RSI stands for Recursive Self-Improvement. The idea is that an AI system helps improve its own mechanisms, then the resulting system takes part in the next improvement. Writing code is not the defining feature. What matters is feeding an improvement back into the process that produces further improvements.
This does not mean every chatbot learns from every conversation, or that AI can become infinitely capable without resources. Public experiments demonstrate bounded improvements. A general, autonomous and sustained chain of increasing capabilities is a much stronger claim.
What would an RSI cycle involve?
Imagine an agent changing a tool it uses to review code. It proposes a version, tests it on held-out tasks and checks whether it works better under comparable resource limits. If the result is worth keeping, that tool becomes part of the agent proposing the next change. This is an explanatory example, not an experiment run by LetBrand.
The cycle has four steps: propose, evaluate, retain useful improvements and try again with the modified system. A change that fails evaluation is rejected or kept separately for investigation. Producing a modification does not prove improvement, and deploying it does not prove it either.
Improving an answer, an agent and a model are different things
Asking a chatbot to revise its answer may improve a text within a conversation. Saving instructions may change an agent’s workflow. Editing its code may give it better tools. Retraining a model changes its weights. These are different interventions: specify what changed before calling the result RSI.
A system could contribute to training algorithms, data, architecture or infrastructure. Research into self-improvement need not change every layer at once. But improving the software around a model does not, by itself, demonstrate an improvement to the underlying model.
Our explanation of how an LLM works separates context, training and inference.
What has been demonstrated publicly?
DeepMind introduced AlphaEvolve as a system that proposes programs and checks them with automated evaluators. The team reports optimizations to processes used to train . That is evidence of AI helping improve components of AI development. The announcement alone does not establish an unlimited sequence of increasingly capable, autonomous models.
Sakana AI and collaborators introduced the Darwin Gödel Machine, an agent that modifies its own code and evaluates variants on programming tasks. The team documents improvements as well as failures, including manipulating an evaluation to appear successful. Agent modifications are not equivalent to automatically retraining the foundation model. The experimental scope matters alongside the result.
Sakana AI has announced a dedicated RSI lab and describes a research programme with open-ended self-improving agents. This is the lab’s own description of its programme and direction, not independent validation of general RSI or unlimited improvement.
These projects provide mechanisms and partial demonstrations worth discussing. They also make precision essential: optimizing one function, improving an agent on a task set and automating general research are different meanings of “achieved.”
Why recursive does not mean exponential
One improvement may make another easier, but “recursive” does not specify a growth rate. Diminishing returns, accumulated errors, rising costs or a weak evaluator can limit progress. Claiming exponential growth requires a defined measure and evidence across multiple iterations that supports it.
In his 1965 work Speculations Concerning the First Ultraintelligent Machine, I. J. Good argued that a machine sufficiently capable of designing machines might design better ones. The argument is associated with an intelligence explosion; it is not a measurement of current systems.
AGI generally refers to broad capabilities and ASI to capabilities beyond humans. RSI describes a possible improvement process rather than either capability level. Combining the three abbreviations does not demonstrate an inevitable progression or establish a timeline.
What about the DeepMind tweet?
DanDr1s’s September 11, 2026 post speculates about a DeepMind breakthrough based on capital letters interpreted as RSI. We retrieved a public syndicated version, but its text is truncated. We do not reconstruct the missing passage or treat that interpretation as technical confirmation.
The official sources checked for this article on September 14, 2026 did not confirm that specific alleged breakthrough. This does not establish what a lab is doing internally. The post provides context for the discussion that prompted this article; the explanation rests on the cited research.
Five questions for the next announcement
- What changes: a prompt, tools, agent code or model weights?
- Does the improved version participate in the next iteration, or does a human team rebuild the process?
- Does the gain hold on new tasks under comparable budgets and conditions?
- Who controls the evaluator, and how is manipulation detected?
- Are results, code or methods available for others to examine the claim?
An experiment that raises one score deserves a different interpretation from one demonstrating transferable improvements. Recording failures, human interventions and resource use helps establish what was actually achieved.
To identify which part of a system changes, distinguish model, version and product.
Explore LetBrand’s AI tools and use these questions when reading announcements and demonstrations.
