
Paradigm Created a Game: Let You Personally Train AI to Autonomous Evolution, Then Experience Losing Control
Written by: Xiao Bing
Well-known crypto VC Paradigm launched a web game yesterday called RSI Simulator.
Players play as the head of an AI lab, with the goal of building superintelligence from scratch.
You have three resources: human capital, compute, and data, and need to decide how much to invest in R&D, how much in training, and how much in expansion. After each round of decisions, model capabilities change, and these changes in turn affect R&D efficiency, until a certain tipping point where AI begins to accelerate self-improvement, and the curve shoots up sharply.
It sounds like an idle game. In reality, this is an educational tool that turns an economics paper into an interactive experience.
Paradigm uses it to answer a question everyone is guessing but few quantify: How fast will AI self-improvement actually be? Will it explode suddenly? 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.
Simply put: when an AI model is powerful enough to help humans improve the next generation of AI models, a cycle is formed. AI participates in creating stronger AI, and stronger AI participates in creating even stronger AI.
Once this cycle starts, it can theoretically produce exponential acceleration, known as an intelligence explosion.
This is no longer science fiction. In February 2026, OpenAI admitted upon releasing GPT-5.3-Codex that early versions of the model "played a significant role" in creating itself, helping debug the training process, manage deployment, and diagnose evaluation failures. This is the first time a top lab has formally acknowledged that a model substantially participated in the engineering process of creating its successor.
How to Play?
The game part of RSI Simulator simulates resource allocation decisions in an AI lab.
Each round, you need to allocate resources among three directions. Investing more compute can train larger models; investing more human research can improve algorithm efficiency (achieving the same effect with less compute); investing in data collection can expand the boundaries of training materials.
The key mechanism is: when model capabilities improve to a certain level, it starts "helping with research," and your human research efficiency improves due to AI assistance. This is the starting point of the RSI cycle. You need to judge when to switch resources from expansion mode to "let AI research itself" mode.
If you switch too early, the model capabilities are insufficient, AI research efficiency is lower than humans, wasting compute. If you switch too late, you may be surpassed by other (hypothetical) labs.
Paradigm also equipped an "Explorer," allowing users to directly adjust parameters of the underlying economic model, such as the "elasticity coefficient of model capabilities on R&D speed," to observe how the AI progress curve changes under different parameter settings.
This part is aimed at a more hardcore audience, essentially turning the mathematical models of academic papers into draggable dashboards.
Four Insights the Game Wants to Convey
In the accompanying research article, Paradigm summarized four insights gained from the model and the game.
Bottlenecks are always in the weakest link. AI R&D requires three complementary factors: human capital, compute, and data. Even if AI intelligence improves infinitely, physical limits on compute and data still exist. Recursive self-improvement may be limited by compute supply or data quality, rather than intelligence itself.
Self-improvement may be pulsed, not continuous exponential growth. AI may experience a period of autonomous acceleration, then stop far from reaching physical limits, if compute becomes a bottleneck. An intelligence explosion may not be a curve shooting straight to the sky, but rather several urgent steps with plateau periods in between.
Narrow-domain intelligence explosion may arrive before general intelligence explosion. AI may first achieve acceleration in autonomous improvement in the narrow field of "AI research," but this capability may not generalize to all cognitive tasks. You can imagine an AI very good at improving AI training code, but without equivalent breakthroughs in managing supply chains or writing novels.
Everything depends on parameter settings. Model output highly depends on a set of parameters called "elasticity coefficients" (elasticity). Elasticity coefficients measure "when input increases by 1%, how many percentage points output increases." The most critical elasticity coefficient in recursive self-improvement is: when current model capabilities improve by 1%, how many percentage points does the speed of new discoveries increase. This number may change over time, and tracking its real-time changes is crucial.
Why Paradigm?
Paradigm is no longer just a simple crypto VC.
A month ago, Paradigm just completed a new fund raising of $1.2 billion. The direction of this money is completely different from the previous three funds. Co-founders Matt Huang (former Sequoia Capital partner) and Managing Partner Alana Palmedo wrote in the announcement that the fund will "start from crypto and expand to AI, robotics, and other frontier fields."
Since its establishment in 2018, Paradigm has raised over $4 billion cumulatively and invested in landmark projects in the crypto industry such as Uniswap, Optimism, Hyperliquid, and Kalshi.
Today, Paradigm's investment map has obviously crossed boundaries. In March, it led the Series H round of drone delivery company Zipline, in April participated in the Series D round of space defense company True Anomaly, and simultaneously invested in AI research lab Nous Research.
Internal tools are also extending towards AI. Paradigm developed Centaur (AI agent toolkit) and EVMbench (blockchain security evaluation project in collaboration with OpenAI), with one leg standing on-chain and the other already stepping into AI infrastructure.
Against this background, RSI Simulator is not just a PR project. Paradigm is betting $1.2 billion in new funds on the "intersection of frontier technologies," namely AI acceleration, decentralized computing, and automated infrastructure.
Understanding the economics of RSI is understanding the underlying logic of its biggest bet. If AI's self-improvement speed matches the curve of certain parameter settings in the game, the update cycle of the entire technology stack will be compressed, and projects that look three years ahead today may be eliminated in six months.
Paradigm said at the end of the article that they are interested in any work pushing the frontier of RSI modeling and encourage researchers in related fields to contact them. This game is a business card, and behind the business card is $1.2 billion and an investment institution that is transforming.
Game Address: https://www.paradigm.xyz/research/rsi/game
Model Explorer: https://www.paradigm.xyz/research/rsi/explorer














