We created a web game to demonstrate the economics of AI research and development. You play an AI lab working to bootstrap an artificial superintelligence from scratch, investing labor, compute, and data into R&D until you are able to achieve self-sustaining acceleration.

The game is inspired by a recent paper, The Economics of Recursive Self-Improvement, as well as other foundational research papers from economics and computer science.

The game is built on the actual economic models from those papers, but is not meant to be a realistic forecast. The models depend heavily on their parameterization, and in the game, the parameters are calibrated for pedagogy rather than predictive accuracy. To understand how the course of the future could depend on some of the relevant parameters, we created an explorer to dig deeper into the underlying models.

AI development is complex, fast-moving, and hard to predict, but it has obeyed some statistical laws (particularly the scaling laws governing model training) with surprising fidelity. We are excited about the potential for games and simulators to help us find and understand new useful models for the trajectory of AI research.

Background

Understanding the trajectory of AI capabilities is one of the most important questions for predicting the future. In particular, understanding how AI itself accelerates AI research—often called recursive self-improvement—might be the most important component to understand, since it could lead to sharp inflection points in the rate of improvement.

We are interested in ways to quantify recursive self-improvement and predict its trajectory, and are particularly excited about the Economics of Recursive Self-Improvement paper, which came from a recently-formed group of economists (including Tom Cunningham at METR) called the Elasticity Institute. We’re excited about their approach, and created the game and explorer to help understand the model and some of its implications more intuitively.

The explorer provides an interface for visualizing and interacting with all of the models in the paper. The game draws on ideas from the paper to create a dynamic model that also incorporates ideas from compute-optimal training, R&D-based models of growth, scale-dependent algorithmic progress, and weak-links in automation.

Takeaways

The game and explorer are tools that can be helpful for understanding the inputs and constraints of recursive self-improvement. Here we share a few insights gained from engaging with the work mentioned above and developing these tools:

Weak links dominate. AI research uses complementary inputs: human researchers, compute, and data. If intelligence is plentiful, other factors may still bottleneck progress. Recursive self-improvement may be compute- or data-constrained. This could particularly be true if algorithmic progress continues to be dependent on increasing scale, as observed in Gundlach et al. (2025) (another model incorporated into the game).

Even if you could build an AI that is better than any human AI researcher, it would still be limited as a researcher by its access to compute for experiments and training, as well as by data (at least as long as the current paradigm holds).

Recursive self-improvement may come in spurts. It is possible for AI to experience self-sustaining acceleration for a period, and then stop long before reaching the physical limits of intelligence. In fact, this seems likely if compute remains a bottleneck.

We might have a "narrow" intelligence explosion first. We might achieve recursive self-improvement first through "narrow" capabilities (specific to AI research or optimization) that don't fully generalize.

Predictions depend on parameterization. The economic model outputs depend on parameters called elasticities, which tell you how much a quantity increases in response to an increase in a given input. The critical elasticity powering recursive self-improvement is the elasticity of the rate of discovery to current model capabilities. It is the product of other elasticities and dependent on other inputs, and may change over time. This makes tracking up-to-date metrics for these values important.

Conclusion

The future may look very different depending on how the speed of AI progress changes. To better predict where we are headed, it is important to understand this progress.

Early work has provided economic models for measuring recursive self-improvement. However, there are many questions related to the pace of progress that are still difficult to answer. In these cases, metrics and models can provide important information that helps calibrate responses. We are interested in work that pushes the frontier on modeling recursive self-improvement, and hope that our game and simulator make recursive self-improvement dynamics more intuitive.

Please reach out to [email protected] and [email protected] if you are working on similar topics!

Thank you to Tom Cunningham, Basil Halperin, Nate Rush, Will Robinson, Kevin Liu, Chris Tonetti, Hart Lambur, transmissions11, and Dave White for feedback.

Disclaimer: This post is for general information purposes only. It does not constitute investment advice or a recommendation or solicitation to buy or sell any investment and should not be used in the evaluation of the merits of making any investment decision. It should not be relied upon for accounting, legal or tax advice or investment recommendations. This post reflects the current opinions of the authors and is not made on behalf of Paradigm or its affiliates and does not necessarily reflect the opinions of Paradigm, its affiliates or individuals associated with Paradigm. The opinions reflected herein are subject to change without being updated.