The question of when artificial intelligence would begin improving itself has been floating around for decades, and it has become increasingly salient as modern AI improves. It carries a kind of mythology: AI progress takes off, and the hardest problems the world faces fall before the power of an ever-increasing, infinitely-scaling intelligence. So when does it begin?

Maybe it already has. 

Anthropic recently claimed that Claude has measurably contributed to Claude’s own development. That sentence is worth reading slowly. An AI system helped build a better version of itself. If that claim holds, then an early form of Recursive Self-Improvement (RSI) has arrived.

RSI is the process by which an AI system enhances its own capabilities. The system is used to make an improved AI system to replace itself. That new version is then put to work to create its own replacement in a recursive process which continues until no further improvements can be made. The concept has long been considered a theoretical threshold, a line beyond which predictions about AI development become exponentially harder to make.

What has emerged at Anthropic (and possibly at OpenAI) isn’t the dramatic, runaway scenario from science fiction. AI systems contributing meaningfully to their own research and development cycles ultimately means shorter timelines – these systems will be able to surface insights their human counterparts might have taken months longer to find.

RSI and the Singularity

RSI sits at the start of a canonical event in techno-futurism: the Singularity. The term, popularized by futurist Ray Kurzweil, describes a hypothetical point at which technological growth becomes so rapid and so self-reinforcing that the future becomes unpredictable. As with physical singularities, there is an event horizon beyond which we can’t yet see.

This leaves us with questions we aren’t equipped to answer. Does self-improvement accelerate? For how long? When will it decelerate? Will large language models hit a wall, and, if so, will they be capable of designing their successors? Will RSI be the “phase change” in human development it has been speculated to be, or will it just steepen the curve?

Many processes follow the pattern “slowly, then all at once.” If RSI is indeed beginning, the world may be entering the “then all at once” phase of AI development. Intelligence may become too powerful and cheaply deployable for humans to match. That is a complicated and uncertain future.

The Near-Term Paradox

In a world with RSI, a near-term effect for most users may actually be worse, not better. The highest-value use of compute and research resources could shift from generating revenue by answering customer queries to feeding the RSI process itself. The marginal unit of GPU time is worth more inward than outward.

This creates a strange economic dynamic. Consider the energy question: are there industrial power consumers – aluminum smelters, data-intensive manufacturers – where it would make economic sense for an AI company to pay them more than their current customers do, simply to free up grid capacity? This isn’t as crazy as it sounds. When the potential value of accelerated self-improvement is large enough, the cost of acquiring power through unconventional means changes materially. 

With those resource allocation shifts, will there be “intelligence brownouts”? Periods where frontier model capability is effectively rationed because the compute is occupied with the next previously impossible problem? The calculus of who gets access to what, and when, will shift in ways that can be modeled, with so much left uncertain, can’t be predicted. 

Will we, the users, ever even interact with our post-RSI partners? Will we need to? The most powerful models may never reach end users directly – what gets deployed could be scaled-down versions, or a purpose-built models the frontier system helped create. Being the beneficiaries of a hidden intelligence might not be so different from today.

The Model Web, Revisited

In a previous blog post, we suggested that we are moving toward a bifurcation: A human web, increasingly shaped by what AI surfaces and recommends, and a Model Web, purpose-built for how AI systems ingest, share, and act on information. If models become more capable at extracting and synthesizing web information than any human researcher, the natural endpoint is routing most information retrieval through a model rather than doing it directly.

The web has long been a major part of building human connections, directly and indirectly. We interact with it and each other in largely unchanging ways. Ultimately, we are constrained by the data we collect, how we collect it, what static systems can do with it, and how those interactions can affect the real world. In a world where intelligence exists at the edge, the way we connect over the internet may change in unpredictable ways.

It’s not certain that humans will be significant, direct users of the web in the way we are today. The immediate next question is will the models build their own version of the web? Agents don’t need to communicate the same way that systems in the modern web do. It’s not even certain that we should intervene in the creation of the Model Web. Will we even accelerate its development, or will the agentic future be waiting on us?

Embracing the Transition

For businesses, the immediate implication of RSI is resource contention; compute resources will be more expensive, allocation will be more competitive, and the most capable models’ capacity will be partially occupied with their own improvement. Organizations that treat AI access as a commodity will be surprised when it isn’t. The strategic advantage goes to those who understand what they actually need versus what’s merely available.

For regulators, RSI changes the cost-benefit balance in domains where AI has been held at arm’s length – medicine being the obvious example. When a system can design, test, and iterate on drug candidates faster than any human institution, the process of an “adequate safety review” has to be re-examined. The black market for unapproved, AI-derived therapeutics is a predictable outcome if unregulated alternatives are fast and cheap.

For individuals, the near-term may bring little visible change, and then a great deal of it at once. The practical question is less about understanding the underlying technology and more about positioning: Are the tools, workflows, and information environments you depend on being built to embrace the transition from a human web to a shared one?

It’s important to recognize that the future of AI is likely to be less of a steady climb and more of a series of plateaus and breakthroughs. Models may appear to be approaching their limits, only to surge forward when a new technique or architecture unlocks another level of capability. For those living through it, progress may feel uneven. Gradual most of the time, then suddenly transformative.

Ultimately, machines are getting better at getting better. The line between the tool and the toolmaker is blurring, and the question confronting us is no longer whether RSI is possible or when it will begin. It’s what happens next.