Modeling The AGI Economy

Competition, Redistribution and the Fork Ahead

Before building anything, we -- and by we I mean Claude -- surveyed the existing academic work. There is a lot of fantastic work to build on, but apparently nobody has combined all the elements needed to address both competition and redistribution at the same time (if someone has done that already I would love to see it):

The task-based framework comes from Acemoglu and Restrepo. Their 2022 Econometrica paper models production as a continuum of tasks that can be allocated between capital and different types of labor, with automation expanding the set of tasks capital can perform. This gives you displacement effects, productivity effects, and ripple effects through the wage structure. But it assumes competitive markets and has no redistribution policy.

For wealth dynamics, Moll, Rachel, and Restrepo's "Uneven Growth" shows how automation raises returns to wealth, generating endogenous wealth concentration through a heterogeneous-agent model. The April 2025 IMF working paper by Broadberry et al. builds on this to study AI diffusion. But again, no market power lever and no redistribution.

Korinek and Stiglitz (2019) provide the conceptual architecture closest to what we need. They explicitly discuss both anti-trust policy (lowering monopoly rents so competition passes cost savings to consumers) and redistribution (non-distortionary taxation to compensate losers). But their models are deliberately simple illustrations rather than a full GE model.

Saint-Paul (2025) comes closest to combining all the pieces. He models an economy where oligarchs who own proprietary technology choose between UBI and blocking AI to preserve the middle class as consumers. His key finding, that UBI plus AI can dominate blocking AI, resonates with my intuition. But his model is more political economy than full GE.

On the growth side, Aghion, Jones, and Jones (2017) and Trammell and Korinek (2024) provide frameworks for thinking about explosive growth and the Baumol cost disease constraint. Growth may be constrained not by what AI is good at, but by what remains essential yet hard to automate. Trammell and Korinek is especially useful as a comprehensive survey mapping the parameter space.

On the redistribution side specifically, Lopes (2024) builds a large-scale overlapping generations GE model to study UBI, finding that an expenditure-neutral reform can increase capital accumulation and reduce inequality but doesn't model AI automation.

And on the neglected question of market structure, Barkan (2024) demonstrates that under imperfect competition, AI productivity increases can actually decrease GDP, the opposite of what competitive models predict.

Based on this work Claude constructed a GE model that simultaneously includes heterogeneous capital ownership generating endogenous inequality, market concentration as a policy variable, and a negative income tax or basic income.

The model combines four building blocks:

1. Production uses a CES task framework (following Acemoglu-Restrepo). A fraction α of tasks are automated and performed by AI-capital with productivity A; the rest require labor. The elasticity of substitution σ is critical. When σ > 1, capital and labor are good substitutes, enabling the "everything gets cheap" scenario; when σ < 1, labor remains essential as a Baumol bottleneck.

2. Market power is modeled as Cournot competition with N symmetric firms. This gives a markup μ that ranges from near-monopoly to near-perfect competition. The markup compresses the effective labor share and creates a wedge between productivity and consumer prices. Competition policy — antitrust, regulation, open standards — is captured by the slider for N.

3. Capital dynamics are endogenous. Ten household deciles hold unequal capital stocks, with concentration governed by a parameter θ. Each period, households save a fraction of their post-tax income, with richer households saving more. Capital accumulates: k_{i,t+1} = (1−δ)k_i + s_i·y_net. This is the mechanism through which inequality compounds over time. The Piketty r > g dynamic emerges endogenously.

4. Negative income tax implements budget-balanced redistribution: y_net = (1−t)·y + t·ȳ, where t is the tax rate and ȳ is mean income. Below-mean earners receive transfers; above-mean earners pay. A labor supply elasticity captures disincentive effects.

The simulation runs for 40 periods, with automation ramping logistically from 30% toward the target level. AI productivity compounds at a user-selected growth rate. The model tracks output, prices, the Gini coefficient, real purchasing power by decile, capital accumulation, and the labor share over time.

The five presets tell the story concisely:

AI Dystopia (high automation, duopoly, no redistribution): The labor share collapses, the Gini climbs relentlessly over time, and the bottom decile's real purchasing power actually falls even as total output soars. Capital ownership concentrates as the rich save more and accumulate more capital each period. This is the permanent precariat.

Competition Only (high automation, 30 firms, no redistribution): Prices fall dramatically as competition forces cost savings through to consumers. The bottom decile does better than under dystopia because goods are cheaper. But inequality still compounds through differential capital accumulation. Competition helps with the price level but doesn't fix the income distribution.

Redistribution Only (high automation, duopoly, 40% NIT): The Gini compresses, but prices stay elevated because the monopoly markup prevents productivity gains from reaching consumers. The transfers are effectively subsidizing monopoly rents.

AI Utopia (high automation, competitive markets, moderate NIT): Output grows substantially, prices fall, the Gini stays low, and the bottom decile's real purchasing power rises alongside everyone else's. The combination of competition (driving prices down) and redistribution (maintaining purchasing power) produces broadly shared prosperity.

The dynamic charts make something visceral that static analysis obscures: inequality compounds. Even starting from moderate initial conditions, the differential savings mechanism means that capital ownership concentrates over time unless policy actively counteracts it.

This is a first-cut model meant to build intuition and invite discussion, not a calibrated forecasting tool. Also as stated above, it is entirely vibe coded, so YMMV.

Among the things worth considering:

Most importantly: checking whether the model appears to be correct. I would love to see other agents pointed at this for analysis.

The NIT currently taxes all income uniformly. This is obviously far from the current policy which taxes labor income more heavily than capital income. So effectively an additional policy choice is at present hardcoded.

The model treats the number of firms as static to simulate a policy choice on competition. There are more dynamic ways of modeling concentration and more sophisticated approaches to examining different types of regulation.

Capital ownership broadening (sovereign wealth funds, stakeholder ownership, broad-based equity participation) is arguably a third policy dimension beyond competition and redistribution. The model doesn't currently include it.

On the production side, the model doesn't capture the Baumol bottleneck dynamics that Aghion-Jones-Jones emphasize: sectors where automation is hard may constrain overall growth regardless of how productive AI becomes elsewhere.

Finally, the model doesn't include the emergence of a "human qua human" sector of the economy: jobs that could be automated but where consumer preferences choose the "artisanal" option instead (e.g. in hospitality).

I'm sharing this model because I think the debate about the AGI economy is too often conducted through competing anecdotes and vibes rather than through structured analysis. Please try the model yourself, find bugs in it, criticize the assumptions, propose extensions. What other policy levers should be in there? What dynamics are missing? You can fork the repo or make a pull request with suggestions.

My intuition that you need both competition and redistribution via a basic income or negative income tax, and that neither alone is sufficient seems to be supported by this model. Looking forward to feedback!

Before building anything, we -- and by we I mean Claude -- surveyed the existing academic work. There is a lot of fantastic work to build on, but apparently nobody has combined all the elements needed to address both competition and redistribution at the same time (if someone has done that already I would love to see it):

The task-based framework comes from Acemoglu and Restrepo. Their 2022 Econometrica paper models production as a continuum of tasks that can be allocated between capital and different types of labor, with automation expanding the set of tasks capital can perform. This gives you displacement effects, productivity effects, and ripple effects through the wage structure. But it assumes competitive markets and has no redistribution policy.

For wealth dynamics, Moll, Rachel, and Restrepo's "Uneven Growth" shows how automation raises returns to wealth, generating endogenous wealth concentration through a heterogeneous-agent model. The April 2025 IMF working paper by Broadberry et al. builds on this to study AI diffusion. But again, no market power lever and no redistribution.

Korinek and Stiglitz (2019) provide the conceptual architecture closest to what we need. They explicitly discuss both anti-trust policy (lowering monopoly rents so competition passes cost savings to consumers) and redistribution (non-distortionary taxation to compensate losers). But their models are deliberately simple illustrations rather than a full GE model.

Saint-Paul (2025) comes closest to combining all the pieces. He models an economy where oligarchs who own proprietary technology choose between UBI and blocking AI to preserve the middle class as consumers. His key finding, that UBI plus AI can dominate blocking AI, resonates with my intuition. But his model is more political economy than full GE.

On the growth side, Aghion, Jones, and Jones (2017) and Trammell and Korinek (2024) provide frameworks for thinking about explosive growth and the Baumol cost disease constraint. Growth may be constrained not by what AI is good at, but by what remains essential yet hard to automate. Trammell and Korinek is especially useful as a comprehensive survey mapping the parameter space.

On the redistribution side specifically, Lopes (2024) builds a large-scale overlapping generations GE model to study UBI, finding that an expenditure-neutral reform can increase capital accumulation and reduce inequality but doesn't model AI automation.

And on the neglected question of market structure, Barkan (2024) demonstrates that under imperfect competition, AI productivity increases can actually decrease GDP, the opposite of what competitive models predict.

Based on this work Claude constructed a GE model that simultaneously includes heterogeneous capital ownership generating endogenous inequality, market concentration as a policy variable, and a negative income tax or basic income.

The model combines four building blocks:

1. Production uses a CES task framework (following Acemoglu-Restrepo). A fraction α of tasks are automated and performed by AI-capital with productivity A; the rest require labor. The elasticity of substitution σ is critical. When σ > 1, capital and labor are good substitutes, enabling the "everything gets cheap" scenario; when σ < 1, labor remains essential as a Baumol bottleneck.

2. Market power is modeled as Cournot competition with N symmetric firms. This gives a markup μ that ranges from near-monopoly to near-perfect competition. The markup compresses the effective labor share and creates a wedge between productivity and consumer prices. Competition policy — antitrust, regulation, open standards — is captured by the slider for N.

3. Capital dynamics are endogenous. Ten household deciles hold unequal capital stocks, with concentration governed by a parameter θ. Each period, households save a fraction of their post-tax income, with richer households saving more. Capital accumulates: k_{i,t+1} = (1−δ)k_i + s_i·y_net. This is the mechanism through which inequality compounds over time. The Piketty r > g dynamic emerges endogenously.

4. Negative income tax implements budget-balanced redistribution: y_net = (1−t)·y + t·ȳ, where t is the tax rate and ȳ is mean income. Below-mean earners receive transfers; above-mean earners pay. A labor supply elasticity captures disincentive effects.

The simulation runs for 40 periods, with automation ramping logistically from 30% toward the target level. AI productivity compounds at a user-selected growth rate. The model tracks output, prices, the Gini coefficient, real purchasing power by decile, capital accumulation, and the labor share over time.

The five presets tell the story concisely:

AI Dystopia (high automation, duopoly, no redistribution): The labor share collapses, the Gini climbs relentlessly over time, and the bottom decile's real purchasing power actually falls even as total output soars. Capital ownership concentrates as the rich save more and accumulate more capital each period. This is the permanent precariat.

Competition Only (high automation, 30 firms, no redistribution): Prices fall dramatically as competition forces cost savings through to consumers. The bottom decile does better than under dystopia because goods are cheaper. But inequality still compounds through differential capital accumulation. Competition helps with the price level but doesn't fix the income distribution.

Redistribution Only (high automation, duopoly, 40% NIT): The Gini compresses, but prices stay elevated because the monopoly markup prevents productivity gains from reaching consumers. The transfers are effectively subsidizing monopoly rents.

AI Utopia (high automation, competitive markets, moderate NIT): Output grows substantially, prices fall, the Gini stays low, and the bottom decile's real purchasing power rises alongside everyone else's. The combination of competition (driving prices down) and redistribution (maintaining purchasing power) produces broadly shared prosperity.

The dynamic charts make something visceral that static analysis obscures: inequality compounds. Even starting from moderate initial conditions, the differential savings mechanism means that capital ownership concentrates over time unless policy actively counteracts it.

This is a first-cut model meant to build intuition and invite discussion, not a calibrated forecasting tool. Also as stated above, it is entirely vibe coded, so YMMV.

Among the things worth considering:

Most importantly: checking whether the model appears to be correct. I would love to see other agents pointed at this for analysis.

The NIT currently taxes all income uniformly. This is obviously far from the current policy which taxes labor income more heavily than capital income. So effectively an additional policy choice is at present hardcoded.

The model treats the number of firms as static to simulate a policy choice on competition. There are more dynamic ways of modeling concentration and more sophisticated approaches to examining different types of regulation.

Capital ownership broadening (sovereign wealth funds, stakeholder ownership, broad-based equity participation) is arguably a third policy dimension beyond competition and redistribution. The model doesn't currently include it.

On the production side, the model doesn't capture the Baumol bottleneck dynamics that Aghion-Jones-Jones emphasize: sectors where automation is hard may constrain overall growth regardless of how productive AI becomes elsewhere.

Finally, the model doesn't include the emergence of a "human qua human" sector of the economy: jobs that could be automated but where consumer preferences choose the "artisanal" option instead (e.g. in hospitality).

I'm sharing this model because I think the debate about the AGI economy is too often conducted through competing anecdotes and vibes rather than through structured analysis. Please try the model yourself, find bugs in it, criticize the assumptions, propose extensions. What other policy levers should be in there? What dynamics are missing? You can fork the repo or make a pull request with suggestions.

My intuition that you need both competition and redistribution via a basic income or negative income tax, and that neither alone is sufficient seems to be supported by this model. Looking forward to feedback!