Paradigm has released RSI Simulator, a web game built to explore a question that usually lives in research papers and AI safety debates: how quickly AI could begin improving itself, whether that process could break into sudden acceleration, and what factors set the ceiling.
In the simulator, the player runs an AI lab and tries to build superintelligence from scratch. The lab has three resources to allocate — labor, compute, and data. Each round asks the player to decide how much to devote to research, training, and expansion. Model capability changes after each decision, and those changes feed back into research productivity. Once a threshold is reached, the AI starts accelerating its own improvement and the curve turns sharply upward.
An economics paper turned into an interactive tool
On the surface, RSI Simulator resembles an idle game. Paradigm frames it as an educational tool. The goal is to make a hard-to-quantify issue more concrete: how fast recursive self-improvement might move, whether it can arrive as a breakout event, and which variables matter most.
Recursive self-improvement, or RSI, sits near the center of AI safety research. In simple terms, the loop begins when an AI model becomes capable enough to help humans improve the next generation of AI systems. AI then contributes to building a stronger successor, and that stronger system helps build the next one after that. In theory, the loop can create exponential acceleration — the scenario often described as an intelligence explosion.
Paradigm argues this is no longer only a science-fiction thought experiment. In February 2026, when OpenAI introduced GPT-5.3-Codex, the company said early versions of the model had “played an important role” in creating itself by helping debug the training process, manage deployment, and diagnose failures in evaluation. Paradigm presents that as the first formal acknowledgment from a leading lab that a model materially participated in the engineering workflow that produced its successor.
The game centers on resource allocation and timing
The simulator models the basic choices an AI lab faces when assigning resources. Every round, the player must split resources across three directions.
- More compute allows the lab to train larger models.
- More labor in research improves algorithmic efficiency, meaning the same result can be reached with less compute.
- More investment in data collection expands the range of training material available.
The central mechanic appears when model capability climbs high enough for the system to start helping with research. At that point, AI assistance raises the productivity of human researchers, which marks the start of the RSI loop. The player then has to decide when to shift from expansion mode to a mode where the AI is effectively researching itself.
Switch too early and the model is still too weak, so AI-assisted research is less efficient than human work and compute gets wasted. Switch too late and a rival lab — hypothetical inside the game, but strategically important in the model — may move ahead first.
Explorer opens the underlying parameters
Paradigm paired the game with an Explorer that lets users directly adjust the parameters in the underlying economic model. One example mentioned in the article is the elasticity between model capability and research speed. Users can change those settings and watch how the projected AI progress curve responds.
This part is aimed at a more technical audience. In effect, it turns the math from an academic paper into a drag-and-drop dashboard.
Four takeaways from the model
In the related research article, Paradigm summarizes four main conclusions drawn from the model and the game.
The bottleneck sits at the weakest link
AI development depends on three complementary inputs: labor, compute, and data. Even if intelligence keeps rising, the physical limits tied to compute supply and data quality do not disappear. In that framing, recursive self-improvement may be capped by compute or data rather than intelligence itself.
Self-improvement may come in pulses, not as a steady exponential line
Paradigm says AI could go through a period of autonomous acceleration and then stall long before it hits physical limits if compute becomes the bottleneck. That would make an intelligence explosion look less like a single uninterrupted ascent and more like a series of sharp steps separated by plateaus.
Narrow intelligence explosions may arrive before general ones
The article also argues that AI might first achieve accelerated self-improvement in the narrow domain of AI research itself, without that capability spreading evenly across every cognitive task. A system could become highly effective at improving AI training code while showing no comparable jump in supply-chain management or novel writing.
Everything depends on the parameters
Model output depends heavily on a set of variables referred to as elasticities. Those values measure how much output rises when input increases by 1%. In the RSI setting, the critical elasticity is how much an increase of 1% in current model capability raises the rate of new discoveries. Paradigm says that number may change over time, making it important to track in real time.
Why this matters for Paradigm
The article places RSI Simulator inside a broader shift at Paradigm. One month ago, the firm said it had raised a new $1.2 billion fund. In the announcement, co-founder Matt Huang and managing partner Alana Palmedo wrote that the fund would expand “from crypto into AI, robotics, and other frontier areas.”
Since its 2018 launch, Paradigm has raised more than $4 billion and backed crypto projects including Uniswap, Optimism, Hyperliquid, and Kalshi. Its investment footprint now extends well beyond crypto. In March, the firm led the Series H round for drone delivery company Zipline. In April, it joined the Series D round for space defense company True Anomaly and also invested in AI research lab Nous Research.
The same pattern shows up in internal tooling. The article says Paradigm has developed Centaur, an AI agent toolkit, and EVMbench, a blockchain security benchmarking project built with OpenAI. In the wording of the piece, one leg remains on-chain while the other has already stepped into AI infrastructure.
A public signal behind a $1.2 billion expansion
Seen through that lens, RSI Simulator is presented as more than a PR exercise. Paradigm is using its new $1.2 billion fund to back what the article calls the intersection of frontier technologies: AI acceleration, decentralized computing, and automated infrastructure.
The piece argues that understanding the economics of RSI also means understanding the logic behind Paradigm’s largest current bet. If AI self-improvement follows curves similar to some of the parameter settings shown in the simulator, the refresh cycle across the technology stack could compress sharply. A project that looks three years ahead today could be obsolete six months later.
At the end of the article, Paradigm says it is interested in any work that pushes the frontier of RSI modeling and encourages researchers in the area to get in touch. The game, in that sense, doubles as a calling card for a firm in the middle of a broader transition.
Game: https://www.paradigm.xyz/research/rsi/game
Explorer: https://www.paradigm.xyz/research/rsi/explorer