Economics of TAI Sprint Submission: A Recursively-Inspired Framework and Simulation Results for Evaluating a Value-Based UBI Policy

Ram Potham

The potential for Transformative AI (TAI) to exacerbate inequality necessitates exploring novel redistributive policies. This document presents a comprehensive framework and reports results from a conceptual simulation study evaluating the economic implications of a Universal Basic Income (UBI) funded by a significant annual tax (e.g., 4%) on company value. Drawing insights from recursive alignment (RA) and functional models of intelligence (FMI), this framework emphasizes policy adaptability and potential failure modes arising from rigid design under TAI's deep uncertainty. Developed for the Economics of TAI Sprint (Tracks 3 & 5), this document details the core research questions, analytical focal points, conceptual model structure, RA-inspired simulation design, and interprets the simulation outputs. It serves as a proof-of-concept and guide for future research into designing resilient economic policies for the TAI era.

Reviewer's Comments

Reviewer's Comments

Arrow
Arrow
Arrow

Joel Christoph

The paper proposes a value based UBI funded by a four percent annual tax on company equity and evaluates it with an agent based model inspired by recursive alignment ideas. The concept of viewing policy design through adaptability, rigidity, and failure modes is creative and fits the sprint’s call for resilient governance tools. The author outlines a clear modular framework with heterogeneous households, strategic firms, and government rules, and implements five scenarios in Mesa, reporting time series for inequality, investment, and fiscal balance on page 6 and the triplet of panels in Figure 1 on page 8 ​

.

Nonetheless the contribution remains exploratory. Key parameters that drive the simulated paths, such as the elasticity of investment to after tax returns or the stochastic process for firm value manipulation, are only described qualitatively. No calibration to empirical data, robustness checks, or confidence bands accompany the headline Figure 1 results. The decision to average just five runs for each scenario limits statistical power, and the model horizon of thirty years may not capture longer term capital accumulation or demographic feedbacks. The literature engagement focuses on recursive alignment and functional models of intelligence but omits the extensive empirical and theoretical work on UBI finance, wealth taxes, and adaptive fiscal rules.

Links to AI safety are present but indirect. The paper claims that adaptive fiscal policy can mimic recursive alignment principles and avoid stability failures, yet it does not show how the tax design interacts with advanced AI incentives, ownership concentration of compute, or catastrophic misuse channels. Adding explicit pathways from tax funded redistribution to safety outcomes, for example reduced race dynamics or funding for alignment research, would raise the safety relevance.

Technical documentation is thin. The Mesa code, parameter file, and raw outputs are not shared, preventing replication. The framework is verbally specified but lacks formal equations or a flowchart. Tables summarising baseline values, shock sizes, and adaptive rule algorithms would help readers evaluate validity. A sensitivity sweep over tax rates and adaptive triggers, together with comparisons to simpler lump sum or payroll financed UBI baselines, would illustrate the marginal value of the proposed policy.

Suggested next steps

1. Release the model repository with a reproducible environment and add Monte Carlo sensitivity around growth, gaming probability, and tax elasticity.

2. Calibrate initial distributions and TAI productivity shocks to stylised facts from AI macro projections.

3. Expand the literature review to include recent papers on wealth taxation feasibility and dynamic fiscal rules.

4. Map the policy’s impact on safety relevant metrics such as concentration of AI profits or funding for public goods that reduce existential risk.

5. Provide formal notation for production, household, and government behaviour to allow integration with general equilibrium or optimal control extensions.

Cecilia Wood

Interesting approach with clear policy implication. I thought Section 5 was particularly strong and introduced some interesting and potentially important messages. I would have liked to have understood the methodology in a lot more detail - it's difficult to assess the validity of your conclusions without understanding how you generated each of the graphs. (There should be enough detail to essentially reproduce them)

Some more specific remarks below.

Define what you mean by tax margin – do you mean the marginal rate of taxation?

When you talk about incentives, I would talk about efficiency losses at the equilibrium (which implies

strategic behaviour). This project felt very ambitious at first - it seemed like you were introducing a lot of non-standard assumptions unless you were basing this on an existing model (this isn't my area of expertise). My guess is including all of them at once would make it hard both to solve analytically and to draw conclusions about how each addition changes the dynamics, but a simulation based approach is more tractable.

Explain recursive alignment and FMI (you define in introduction but don’t explain). I didn't quite understand how this impacted your approach, and how your approach differs from standard macro models.

Fiscal neutrality is a nice goal but hard in practice – do you mean scenario 2 as a benchmark, ie what’s achievable in an ideal scenario? Worth discussing limitations (e.g. governments might need to accurately understand productivity changes, which are hard to estimate even if you did have full access to corporate data)

Key parameters set to plausible values – ideally, find another well-known paper which you can point to, or otherwise need to justify why these are plausible.

5 runs aren’t generally enough. Unclear where you introduce stochasticity in your modelling

Section 5 has some thoughtful discussion of the results - negligible policy impact on investment rate is very interesting.

Yulu Pi

This paper investigates a UBI policy funded through a substantial annual tax on company value, as a potential response to the growing inequality that TAI might exacerbate. While the concepts of RA and FMI are mentioned frequently, they are never clearly defined, making it hard for to fully understand the framework or assess the reasoning behind the simulations.

Your project rightly highlights the uncertainty around TAI development as a key factor in designing and evaluating UBI policy. This focus is well-placed and necessary. However, the simulation could do more to reflect that uncertainty. Expanding the scenarios to include not just different policy designs, but also how the same policy performs under varied TAI trajectories, would provide a more robust test of the model.

For readers who aren’t already familiar with the Mesa simulation library, the technical setup could use more clarity. It would be helpful to explain how key assumptions were made, how agent behavior was modeled, and what led to specific outcomes—such as the finding that the tax had little effect on investment.

This is a very timely contribution which tackles important issues TAI might rise. It asks important questions and presents a really creative modeling approach. Addressing issues I mentioned above can help the paper become more robust.

Rubi Hudson

This project could be improved by writing out the equations for the model and justifying why those modelling choices were made. The parameter values used for simulations should be specified.

Engaging with existing literature and citing it would help improve this work.

It's unclear how this model was inspired by recursive alignment and functional models of intelligence.

Strong research questions often involve a comparison along a single axis. What's the state of the art or default policy, and how does your work improve on it? Answering this question clearly would make for a stronger project.

Cite this work

@misc {

title={

Economics of TAI Sprint Submission: A Recursively-Inspired Framework and Simulation Results for Evaluating a Value-Based UBI Policy

},

author={

Ram Potham

},

date={

4/28/25

},

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

In this work, we investigate the emergence of capabilities in reinforcement learning (RL) by framing them as developmental phase transitions. We propose that the individual components of the reward function can serve as direct observables for these transitions, avoiding the need for complex, derived metrics. To test this, we trained a language model on an arithmetic task using Group Relative Policy Optimization (GRPO) and analyzed its learning trajectory with the Local Learning Coefficient (LLC) from Singular Learning Theory. Our findings show a strong qualitative correlation between spikes in the LLC—indicating a phase transition—and significant shifts in the model's behavior, as reflected by changes in specific reward components for correctness and conciseness. This demonstrates a more direct and scalable method for monitoring capability acquisition, offering a valuable proof-of-concept for developmental interpretability and AI safety. To facilitate reproducibility, we make our code available at \url{github.com/ilijalichkovski/apart-physics}.

Read More

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

We estimate parameters from real-world data (e.g., adversarial success rates, detection latency, patching delays) and simulate attack propagation scenarios across 8 sectors (enterprise, retail, trading, development, customer service, academia, medical, and critical infrastructure AI tools).

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Read More

Jul 28, 2025

Momentum–Point-Perplexity Mechanics in Large Language Models

This work analyzes the hidden states of twenty different open-source transformer language models, ranging from small to medium size and covering five major architectures. The key discovery is that these models show signs of "energy conservation" during inference—meaning a certain measure combining changes in hidden states and token unpredictability stays almost constant as the model processes text.

The authors developed a new framework inspired by physics to jointly analyze how hidden states and prediction confidence evolve over time. They propose that transformers' behavior can be understood as following certain mechanical principles, much like how physical systems follow rules like conservation of energy.

Their experiments show that this conserved quantity varies very little between tokens, especially in untrained (random-weight) models, where it's extremely stable. In pre-trained models, the average energy drops more due to training, but there are larger relative fluctuations from token to token.

They also introduce a new method, based on this framework, for controlling transformer outputs by "steering" the hidden states. This method achieves good results—producing completions rated as higher in semantic quality, while still maintaining the same kind of energy stability.

Overall, the findings suggest that viewing transformer models through the lens of physical mechanics gives new, principled ways to interpret and control their behavior. It also highlights a key difference: random models behave more like balanced systems, while trained models make quicker, more decisive state changes at the cost of less precise energy conservation.

Read More

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

In this work, we investigate the emergence of capabilities in reinforcement learning (RL) by framing them as developmental phase transitions. We propose that the individual components of the reward function can serve as direct observables for these transitions, avoiding the need for complex, derived metrics. To test this, we trained a language model on an arithmetic task using Group Relative Policy Optimization (GRPO) and analyzed its learning trajectory with the Local Learning Coefficient (LLC) from Singular Learning Theory. Our findings show a strong qualitative correlation between spikes in the LLC—indicating a phase transition—and significant shifts in the model's behavior, as reflected by changes in specific reward components for correctness and conciseness. This demonstrates a more direct and scalable method for monitoring capability acquisition, offering a valuable proof-of-concept for developmental interpretability and AI safety. To facilitate reproducibility, we make our code available at \url{github.com/ilijalichkovski/apart-physics}.

Read More

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

We estimate parameters from real-world data (e.g., adversarial success rates, detection latency, patching delays) and simulate attack propagation scenarios across 8 sectors (enterprise, retail, trading, development, customer service, academia, medical, and critical infrastructure AI tools).

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Read More

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

In this work, we investigate the emergence of capabilities in reinforcement learning (RL) by framing them as developmental phase transitions. We propose that the individual components of the reward function can serve as direct observables for these transitions, avoiding the need for complex, derived metrics. To test this, we trained a language model on an arithmetic task using Group Relative Policy Optimization (GRPO) and analyzed its learning trajectory with the Local Learning Coefficient (LLC) from Singular Learning Theory. Our findings show a strong qualitative correlation between spikes in the LLC—indicating a phase transition—and significant shifts in the model's behavior, as reflected by changes in specific reward components for correctness and conciseness. This demonstrates a more direct and scalable method for monitoring capability acquisition, offering a valuable proof-of-concept for developmental interpretability and AI safety. To facilitate reproducibility, we make our code available at \url{github.com/ilijalichkovski/apart-physics}.

Read More

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

We estimate parameters from real-world data (e.g., adversarial success rates, detection latency, patching delays) and simulate attack propagation scenarios across 8 sectors (enterprise, retail, trading, development, customer service, academia, medical, and critical infrastructure AI tools).

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Read More

Jul 28, 2025

Local Learning Coefficients Predict Developmental Milestones During Group Relative Policy Optimization

In this work, we investigate the emergence of capabilities in reinforcement learning (RL) by framing them as developmental phase transitions. We propose that the individual components of the reward function can serve as direct observables for these transitions, avoiding the need for complex, derived metrics. To test this, we trained a language model on an arithmetic task using Group Relative Policy Optimization (GRPO) and analyzed its learning trajectory with the Local Learning Coefficient (LLC) from Singular Learning Theory. Our findings show a strong qualitative correlation between spikes in the LLC—indicating a phase transition—and significant shifts in the model's behavior, as reflected by changes in specific reward components for correctness and conciseness. This demonstrates a more direct and scalable method for monitoring capability acquisition, offering a valuable proof-of-concept for developmental interpretability and AI safety. To facilitate reproducibility, we make our code available at \url{github.com/ilijalichkovski/apart-physics}.

Read More

Jul 28, 2025

AI agentic system epidemiology

As AI systems scale into decentralized, multi-agent deployments, emergent vulnerabilities challenge our ability to evaluate and manage systemic risks.

In this work, we adapt classical epidemiological modeling (specifically SEIR compartment models) to model adversarial behavior propagation in AI agents.

By solving systems of ODEs describing the systems with physics-informed neural networks (PINNs), we analyze stable and unstable equilibria, bifurcation points, and the effectiveness of interventions.

We estimate parameters from real-world data (e.g., adversarial success rates, detection latency, patching delays) and simulate attack propagation scenarios across 8 sectors (enterprise, retail, trading, development, customer service, academia, medical, and critical infrastructure AI tools).

Our results demonstrate how agent population dynamics interact with architectural and policy design interventions to stabilize the system.

This framework bridges concepts from dynamical systems and cybersecurity to offer a proactive, quantitative toolbox on AI safety.

We argue that epidemic-style monitoring and tools grounded in interpretable, physics-aligned dynamics can serve as early warning systems for cascading AI agentic failures.

Read More

This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.