We created a web game to demonstrate the economics of AI R&D. You play an AI lab bootstrapping an artificial superintelligence by investing labor, compute, and data. Inspired by a recent paper, The Economics of Recursive Self-Improvement, and foundational research, the game uses actual economic models but is calibrated for pedagogy rather than prediction.

We also created an explorer to dig deeper into the underlying models. Background: understanding recursive self-improvement is crucial for predicting AI trajectory. The paper from the Elasticity Institute (including Tom Cunningham) provides a framework. The game incorporates ideas from compute-optimal training, R&D-based growth, scale-dependent algorithmic progress, and weak links in automation.

Takeaways: weak links dominate; recursive self-improvement may be compute- or data-constrained. If algorithmic progress continues to depend on increasing scale, as observed in Gundlach et al. (2025), bottlenecks remain. Recursive self-improvement may come in spurts, perhaps achieving narrow explosion before generalizing. Predictions depend on elasticities, particularly the elasticity of discovery rate to current capabilities.

Conclusion: better understanding of AI progress speed is important. Metrics and models help calibrate responses. Reach out to Paradigm ([email protected], [email protected]). Thanks to Tom Cunningham, Basil Halperin, Nate Rush, Will Robinson, Kevin Liu, Chris Tonetti, Hart Lambur, transmissions11, and Dave White. This post is for general information only and does not constitute investment advice.