Source-linked AI summary
Scenarios for the Transition to AGI
Anton Korinek, Donghyun Suh
TL;DR
The paper examines how a transition toward AGI may affect output and wages as increasingly complex tasks become automatable. It models work as tasks with varying computational complexity and studies alternative task distributions and automation paths. Wages can rise indefinitely with a sufficiently thick unbounded task tail, but collapse under bounded complexity or when automation outpaces capital accumulation.
Problem
The paper asks how technological progress culminating in AGI may affect output and wages, including whether wages rise or collapse.
Source: §1 · p. 2Method
The paper models work as atomistic tasks of differing computational complexity and studies how an advancing automation index affects their automation.
Source: §1 · p. 2, §5 · p. 37Results
With an unbounded Pareto task tail, wages can rise indefinitely if the tail is sufficiently thick, whereas bounded task complexity leads wages to surge initially and then collapse.
Source: §5 · p. 37Takeaways & Limitations
The wage outlook depends on task-complexity assumptions and capital accumulation, while fixed factors can depress wages and automating innovation can restart wage growth.
Source: §5 · p. 37The baseline describes a static economy with a representative household endowed with labor and capital, while societal automation choices are separately represented by Ψ(I).
Source: §2.2 · p. 9, §4.3 · p. 31
Abstract
from arXiv · showhide
We analyze how output and wages behave under different scenarios for technological progress that may culminate in Artificial General Intelligence (AGI), defined as the ability of AI systems to perform all tasks that humans can perform. We assume that human work can be decomposed into atomistic tasks that differ in their complexity. Advances in technology make ever more complex tasks amenable to automation. The effects on wages depend on a race between automation and capital accumulation. If the distribution of task complexity exhibits a sufficiently thick infinite tail, then there is always enough work for humans, and wages may rise forever. By contrast, if the complexity of tasks that humans can perform is bounded and full automation is reached, then wages collapse. But declines may occur even before if large-scale automation outpaces capital accumulation and makes labor too abundant. Automating productivity growth may lead to broad-based gains in the returns to all factors. By contrast, bottlenecks to growth from irreproducible scarce factors may exacerbate the decline in wages.
1 Introduction
The paper develops a compute-centric framework for analyzing how automation toward AGI affects output and wages, asking whether wages rise or collapse as tasks become automatable. Its results depend on task complexity, labor scarcity, and the pace of capital accumulation.
1 Introduction
- The model treats work as atomistic tasks of differing complexity and assumes technological progress raises the maximum task complexity that can be automated.Task complexity is interpreted in compute space; labor performs all tasks, while capital can perform automated tasks.Source: §1 · pp. 2–3
- With fixed capital and labor, wages remain high until automation passes a threshold that alleviates labor scarcity, after which wages fall to capital’s marginal product despite remaining human tasks.In this second region, labor and capital are perfect substitutes at the margin.Source: §1 · p. 3
- For given factor endowments, wages respond to automation in a hump-shaped pattern, while the factor-price frontier expands during incomplete automation and collapses when all tasks are automated.The frontier summarizes possible factor-price combinations across capital/labor ratios.Source: §1 · p. 3
- In dynamic settings, wage outcomes reflect a race between automation and capital accumulation, with results bounded by constant capital and instantaneous adjustment to optimal capital.For optimizing representative agents with linearly separable intertemporal preferences, output and wage effects lie between those bounds.Source: §1 · p. 4
- When task complexity is bounded, full automation arrives in finite time and wages collapse; with an unbounded distribution, wages rise forever only if the remaining-task tail is sufficiently thick.If the unbounded tail is too thin, wages eventually collapse.Source: §1 · p. 4
- In the business-as-usual simulation, unbounded task complexity prevents finite-time AGI, and output and wages rise forever at a pace similar to advanced-country experience over the past century.The scenario combines a Pareto task-complexity distribution with exponential automation-index growth.Source: §1 · p. 4
- The baseline and aggressive AGI simulations reach full automation within 20 and five years, respectively, and show ten times faster growth than business as usual, while wages collapse near full automation.The aggressive scenario assumes a shorter-tailed task distribution.Source: §1 · p. 4
- Extensions show that fixed-factor bottlenecks can worsen wage declines, automating innovation can lift returns to all factors, and retaining enough human-only jobs or slowing automation can support wages.Slowing automation may bring significant worker gains at the cost of forgoing a growing fraction of output; heterogeneous skills can leave a shrinking superstar share with rising wages while most workers are devalued.Source: §1 · p. 5
2 A Compute-Centric Model of Automation
The model treats work as atomistic tasks with different machine-compute requirements and asks how automation changes their allocation across capital and labor. It derives a threshold for labor scarcity and characterizes how automation and technology shift factor returns.
2.1 Tasks in Compute Space
- Human work is modeled as recombinable atomistic tasks whose automation difficulty is measured by machine compute intensity, while an exogenous exponential index expands the automatable complexity frontier.Machine compute intensity may differ sharply from how difficult a task is for humans.Source: §2.1 · pp. 6–8
2.2 Baseline Model
- A continuum of complexity-ranked tasks is aggregated with CES production; tasks below threshold I can use capital or labor, while higher-complexity tasks require labor.The baseline assumes task inputs are gross complements, with elasticity of substitution σ < 1.Source: §2.2 · p. 9
2.3 Equilibrium: Characterizing Two Regions
- For fixed factor supplies, automation crosses threshold Î where labor stops being scarce: below it w > R and labor serves only unautomated tasks; above it w = R and may join automated tasks.The threshold rises with K/L; once labor is no longer scarce, it is interchangeable with capital at the margin even when tasks are highly complementary.Source: §2.3 · pp. 10, 12
2.4 Factor Price Frontier (FPF)
- The factor-price frontier maps feasible returns: greater automation raises limiting wages and rotates it, while higher A proportionately raises returns to both factors.In the labor-scarce region, greater capital intensity raises wages and lowers the return to capital; when labor scarcity is relieved, the frontier is the point w = R = A.Source: §2.4 · pp. 12–13
The Automation Path on the Factor Price Frontier
With capital and labor fixed, automation raises capital returns and can raise output, while its effect on wages eventually reverses as labor displacement dominates. The factor-price response depends on capital abundance and the extent of automation.
The Automation Path on the Factor Price Frontier
- Output rises with automation only below the threshold; the productivity gain requires capital to be sufficiently abundant relative to labor.Once the threshold is reached, output is unaffected by further automation.Source: The Automation Path on the Factor… · p. 14
- Wages rise with automation below a capital-intensity-dependent threshold and fall above it, while capital returns rise whenever automation is incomplete.At sufficiently high automation, both factor returns equal A.Source: The Automation Path on the Factor… · p. 14
- Figure 5 shows a clockwise-rotating factor-price frontier: along the automation path, capital returns rise continuously while wages first rise and then fall.At the threshold, both returns equal A.Source: Figure 5 · p. 15, The Automation Path on the Factor… · p. 15
2.5 Automation and Factor Earnings
- With equal capital and labor endowments, the wage bill eventually declines as automation rises, while output continues rising and capital returns grow faster.At Φ = 0.5, factor returns are equal and factor shares are one-half each.Source: §2.5 · p. 16, Figure 6 · p. 16
- With capital ten times labor and strong complementarity, wages fall about 85% from around Φ ≈ 0.83, after growing until roughly Φ ≈ 0.80.At Φ = 10/11, factor returns equalize; the capital share is ten times the labor share.Source: §2.5 · p. 16, Figure 6 · p. 16
- The output gain from full automation depends on available capital: output doubles in the equal-endowment illustration and grows eleven-fold when capital is ten times labor.This motivates analyzing automation together with capital accumulation.Source: §2.5 · p. 16
3 Dynamics: The Race between Automation and Capital Accumulation
The section studies how automation interacts with capital accumulation, using alternative task-complexity distributions and a dynamic household model. Its results show that wages depend on the pace of automation relative to capital accumulation, and that bounded-task scenarios can produce wage collapse before full automation.
3.1 Automation Scenarios
- The scenarios contrast an unbounded Pareto task distribution, which always leaves tasks unautomated, with bounded distributions that reach full automation in 20 or 5 years.The bout-of-automation case combines rapid task automation with an unbounded tail.Source: §3.1 · p. 17
3.2 Consumer Problem
- The model bounds output and wages between fixed-capital lower paths and upper paths, with the lower wage bound first rising and then declining as automation advances.The upper bounds rise with automation before full automation, when the upper wage bound collapses to A.Source: §3.2 · p. 19
- With bounded task complexity, the wage upper bound collapses discontinuously to A at full automation, independently of preferences and savings behavior.For bounded distributions, full automation is reached in finite time; the lower wage bound also declines to A.Source: §3.2 · pp. 19–20
- The savings rate needed to offset automation’s wage effect rises with net displacement, the capital-output ratio, the automated-task mass, and automation growth.The resulting threshold separates wage growth from wage decline in the short run.Source: §3.2 · pp. 20–21
- In the Pareto scenario, wage growth ranges from collapse to continued growth depending on automation’s rate relative to capital accumulation.If automation is too fast, wages converge to A; at intermediate rates capital accumulation constrains wage growth, while sufficiently slow automation makes automation the constraint.Source: §3.2 · pp. 22–23
3.3 Numerical Illustration
- The business-as-usual simulation yields about 2% annual output growth, with capital returns and the wage bill rising roughly in tandem.Its automation rate is λg = 0.01 per year, and the labor share declines slightly.Source: Figure 8 · p. 24, §3.3 · p. 25
- In the baseline AGI simulation wages collapse before full automation, while the aggressive scenario’s collapse occurs after about 3 years and its growth take-off occurs earlier.The baseline scenario reaches full automation in 20 years and has steady-state growth of 18% per year after collapse; the aggressive scenario reaches it in 5 years.Source: §3.1 · p. 17, §3.3 · p. 25
- In the bout-of-automation simulation, wages collapse initially but rise above A and grow again around year 9 as capital accumulation restores labor scarcity.The scenario retains a long tail of tasks that cannot be automated.Source: §3.1 · p. 17, §3.3 · p. 25
4 Extensions
The extensions examine how automation affects growth, factor scarcity and returns, and the distribution of gains and losses. They consider fixed-factor scarcity, automating R&D, social limits and worker-skill differences, as well as task-specific capital and the pace of compute accumulation.
4.1 Fixed Factors and the Return of Scarcity
- With fixed factors, automation eventually outpaces capital accumulation, and wages converge to capital returns when the economy enters the labor-abundant region in finite time.For a given capital intensity, automation initially raises returns to capital, while its effect on wages is hump-shaped; the fixed factor lowers the automation level at which wages begin to decline.Source: §4.1 · p. 27
4.2 Automating Technological Progress
- Once automation passes a threshold, automating ideas production can drive technology and output to infinity in finite time, with labor’s marginal product rising alongside output.The result assumes a constant savings rate, fixed capital allocation across output and ideas, and no depreciation; sufficient automation of final-output production lets the other automation and knowledge-return parameters be any finite values. The authors note that fixed factors may eventually bottleneck both output and ideas production.Source: §4.2 · pp. 29, 31
4.3 Nostalgic Jobs or Limits on Automation
- A wage-maximizing automation policy yields exponentially growing wages instead of collapse, but eventually forgoes almost 100% of output potential.The planner follows the feasible pace of automation through period 14, then slows it; for Pareto automation paths, the proposition characterizes the constant rate that maximizes long-run wage growth.Source: §4.3 · p. 32, Figure 11 · p. 33
4.4 Heterogeneous Worker Skills
- With heterogeneous skills, automation can make lower-skill workers substitutable first; if skill levels approach full coverage only asymptotically, a shrinking top group may gain without bound while most wages fall to capital returns.The model treats a worker’s skill as the maximum task complexity they can perform; under the stated assumption, non-substitutable workers are effective substitutes for one another.Source: §4.4 · pp. 33–34
4.5 Compute as Specific Capital
- Compute scarcity makes task-specific, hard-to-reallocate capital important, unlike the baseline model’s uniform capital that can be smoothly deployed across newly automatable tasks.The paper identifies compute as a particularly salient, limited-supply form of specific capital that slows AI deployment.Source: §4.5 · p. 34
- Capital specific to a task can be invested only after that task is automated, so sudden automation can outpace accumulation of the required capital.During slow, steady automation, sufficient capital is accumulated as new tasks become automatable, leaving redeployment constraints slack.Source: §4.5 · p. 35
- When compute for newly automatable tasks has not yet accumulated, humans still perform those tasks at a wage despite their technical automability.At this initial point, labor and compute are marginal substitutes and their returns are equated.Source: §4.5 · p. 36
- As compute accumulates, its return declines between k1 and k2; beyond k2, investment spreads across capital types and returns on capital decline.Once compute replaces humans on the newly automated tasks, wages and traditional-capital returns rise while compute’s return falls sharply until it matches other capital’s marginal product.Source: §4.5 · p. 36
- Below k1, wages fall and traditional-capital returns rise; between k1 and k2, both rise; at or above k2, wages rise with capital accumulation.Proposition 13 also states that beyond k2, specific capital is accumulated alongside traditional capital.Source: §4.5 · p. 36
- Rapid automation can give specific capital such as compute temporarily high returns, but reproducible accumulation ultimately dissipates the excess.After adjustment, specific capital for newly automated processes earns the market rate of return like other capital.Source: §4.5 · p. 36
5 Conclusions
The paper uses a compute-centric task framework to analyze how automation and capital accumulation shape output and wages during a possible transition to AGI. Outcomes depend on task complexity and growth assumptions: wages may rise indefinitely, collapse before full AGI, or resume growth when innovation is automated.
5 Conclusions
- A compute-centric framework represents work as tasks of varying computational complexity and studies how computing growth affects automation and AGI.The paper also identifies labor and capital adjustment costs, endogenous innovation, distributional impacts, macroeconomic dynamics and policies, and an intelligence explosion as directions for further work.Source: §5 · p. 37
- With a sufficiently thick infinite Pareto tail of task complexity, capital accumulation can automate ever more complex tasks while leaving enough work for wages to rise indefinitely.If the tail is too thin, automation outpaces capital accumulation and wages collapse.Source: §5 · p. 37
- When human-performable task complexity is bounded, wages can surge initially but eventually collapse even before full AGI is reached.Source: §5 · p. 37
- Automation has an inverse-U effect on wages; sufficient capital accumulation can prevent wage declines, while fixed factors can eventually drive wages down.Automating innovation can restart wage growth after an initial automation-driven collapse.Source: §5 · p. 37
- The authors present scenarios for examining alternative futures and preparing for technological transformations associated with AGI.Source: §5 · p. 37
A.1 Proof of Proposition 7
The proof compares capital accumulation with automation to characterize wage growth under a Pareto task-complexity assumption. Wages rise when capital accumulation and automation's productivity effect outweigh labor displacement, but can decline and collapse when automation advances too quickly.
A.1 Proof of Proposition 7
- Capital accumulation must be sufficiently fast relative to automation to keep the economy in region 1; otherwise, the automation index eventually crosses the threshold and wages collapse to A.Source: §A.1 · p. 42
- Wages rise while capital accumulation and automation's productivity effect dominate the displacement effect.The proof identifies this balance as determining wage growth.Source: §A.1 · p. 43
- If s∞A − δ > λg, wages grow exponentially; when s∞A − δ ≤ λg, wages decline until automation crosses the threshold and collapse to A.Source: §A.1 · p. 44
A.2 Proof of Proposition 13
The proof considers alternative production-function cases depending on which inputs are perfect substitutes. It distinguishes labor from traditional capital as the input substitutable with specific capital.
A.2 Proof of Proposition 13
- When FL < FK(It), specific capital and labor are perfectly substitutable.Source: §A.2 · p. 45
- The proof rewrites the production function for each of these two substitution cases.Source: §A.2 · p. 45
- When FK > FK(It), specific capital and traditional capital are perfectly substitutable.Source: §A.2 · p. 45