Written by: Little Cake

Famous crypto VC Paradigm launched a web game yesterday, called RSI Simulator.

Players take on the role of the head of an AI laboratory, with the goal of building superintelligence from scratch.

You have three types of resources: manpower, computing power, and data, and you need to decide how much to invest in research and development, how much in training, and how much in expansion. After each round of decisions, the model's capabilities will change, and these changes will in turn affect research efficiency, until a certain critical point is reached, at which point the AI begins to rapidly improve itself, and the curve skyrockets.

It sounds like an idle game. In reality, this is an educational tool that turns academic economics papers into an interactive experience.

Paradigm uses it to answer a question that everyone is guessing but few are thinking about quantitatively: How fast will AI self-improvement occur? Will it suddenly explode? What factors determine its ceiling?

What is Recursive Self-Improvement?

Recursive self-improvement, abbreviated as RSI, is one of the core topics in the field of AI safety research.

In simple terms: when an AI model becomes powerful enough to help humans improve the next generation of AI models, a loop is formed, where AI participates in creating stronger AI, and stronger AI further participates in creating even stronger AI.

Once this loop is initiated, it can theoretically produce exponential acceleration, also known as the intelligence explosion.

This is no longer science fiction. In February 2026, OpenAI admitted that the early versions of the GPT-5.3-Codex model "played a significant role" in creating itself by helping debug the training process, manage deployment, and diagnose evaluation failures. This is the first time a leading lab has officially acknowledged that a model played a substantial role in the engineering process of making its successor.

How to Play the Game?

The gameplay of RSI Simulator simulates the resource allocation decisions of an AI laboratory.

Each round, you need to allocate resources among three directions. Investing more computing power allows training of larger models; investing more manpower in research can improve algorithm efficiency (achieving the same results with less computing power); investing in data collection can expand the boundaries of training materials.

The key mechanism is: when the model's capabilities reach a certain level, it begins to "help with research," and your manpower research efficiency is improved thanks to AI assistance. This is the starting point of the RSI loop. You need to determine when to switch resources from expansion mode to "letting AI research itself" mode.

If you switch too early, the model's capabilities may not be sufficient, and AI's research efficiency may be lower than that of humans, wasting computing power. If you switch too late, you may be surpassed by other (hypothetical) laboratories.

Paradigm has also equipped a "Explorer," allowing users to directly adjust the parameters of the underlying economic model, such as "the elasticity coefficient of model capability to research speed," to observe how the AI advancement curve changes under different parameter settings.

This part caters to a more hardcore audience, essentially turning the mathematical models of academic papers into a draggable dashboard.

Four Insights the Game Aims to Convey

Paradigm summarized four insights gained from the model and the game in the accompanying research paper.

The bottleneck is always in the weakest link. AI research and development requires three complementary elements: manpower, computing power, and data. Even if AI intelligence improves infinitely, the physical limits of computing power and data still exist. Recursive self-improvement may be limited by computing power supply or data quality, rather than intelligence itself.

Self-improvement may be pulsating, not continuously exponential growth. AI may experience a period of self-acceleration and then stop at a place far from physical limits if computing power becomes the bottleneck. The intelligence explosion may not be a straight line shooting into the sky, but more likely a series of rapid steps with intervals of stagnation.

Narrow domain intelligence explosions may occur before general intelligence explosions. AI may first achieve accelerated self-improvement in the narrow field of "AI research," but this capability may not generalize to all cognitive tasks. You can imagine an AI that is very good at improving AI training code but has no equal breakthroughs in managing supply chains or writing novels.

Everything depends on parameter settings. Model outputs are highly dependent on a set of parameters called "elasticity." The elasticity coefficient measures "how much output increases as input increases by 1%." The most critical elasticity coefficient in recursive self-improvement is: how much percentage increase in discovery speed comes from 1% improvement in current model ability. This number may change over time, so tracking its real-time fluctuations is crucial.

Why Paradigm?

Paradigm is no longer just a simple crypto VC.

A month ago, Paradigm just completed a new fundraise of $1.2 billion. The focus of this money is entirely different from the previous three rounds; co-founder Matt Huang (former partner at Sequoia Capital) and managing partner Alana Palmedo wrote in the announcement that the fund will "start from crypto and extend to AI, robotics, and other cutting-edge fields."

Since its establishment in 2018, Paradigm has raised over $4 billion, investing in iconic projects in the crypto industry such as Uniswap, Optimism, Hyperliquid, and Kalshi.

Now, Paradigm's investment landscape has clearly crossed boundaries. In March, it led a Series H round for drone delivery company Zipline, in April it co-invested in space defense company True Anomaly's Series D, and simultaneously invested in AI research lab Nous Research.

Internally, tools are also extending into AI. Paradigm developed Centaur (AI agent toolkit) and EVMbench (a blockchain security evaluation project in collaboration with OpenAI), with one leg standing on-chain and the other already stepping into AI infrastructure.

In this context, RSI Simulator is no longer just a PR project; Paradigm is betting $1.2 billion on the "intersection of frontier technologies," namely AI acceleration, decentralized computing, and automated infrastructure.

Understanding the economics of RSI is to grasp the underlying logic of its biggest bet. If the speed of AI's self-improvement aligns with the curve set by certain parameters in the game, the entire technology stack's update cycle will be compressed, with projects that seem three years ahead today potentially being eliminated in just six months.

Paradigm stated at the end of the article that they are interested in any work that advances the frontier of RSI modeling, encouraging researchers in related fields to get in touch. This game serves as a business card, behind which lies $1.2 billion and a transforming investment institution.

Game address: https://www.paradigm.xyz/research/rsi/game

Model Explorer: https://www.paradigm.xyz/research/rsi/explorer