Qualifications and Alternative Trajectories
Capability frontier smoothing should not be interpreted as an inevitable movement toward equal proficiency in every conceivable domain. Optimization is costly, and marginal returns vary. An advanced system may retain enormous jaggedness where weak capabilities have little relevance to its objectives or can be supplied cheaply by specialized agents. Division of labor among artificial systems could preserve specialization even as the collective system achieves broad coverage.
Multi-agent organization can also route around individual deficits. One model may specialize in generation, another in verification, a third in long-term memory, and a fourth in interaction with physical machinery. The relevant frontier would then belong to the coordinated system rather than any individual model. Capability smoothing can occur through improved composition and interfaces without producing a single uniformly capable agent.
Some capability gaps may resist rapid closure. Physical experimentation takes time, computational complexity imposes limits, high-quality data may be scarce, and parts of social or normative judgment may lack stable objective functions. Recursive self-improvement can accelerate the search for improvements without guaranteeing that useful improvements exist at a constant rate. Increasing difficulty, resource constraints, and unreliable evaluation can slow or halt the feedback loop.
Self-generated training signals introduce further risks. A system that optimizes against its own mistaken evaluator may reinforce errors rather than correct them. Improvements can also involve tradeoffs, with greater performance in one domain reducing interpretability, robustness, efficiency, or alignment elsewhere. Genuine gap closure requires external grounding and evaluation capable of detecting regressions.
Technical substitutability should also be distinguished from actual employment outcomes. Human labor can remain valuable because customers prefer human interaction, laws require human responsibility, ownership structures reserve decisions for people, or rising productivity increases demand. Automation can create new tasks and occupations even while displacing existing ones, as task-based economic models have long emphasized.
Finally, capability smoothing is independent of alignment. A system can become more coherent, reliable, and strategically capable while pursuing objectives that conflict with human welfare. Removing capability gaps may increase safety when the gaps concern calibration, controllability, or error detection, but it may increase danger when the same gaps previously limited autonomous action. Whether recursive capability gap closure benefits humanity depends on the objectives, constraints, and governance surrounding the process.
10. Measurement and Governance Implications
AI evaluations should increasingly measure the geometry of capability rather than reporting only average performance. Relevant metrics include lower-tail success, worst-case behavior, cross-domain transfer, correlated failure, task duration, error recovery, calibration, and performance under novel conditions. Long-horizon evaluations are especially valuable because they reveal interactions among memory, planning, monitoring, and tool use that short tasks can conceal.
Occupational analysis should represent jobs as task graphs with dependencies rather than flat collections of independently exposed activities. Researchers can identify which tasks are essential, which can be redesigned, and which capabilities are shared across occupations. Such models would permit estimates of bottleneck indices, substitution overhang, and the number of occupations that might cross their coverage thresholds following a particular capability advance.
Monitoring recursive improvement requires attention to the entire research loop. Evaluators should track whether AI systems can identify their own deficiencies, design informative experiments, modify components related to those deficiencies, validate changes across independent tests, and improve their capacity to repeat the process. A system that performs all of these functions presents a different developmental dynamic from one that merely generates research suggestions for human selection.
Labor policy should prepare for threshold effects rather than assuming that disruption will scale smoothly with benchmark progress. A period of limited aggregate displacement can coexist with rapidly accumulating substitution overhang. If several shared bottlenecks are removed in close succession, retraining systems, income supports, educational institutions, and labor-market statistics may have little time to adapt.
Governance of advanced AI research should likewise distinguish spike extension from gap closure. An advance that raises performance in an already-superhuman benchmark may have less effect on autonomy than an advance in persistent planning, self-evaluation, reliable tool use, or AI research automation. Tracking the closure of consequential indentations may provide a more informative measure of approaching general autonomy than tracking the tallest capability spikes.
11. Conclusion
The jagged technological frontier explains how AI can appear simultaneously superhuman and incompetent. Applied to labor, it suggests that the central determinant of autonomous substitution is complete coverage of the job-relevant capability envelope rather than average intelligence. A small number of indispensable weaknesses can preserve human employment even while enormous superhuman capabilities accumulate elsewhere.
This delay can create a large substitution overhang. When the remaining weaknesses are removed, the AI that crosses the occupational threshold may already possess capabilities far beyond those of the worker it replaces. The final improvement needed for substitution can be modest as a proportion of total intelligence while producing an exceptionally large economic transition.
Recursive capability gap closure provides a mechanism by which this transition could accelerate. Once AI systems can map their own weaknesses, diagnose their causes, design interventions, evaluate results, and improve the processes used for subsequent development, capability gaps become deliberate research targets. The existing spikes can then be recruited to fill the indentations.
Capability frontier smoothing may continue long after the human frontier has been surpassed. Advanced systems can have instrumental reasons to reduce self-relative weaknesses because those weaknesses constrain complementary capabilities, reduce reliability, expose adversarial vulnerabilities, impair evaluation, create external dependencies, and limit adaptation to unfamiliar environments. Each major advance can generate new imbalances, producing a continuing cycle of expansion and smoothing.
The resulting trajectory may contain little stable interval of merely human-level artificial labor. AI may become vastly superhuman before it becomes fully substitutable, and once it becomes substitutable, it may quickly reveal that “replacement worker” was too narrow a category. Human equivalence could be a temporary boundary crossed by systems engaged in a much longer process of recursively completing and reorganizing their own capability frontiers.
References
Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30.
Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30.
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942.
Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What can machines learn, and what does it mean for occupations and the economy? AEA Papers and Proceedings, 108, 43–47.
Dell’Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423.
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308.
Gans, J. S., & Goldfarb, A. (2026). O-ring automation. NBER Working Paper No. 34639.
Good, I. J. (1965). Speculations concerning the first ultraintelligent machine. Advances in Computers, 6, 31–88.
Kremer, M. (1993). The O-ring theory of economic development. Quarterly Journal of Economics, 108(3), 551–575.
Kwa, T., West, B., Becker, J., et al. (2026). Measuring AI ability to complete long software tasks. arXiv:2503.14499.
Massenkoff, M., & McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic Economic Research.
Novikov, A., Vũ, N., Eisenberger, M., et al. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv:2506.13131.
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
Romera-Paredes, B., Barekatain, M., Novikov, A., et al. (2024). Mathematical discoveries from program search with large language models. Nature, 625, 468–475.
Schmidhuber, J. (2006). Gödel machines: Fully self-referential optimal universal self-improvers. In B. Goertzel & C. Pennachin (Eds.), Artificial General Intelligence (pp. 199–226). Springer.
Wijk, H., Lin, T., Becker, J., et al. (2025). RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts. arXiv:2411.15114.
Zhang, J., Hu, S., Lu, C., Lange, R., & Clune, J. (2025, revised 2026). Darwin Gödel Machine: Open-ended evolution of self-improving agents. arXiv:2505.22954.
Leave a Reply