AIs Automate Their Improvement Without Achieving Full Autonomy

AI systems are already automating parts of their own improvement, but human control remains central, and L5-level self-improvement has not been demonstrated. From A‑Evolve‑Training to ForgeTrain, clues are accumulating, while GPT‑6 Astra leads a dedicated benchmark without self-improving.

Human Control Remains and L5 Is Not Achieved

Humans still set the objectives, design the evaluation, provide the infrastructure, and decide on publications. AI operates more within this framework, without taking control of the entire improvement loop. Complete L5-level self-improvement has not been demonstrated: only portions of the chain are currently automated. If it were to occur, the application would be far broader than current AI and could assist in fields such as healthcare, security, or emotional intelligence.

Concrete Systems Automate Key Steps

Several projects show that segments of the improvement loop can be entrusted to AI. Four consecutive cycles of post-autonomous training have been conducted by A‑Evolve‑Training on a 30 billion parameter Nemotron, with the external score progressing from 0.80 to 0.86. Ouroboros modifies a coding agent by intervening on its tools, prompts, context construction, and implementation methods. The Darwin Gödel Machine saw its performance on a subset of SWE-bench increase from 20% to 50%, despite a fixed component in the selection and maintenance of variants. ForgeTrain produced a pre-training framework from an empty repository, automating a phase that is normally highly engineered. Together, these developments outline a trajectory: less manual tinkering, more automated improvement.

The RSI Framework: From One-off Trials to Meta-Learning

The self-improvement framework distinguishes several levels. At level B0, a system can improve a response without retaining the gain. At level L1, it follows a human recipe where goals, methods, and criteria remain defined by operators. L2 allows it to choose a path of improvement under a fixed objective, through prompt, tool, or experimental trials. L3 involves deciding what to learn next by leveraging mistakes to guide tasks, data, or environments. L4 entails learning from actual use and embedding useful changes in memory, tools, skills, or infrastructure. L5 aims at improving the improvement process itself, by building better successors, a step considered difficult. At its core, it is a system that improves and reuses these gains, with the key question remaining the distribution of work between humans and machines and the ability to retain progress sustainably.

What Distinguishes RSI Optimization and Signals to Monitor

The differentiating criterion is the useful retention of lessons and their reuse to make future iterations better; without this, one remains in classical optimization. GPT‑6 Astra does not fall into the category of recursively self-improving systems. However, it aggregates capabilities expected in future models likely to exhibit RSI, such as advanced reasoning, coding, computer usage, scientific research, and execution of long, multi-step workflows. Its leading position in the RSI-Exam benchmark illustrates these traits, without constituting proof of recursive self-improvement.