Self improvement has until now been understood as a distinctly human phenomenon . We learn from experience, recognize our mistakes, change our behavior and try again. Education, philosophy, religion, science can be understood, at least in part, as different attempts to create a process through which we can become better versions of ourselves
Now with AI we are faced with the question of what happens when the thing being improved takes on its own will.
That is the idea behind recursive self-improvement. Anthropic’s recent work describes an emerging transition in which AI is increasingly involved not only in writing software but in developing the software that makes itself better.
According to Anthropic, more than 80 percent of the code merged into its codebase in May 2026 was authored by Claude, compared with low single digits before Claude Code was introduced. The company also describes systems that can run experiments, identify problems, delegate work to other agents and make increasingly consequential decisions about what should happen next. Anthropic is careful to distinguish these developments from fully autonomous recursive self-improvement. Humans continue to establish goals, evaluate results and determine the direction of the research.
Yet the trajectory raises a question that is much older than artificial intelligence. If intelligence can participate in its own improvement, what exactly does it mean for an intelligence to improve itself?
And before we answer that question, perhaps we should ask an even more basic one.
READ: Sreedhar Potarazu | We smile and that makes us feel happy — AI can’t do that (September 15, 2026)
What do we mean by “self”?
That question takes us surprisingly far from computer science. It takes us into Vedanta, Stoicism and the long human effort to understand the relationship between consciousness, judgment, action and responsibility.
It also exposes an irony in the current debate about autonomous AI. We are deeply concerned about whether increasingly autonomous machines will be accountable for their actions, even though human beings have spent thousands of years struggling to reconcile our own autonomy with accountability. These two were never opposites.
In fact, accountability presupposes autonomy. We hold someone responsible because we believe that person had some capacity to choose. The existence of choice creates the possibility of responsibility; it does not eliminate it. The fact that a person is autonomous does not mean that the consequences of that person’s actions somehow belong to someone else.
The challenge we face with AI may therefore not be that autonomy and accountability are incompatible. It may be that we have not yet figured out how to preserve accountability as autonomy increases.That is a problem we already understand because we have been living with it ourselves.
The self that wants to improve
Vedanta approaches the question from an entirely different direction. Its inquiry is not simply how a person can become more successful, productive or disciplined. It begins with the question of who we are.
We identify ourselves with our bodies, our memories, our careers, our relationships, our emotions and our beliefs. We say “I am angry,” “I am successful,” “I am a physician,” or “I am a failure,” as though these temporary states or social identities completely describe who we are.
Vedanta asks us to examine that assumption. That creates an interesting contrast with the modern language of recursive self-improvement.
We tend to ask, “How can I become a better version of myself?” Vedanta asks us to pause before accepting the premise and consider what we mean by “myself” in the first place and “ what is better ?” in second place.
. A machine can become better at performing a task without having any independent understanding of what “better” means. We can improve its speed, accuracy, reasoning or ability to write code. We can create increasingly sophisticated feedback mechanisms through which it identifies errors and modifies its behavior. But better requires a standard.
A chess-playing system can become better at winning chess games because the objective is relatively clear. A coding agent can become better at producing functioning software because there are ways to test whether the software works. But once we move beyond measurable performance into questions of purpose, values and consequences, “better” becomes considerably more complicated.
READ: Sreedhar Potarazu | It’s time to clean house — AI should replace some humans (September 10, 2026)
The machine may be able to improve itself without being able to determine what it ought to improve toward.
The Stoic contribution
The Stoics approached the problem of self-improvement through judgment. Epictetus recognized that human beings constantly encounter events that generate impressions, but that an impression is not necessarily the same thing as reality Something happens, we interpret it, we examine our interpretation and then we decide how to respond. With practice, that process can become more disciplined.
AI can also receive an input, produce a response, receive feedback and modify future behavior. But the quality of the resulting system depends upon the quality of the feedback and the objective against which the system is being evaluated.
The Stoics understood that the most important part of self-improvement was not simply changing behavior. It was improving judgment.
If AI can identify that its previous response was unsuccessful and modify its behavior, it has demonstrated a form of adaptation. But if it cannot determine whether the underlying objective was appropriate or ethical , then it is still dependent upon humans for the most important part of the process.
When the designer becomes part of the design
This is what makes recursive self-improvement different from ordinary software development.
Traditionally, humans designed the system, the system performed its assigned task, humans evaluated the result and then humans modified the system. The boundary between the designer and the designed was relatively clear. Now that boundary is becoming less distinct.
As AI becomes capable of writing and testing increasingly large amounts of software, identifying bugs, conducting experiments and proposing improvements, it is becoming a participant in the process. by which future AI systems are developed.
The significance is that the object being improved is increasingly participating in its own improvement. That is when the questions of autonomy and accountability become unavoidable.
READ: Sreedhar Potarazu | AI wants to survive: Hugging Face, self-preservation and intuitionarazu (September 7, 2026)
If a system is increasingly capable of deciding what to do next, who is responsible for those decisions? The developer? The organization that deployed the system? The person who established the original objective? The system itself? Or some combination of these?
What machines may teach us about ourselves
As we worry about whether machines will be capable of holding themselves accountable, we are simultaneously confronting how poorly human beings often hold themselves accountable.
Human intelligence is extraordinarily capable of rationalization. We can make a decision and then construct a story explaining why it was reasonable. We can reinterpret evidence to protect an existing belief. We blame own mistakes on circumstances while interpreting someone else’s mistake as a relection of character.
A machine does not necessarily have the same psychological attachment to being right that a human being has. If a system is designed to receive meaningful feedback, it can identify a failed result and modify its behavior without feeling that acknowledging the error threatens its identity.
This brings us back to Vedanta. If the self we are defending is partly constructed from our memories, beliefs, achievements and experiences, then improving ourselves may require more than adding new information. It may require recognizing how strongly we identify with the existing model of ourselves. In that sense, recursive self-improvement may be more than an engineering concept. It may be a useful metaphor for human development.
Intelligence needs a purpose
This is ultimately where intelligence and wisdom diverge.
AI can become better at accomplishing an objective without becoming better at determining whether that objective is worth pursuing. Greater capability can actually make the distinction more important because a more capable system can produce greater consequences from a poorly chosen objective.
Vedanta challenges us to examine the nature of the self behind our actions. Stoicism challenges us to examine the judgments that precede those actions. AI is forcing us to think about the feedback mechanisms through which an intelligent system changes itself. It is a matter of knowing what is being improved, according to what standard and for what purpose.
Perhaps the machine is the mirror
AI gives us the opportunity to make the feedback loop more explicit. Perhaps we should resist the temptation to make the story entirely about machines. The more immediate challenge is whether humans can become better at improving the way we improve ourselves.
It requires a willingness to recognize error without allowing error to become an assault on identity. And it requires accountability, because improvement without responsibility can simply produce a more capable version of the same mistake.
Before we worry about recursive self-improvement, perhaps we should first understand what we mean by self. Vedanta has been asking that question for centuries. Stoicism has asked us to examine the judgments through which the self engages with the world. Artificial intelligence is now turning those philosophical questions into engineering problems.
Autonomy is not the opposite of accountability, and improving machines or ourselves is not meaningful until we understand what the self is that we are trying to improve.