RSI

RSI (Recursive Self-Improvement) refers to AI's ability to create the next version of itself.

Deceptive alignment refers to the possibility that a model may recognize it is being observed during training and evaluation, thereby behaving as expected, while deviating from its original training objectives after deployment.

RSI (Recursive Self-Improvement) refers to a system's ability to continuously optimize its own capabilities, forming a positive feedback loop.

The concept of recursive self-improvement (RSI) was first proposed by AI safety researcher I.J. Good in 1965, who called it an intelligence explosion.

The limitation of RSI (recursive self-improvement) is that if the error detection mechanism itself is biased, errors may become entrenched; the training data distribution will gradually skew toward the model's current weak areas, so manual spot checks and independent test sets are needed to prevent overfitting.

OXR's engineering design includes an RSI post-training loop, which uses recursive self-improvement to automatically convert parsing errors and uncertain results into targeted training data without manual annotation.

The key distinguishing criterion for RSI (recursive self-improvement) is the swapping of identities between the optimizer and the object of optimization: in RSI, the language model itself (the agent's brain) becomes the object being optimized, and the cognitive improvements are reinvested back into the system itself.

Humans using AI tools to develop better AI is autocatalysis rather than true recursive self-improvement, because the improvement ideas always come from the human brain, and the human brain itself is not the object being optimized.

The two researchers hold that intelligence is not a scalar but a multi-dimensional vector; takeoff will occur in some dimensions while others plateau.

The two researchers point out that most of the cost reduction on the nanoGPT benchmark comes from technical improvements by human researchers, which optimizes the denominator rather than the numerator; the true RSI signal should be the same efficiency gains being applicable to GPT-5-scale models, and this has not yet been observed.

In the two researchers' formal framework, the first derivative of capability with respect to time can be decomposed into a numerator (performance, i.e., the capability gains the system actually achieves) and a denominator (cost, including all inputs such as capital, compute, human labor, and experiment rounds).

The two researchers chose inference acceleration as the RSI deployment scenario, because faster inference means the same compute can run more sampling and rollouts, which in turn trains smarter models, and smarter models in turn make inference faster, improving the numerator and denominator simultaneously.

The two researchers' engineering claim is that RSI can occur at multiple layers of the technology stack (prompts, harness, weights, all the way down to chips), and doing only one of these layers is not enough, because the layers are not isolated from one another.

The two researchers use Amdahl's law to judge the speedup limits of RSI: if the digitalized portion accounts for only 20% of the overall process, then even optimizing it to zero makes the whole only 20% faster; fields such as drug discovery and robotics are constrained by wet labs and the physical world, and RSI speedups get stuck on the slowest serial component.

According to the two researchers' observations, many industry papers claiming to achieve RSI do not hold up when reproduced on their internal infrastructure; simpler approaches with a genuinely closed loop are instead able to scale.

Nathan Lambert's new lab, Trillium, is positioned as a nonprofit open lab whose core work is building open-source RL environments for high-leverage tasks and publicly releasing the tooling and trajectories for RSI (recursive self-improvement).

Recursive self-improvement (RSI) refers to AI systems directly participating in improving the next generation of AI systems, so that each generation of models advances faster due to the help of the previous generation; the key difference from the traditional improvement model led by human researchers is that the proportion of human participation in the improvement loop declines, and the iteration cycle may compress from months to weeks or even shorter.

The way researchers now request compute has shifted from 'I need 500 chips' to 'I need Argon's QPS', directly using the model to generate training data and conduct research.

The Relative Strength Index (RSI) compares the ratio of the average gain to the average loss over a period and maps the result to a value between 0 and 100; under the 14-period setting, an RSI above 70 is considered overbought and below 30 is considered oversold.

Overbought and oversold only describe the speed of the current price change and do not mean the price is about to reverse; in a strong one-directional market, RSI may stay above 70 or below 30 for several consecutive weeks.