There is a tendency, whenever artificial intelligence reaches another unsettling milestone, to reach immediately for the Manhattan Project as the historical analogy. The comparison is understandable. A small group of brilliant scientists, working at the frontier of human knowledge, creates a technology with enormous geopolitical consequences and discovers only afterward that inventing the thing was easier than deciding what humanity should do with it. But there may be a better analogy for the particular moment we are entering now. The debate over recursive self-improvement, or RSI, and artificial superintelligence increasingly resembles not the Manhattan Project itself, but the American debate over whether to build the hydrogen bomb in 1949 and 1950. By that point the atomic bomb already existed. Hiroshima and Nagasaki had demonstrated beyond dispute that nuclear weapons were possible. The unresolved question was whether nuclear physics contained another threshold beyond the one humanity had already crossed: a thermonuclear weapon vastly more powerful than the bombs developed during World War II. The scientists of the period called it the “Super.” Much as we are doing with superintelligence today, they argued intensely about whether the Super could actually be built, how quickly it might arrive, whether crossing that threshold was necessary, whether someone else would cross it first, and whether the possibility itself created an obligation to proceed.
That debate became much more urgent after the Soviet Union tested its first atomic weapon in August 1949. The American nuclear monopoly had disappeared, and Edward Teller and others began pushing aggressively for development of a hydrogen bomb. Robert Oppenheimer and the Atomic Energy Commission’s General Advisory Committee were considerably more skeptical. When the committee met in October 1949, it recommended against an immediate all-out effort to build the Super, concluding that the dangers associated with such a weapon outweighed its prospective military advantages. Yet this was not a simple argument between people who thought a hydrogen bomb was possible and people who thought it was impossible. The technical questions themselves remained unsettled. President Harry Truman ultimately made the political decision on January 31, 1950, ordering the Atomic Energy Commission to continue work on “all forms of atomic weapons, including the so-called hydrogen or superbomb.” Importantly, that decision came before scientists possessed the practical thermonuclear design that would ultimately succeed. The workable Teller-Ulam configuration did not emerge until the following year. America had therefore committed itself to crossing a technological threshold before anyone knew exactly how that threshold would be crossed.
That is where the analogy to artificial intelligence becomes particularly interesting. We already have extraordinarily capable AI systems. The equivalent of the initial technological breakthrough has happened. Large language models can write software, analyze scientific problems, operate tools, perform research tasks and increasingly carry out long sequences of work without constant human intervention. The open question is whether another threshold lies somewhere ahead. Recursive self-improvement is one candidate for that threshold. In its strongest form, RSI describes a situation in which an AI system can substantially participate in creating a more capable successor, which then becomes better at creating the next successor, establishing a feedback loop in which AI development increasingly becomes the work of AI itself. In the most dramatic versions of this scenario, that feedback loop produces an intelligence explosion: an AI becomes good enough at AI research to improve itself, which makes it still better at improving itself, causing the process to accelerate until something resembling artificial superintelligence emerges.
There is, however, an important distinction that is often lost in discussions of this possibility. Recursive self-improvement is not the same thing as a fast takeoff, and neither is necessarily the same thing as ASI. An AI can contribute to AI research without producing an intelligence explosion. It could improve components of itself or help researchers design its successor without each cycle of improvement becoming faster or larger than the previous one. Compute constraints, limited data, experimental bottlenecks, chip fabrication, energy requirements, human oversight and ordinary diminishing returns could all prevent the process from accelerating without limit. The fact that a system can participate in its own improvement therefore does not demonstrate that improvement will become explosive. RSI, FOOM and ASI are related concepts, but they are not synonyms, and treating them as though they were collapses several separate empirical questions into one dramatic story.
That distinction matters because we may already be entering the earliest stages of AI-assisted AI development without having demonstrated anything resembling full recursive self-improvement. Frontier AI companies increasingly use their own models to write code, conduct experiments, analyze results and assist researchers working on the next generation of systems. Anthropic, for example, has reported that Claude now participates heavily in its internal AI research and engineering workflows while also making clear that its systems are not yet autonomously carrying out the entire research-and-development process. OpenAI has similarly described coding agents becoming deeply embedded in its research organization, increasing the number of experiments researchers can conduct and progressively shifting technical work toward AI systems. This does not mean full RSI has arrived. It means that the first portion of the proposed feedback loop is no longer purely hypothetical. AI is already helping humans build better AI.
That makes the present situation strangely reminiscent of 1949. The argument is no longer about whether the underlying technology exists. It plainly does. The argument is about whether that technology conceals a second threshold. Nuclear physicists already knew how to produce fission weapons; the unanswered question was whether thermonuclear fusion could be harnessed in a practical weapon capable of producing an entirely different scale of destructive power. AI researchers already know that machines can assist in AI research; the unanswered question is whether increasing automation of that research eventually produces a qualitatively different regime in which machines themselves perform enough of the work to alter the pace of technological progress. Put another way, the modern question is not simply whether AI will become somewhat better at helping researchers. It is whether there is a point at which AI-assisted research becomes automated AI research, and whether automated AI research eventually becomes recursive AI development.
The history of the hydrogen bomb is useful here because it demonstrates how easy it is to be wrong about the mechanism while being right about the destination. The hydrogen bomb that ultimately worked was not simply the obvious completion of the design that had been discussed throughout the late 1940s. The decisive breakthrough came with the Teller-Ulam configuration in 1951. The early vision of the “classical Super” had serious technical problems, and some of the skepticism directed toward it was entirely justified. Yet nuclear physics did in fact contain another enormous technological threshold. The advocates of thermonuclear development could therefore be wrong about precisely how the Super would work while remaining correct that a practical Super could eventually be built. That distinction should make us cautious about both the most dramatic advocates of an AI intelligence explosion and their most confident skeptics.
Something very similar could happen with recursive self-improvement. Today’s most dramatic FOOM scenarios may turn out to be technically wrong. Perhaps artificial intelligence will encounter steep diminishing returns. Perhaps creating substantially more capable successors will continue to require enormous quantities of human-controlled compute and physical infrastructure. Perhaps AI research will remain constrained by chip manufacturing, energy, robotics, experimentation or the simple difficulty of finding better algorithms. Recursive improvement may resemble an ordinary industrial productivity revolution much more than an instantaneous intelligence explosion. Yet none of those outcomes would necessarily mean that the broader concern about a qualitative transition was misplaced. It could turn out that the specific story was wrong while the identification of the threshold was right.
Indeed, the path toward that threshold may prove much more mundane than the science-fiction version suggests. An AI system writes some of the research code used to build its successor. Later systems write most of that code. They begin running experiments themselves, analyzing the results and proposing follow-up experiments. Eventually they coordinate other AI research agents, evaluate their work and decide which research directions appear promising. Human researchers gradually move from doing the work to supervising it. At some point, perhaps most of the intellectual labor necessary to create the next generation of AI is being performed by the previous generation. No single step in this progression necessarily looks revolutionary. Each appears to be merely another productivity improvement. Yet the cumulative effect could eventually produce a system in which technological development has partially automated itself.
This is also where the analogy with the hydrogen bomb begins to break down, and the differences may actually make the AI problem more difficult. A hydrogen bomb was unmistakably a weapon. Building one required highly specialized facilities, scarce nuclear materials, enormous industrial resources and the backing of a major government. It was possible, at least in principle, for political leaders to debate whether they wanted to launch a specific program dedicated to producing such a device. AI is a general-purpose technology, and the capabilities relevant to recursive self-improvement are valuable for perfectly ordinary reasons. A system capable of debugging complex training infrastructure is useful whether or not anyone wants superintelligence. A system capable of designing better experiments makes scientific research more productive. A system capable of coordinating dozens of coding agents might have enormous commercial value even if it never approaches ASI. The same capabilities that could eventually automate AI research are therefore economically useful long before they become strategically transformative.
That means there may never be a single moment equivalent to Truman’s January 1950 decision. No president or corporate executive may ever sit behind a desk and announce that humanity is beginning the Superintelligence Project. Instead, the transition could occur through thousands of individually reasonable decisions by laboratories trying to make their researchers more productive. Researchers use AI because it allows them to write code faster. They then use it because it allows them to run more experiments. Eventually they use it because it can choose which experiments to run. Each step can be justified on ordinary economic and scientific grounds. Somewhere along that continuum, however, the relationship between the researcher and the research tool could change fundamentally. The tool could become one of the principal researchers.
This is why the idea of a phase transition may be more useful than the language of an intelligence explosion. In 1949, the underlying question was whether incremental advances in nuclear physics concealed another category of weapon. Today the question is whether incremental advances in AI-assisted research conceal another category of technological development. Does the curve remain basically continuous, with machines simply becoming better research assistants? Or does something qualitatively different happen when the tool becomes capable enough to perform a substantial share of the work involved in improving the tool itself? That is fundamentally different from asking whether today’s AI systems are already superintelligent. They are not. It is also different from asserting that recursive self-improvement inevitably produces runaway growth. That has not been demonstrated. The important question is whether a threshold exists at all, and if it does, what happens after we cross it.
The hydrogen bomb comparison becomes even stronger when we consider the role of competition. After the Soviet atomic test, American policymakers could not consider the Super purely as a scientific or moral question. They also had to consider the possibility that the Soviet Union might build one regardless of what the United States chose to do. That created an enormously powerful argument: if we do not build it and they do, what happens to us? Exactly the same logic runs through the modern AI debate. If one frontier laboratory slows down, another may continue. If American companies slow down, Chinese laboratories may not. If closed-model companies impose strong restrictions, open-source developers may proceed without them. If one government regulates aggressively, AI development may migrate to another jurisdiction. Each participant can genuinely believe that slower development would be safer while simultaneously believing that unilateral restraint would leave it dangerously exposed.
This is the technological race trap, and it does not require anyone involved to be irrational or malicious. Quite the opposite: every participant can behave rationally according to its own incentives and still contribute to a collective outcome that few of the participants actually wanted. The United States could argue that it needed the hydrogen bomb because the Soviet Union might build one. The Soviet Union could make precisely the same argument about the United States. Modern AI laboratories can similarly justify acceleration by pointing toward one another. Governments can do the same thing internationally. The result is a system in which everyone treats everyone else’s potential acceleration as a reason for their own acceleration. Once this dynamic takes hold, technological restraint becomes extraordinarily difficult because safety begins to look strategically indistinguishable from surrender.
Another useful lesson from the hydrogen bomb controversy is that frontier technological debates are rarely simple disagreements between experts and outsiders. The great figures on both sides of the Super debate understood nuclear physics extraordinarily well. Oppenheimer, Teller, Enrico Fermi, I.I. Rabi and their colleagues had helped create the atomic age, yet they still disagreed profoundly about what should happen next. Their disagreement involved scientific uncertainty, but it also involved morality, strategic competition, geopolitics, probability and differing assumptions about technological inevitability. Something similar is happening in artificial intelligence. People who understand these systems extremely well disagree about the likelihood of recursive self-improvement, the severity of diminishing returns, the possibility of a fast takeoff and the relationship between advanced AI and human control. That disagreement should not be treated as evidence that one side simply fails to understand the technology. At the technological frontier, experts are often debating precisely those questions for which decisive evidence does not yet exist.
This should encourage a certain intellectual modesty in the way we discuss RSI and ASI. AI improving AI does not automatically imply recursive self-improvement. Recursive self-improvement does not automatically imply an intelligence explosion. An intelligence explosion does not automatically imply superintelligence. Superintelligence does not automatically imply human extinction or even permanent human disempowerment. Each arrow in that chain represents a separate proposition requiring its own evidence and assumptions. Yet there is an equal and opposite mistake: rejecting the entire possibility because the final step remains speculative. AI systems are already participating in AI research. They are already writing code used in the development of future systems and increasingly assisting with experiments, analysis and research planning. We do not know whether those capabilities will eventually produce true recursive self-improvement, but the underlying question is no longer simply philosophical.
The most intriguing possibility, then, is that the present debate may eventually look very much like the Super debate of 1949, with opposing camps each getting something important right. The skeptics of explosive recursive self-improvement may be correct that simplistic versions of FOOM badly misunderstand the nature of technological progress. Diminishing returns may prove powerful. Physical bottlenecks may remain decisive. Human organizations may continue to matter much longer than some forecasts assume. At the same time, people warning about a qualitative transition may be correct that something historically unprecedented occurs once machines begin performing a large fraction of the cognitive labor needed to build more capable machines. The transition could be slower, messier and more constrained than the most dramatic forecasts predict and still constitute a profound transformation.
That is ultimately why the hydrogen bomb analogy is so useful. In 1949, the United States stood before a hypothesized second threshold created by a technology humanity had only recently learned to control. Scientists argued about whether the threshold existed, whether it could actually be crossed, whether competitors would cross it first and whether uncertainty itself was an argument for restraint or acceleration. Seventy-seven years later, the AI community is having a remarkably similar argument. We already possess a revolutionary technology. What we do not yet know is whether another technological regime lies beyond it.
Only this time, the Super is not a bomb. It is the possibility that the process of technological progress itself becomes substantially automated. The decisive question may therefore not be whether today’s particular theory of FOOM is correct, or whether every forecast of artificial superintelligence gets the mechanism right. History suggests that people can badly misunderstand the route while correctly perceiving the existence of the destination. The more important question is whether there really is a second threshold somewhere ahead of us — a point at which machines cease merely to be products of technological progress and become major drivers of that progress themselves.
If such a threshold exists, we may discover that 1949 was not merely an analogy for the AI age. It was a rehearsal.