When Intelligence Stops Being the Bottleneck: AI and Recursive Self-Improvement

Recursive self-improvement is the mechanism behind many of the most dramatic forecasts about advanced AI. An AI becomes better than humans at AI research, helps create a more capable successor, and thereby produces a system even better at designing the generation after that. Follow the loop far enough and an intelligence explosion can seem almost inevitable.

The loop, however, contains more than intelligence. Models still have to be trained. Experiments have to be run and interpreted. Chips have to be fabricated. Electricity has to be generated. Biological hypotheses eventually encounter cells, animals and patients. Becoming better at thinking does not make every other input to technological progress disappear.

The more interesting possibility may therefore be not an unlimited intelligence explosion, but a repeated migration of scarcity. If AI makes sophisticated reasoning cheap and abundant, the bottleneck moves elsewhere: towards compute and evaluation, then towards experiments, energy, manufacturing and the stubbornly physical business of asking reality whether an idea actually works.

Editorial image generated by the author.

The Recursive Improvement Assumption

Parts of the imagined feedback loop already exist, although not yet in the strong sense of an AI independently designing, training and deploying a more capable successor. The Darwin Gödel Machine, for example, iteratively modifies its own coding-agent implementation and retains variants that perform better on coding benchmarks. Its foundation model is not itself being recursively retrained; the system is improving the machinery built around that model. Anthropic’s own 2026 assessment similarly describes AI taking over a growing share of AI engineering and research while stressing that full recursive self-improvement has not yet been achieved and is not inevitable.

Several quite different developments are easily compressed into the phrase “AI improving AI.” A model can write code used in a training system. An agent can modify its own tools or workflow. AI researchers can use models to design experiments, while more autonomous systems can propose and execute experiments themselves. None of these is quite the same as a system autonomously producing a fundamentally more capable foundation model and then handing AI development over to that successor.

The transition is nevertheless becoming less hypothetical. Anthropic reports that Claude can already match or outperform skilled humans at executing some well-specified AI experiments, while larger gaps remain in deciding which goals and research directions are worth pursuing. In an April 2026 experiment, Claude-powered agents proposed hypotheses, ran experiments and iterated on an open AI-safety research problem. The qualification is as revealing as the result: humans still chose the problem and created the scoring rubric, and the task had a clear performance floor and ceiling.

That hints at an important condition for rapid recursive improvement. Intelligence helps, but so does having an environment that can tell the system cheaply and reliably whether its latest idea is actually better.

Why Software Could Move Fast

Software is unusually favourable terrain for this kind of acceleration. Code can be generated, executed, tested, discarded and replaced without waiting for anything to be manufactured. Thousands of alternatives can be explored in parallel, provided enough compute is available, while better algorithms can make existing hardware more productive without changing a single transistor.

Google DeepMind’s AlphaEvolve offers a useful glimpse of the mechanism. It uses language models to propose programs, automated evaluators to score them and an evolutionary process to preserve and modify promising solutions. DeepMind reports that one result sped up a kernel used in Gemini by 23 per cent, reducing Gemini training time by about 1 per cent. AlphaEvolve has also produced improvements deployed elsewhere in Google’s computing infrastructure.

One per cent sounds modest until it is applied repeatedly to one of the world’s most compute-intensive industrial processes. More importantly, the causal loop is the one recursive-improvement arguments care about: AI discovers a software improvement; the improvement makes AI training cheaper or faster; some of those savings can then be used to develop better AI.

There are many places where similar gains might be found: training algorithms, optimisers, data selection, synthetic data, model architectures, memory systems, inference strategies, distributed computing, compilers and low-level kernels. None requires an AI to build a new factory before benefiting from its own work. Software-based improvement could therefore move much faster than ordinary industrial development for some period, but its success would create another problem: once generating experiments becomes cheap, running them becomes expensive.

The Bottleneck Moves

Imagine an AI research system capable of proposing ten thousand worthwhile experiments in the time a human team currently needs to devise a few dozen. That is an extraordinary increase in research capacity only if those experiments can actually be performed. A model can describe an experiment in seconds while the training run required to evaluate it may occupy expensive hardware for hours or days. The researchers are no longer short of ideas; they are short of compute.

More capable AI could attack that constraint as well. It might devise smaller experiments that predict the results of larger ones, improve scheduling across data centres, identify which experiments carry the most information or discover another round of algorithmic efficiencies. AlphaEvolve’s work on training kernels and data-centre scheduling already belongs to this category. Anthropic describes a related organisational effect: as AI makes producing engineering work cheaper, human review and the choice of which work is worth pursuing become more important constraints.

Removing one bottleneck therefore need not make the system unconstrained. It can simply expose the next one. If machine intelligence dramatically expands the supply of competent scientific and engineering work, technological systems may become much better at generating plausible things to try. The cost of finding out which of them actually work then matters more.

The World Is Expensive to Query

Generating possible answers can become cheap much faster than obtaining reliable feedback from reality. Software again provides the favourable case: a program can be run against a test suite, and many mathematical or algorithmic problems can be checked against explicit rules. AlphaEvolve depends on this property; its proposed programs can be run and scored using quantitative evaluators.

The same asymmetry appears in scientific discovery. I have explored it more directly in AI’s ability to expand the search space faster than experiments can validate what it finds. The relevant point for recursive improvement is broader. In AI research, the evaluator may be a benchmark or a training run. In much of physical science, the evaluator is eventually reality itself.

An AI might generate thousands of promising materials or molecular structures, but prediction is not observation. Some candidates have to be synthesised and measured. Drug development makes the problem especially clear: computational systems can assist discovery, but candidate drugs still pass through laboratory and preclinical work before clinical research in people. Those stages can themselves become better designed and faster, but biological observation does not automatically accelerate in proportion to the speed at which hypotheses are generated.

This does not create a clean boundary between fast software and slow physical science. Better simulations, surrogate models, digital twins and automated laboratories can pull parts of physical experimentation back into computational space. An AI that can reliably simulate a thousand candidate designs may need to manufacture only the most promising ten. Better experimental design can extract more information from each physical test. Advanced AI would, in that sense, repeatedly try to make the world cheaper to query.

The simulation still has to remain connected to what it represents. If a model is weakest precisely in the regime where a genuinely new design becomes interesting, more computation can produce increasingly confident predictions without producing new evidence. Somewhere in the process, reality has to be consulted.

The cost of that consultation varies enormously. A program may return an answer in milliseconds. A semiconductor design may require fabrication. A biological experiment may require cells to grow. A medical question may require following patients over time. An economic policy may never offer a clean controlled experiment at all. Fields with cheap, objective feedback loops may therefore accelerate far more rapidly than fields in which evidence arrives slowly, noisily or ambiguously. Scarcity has not disappeared; it has moved from producing hypotheses towards obtaining trustworthy information.

AI Does Not Need a Body

Hardware introduces a different misconception: that AI would need humanoid robots, autonomous mines and self-operating factories before it could participate materially in its own physical expansion. Humans do not usually manipulate advanced technology that way either. A semiconductor engineer does not personally etch a processor. Designs move through software, simulations, engineering organisations, fabrication plants, logistics networks and highly automated machinery. An advanced AI could operate through much of the same infrastructure.

There are already narrow examples. AlphaEvolve proposed a modification to an arithmetic circuit used for matrix multiplication; after verification, Google says the change was incorporated into an upcoming Tensor Processing Unit. That is nowhere close to an AI autonomously designing and manufacturing an entire frontier processor, but it demonstrates the relevant interface. Software-generated reasoning can alter future hardware without the AI possessing hands.

The first effective “body” of advanced AI may therefore look less like a humanoid machine than the industrial economy surrounding computing. Cloud infrastructure provides computation. Automated laboratories can perform experiments. Semiconductor fabs manufacture designs. Logistics systems move components. Humans remain in whichever parts of the chain have not yet been automated. AI could gain substantial influence over the physical economy through these interfaces long before robotics achieves anything resembling general human dexterity.

Physical production still does not behave like software. An improved algorithm can sometimes be deployed almost immediately; a new processor must be fabricated, packaged, tested, shipped and installed. A new data centre requires electrical infrastructure and cooling, making AI energy demand an infrastructure problem as well as a computing problem. If a new chip depends on a different manufacturing process, production equipment may itself need modification.

Greater intelligence could improve all of these processes. It might design better accelerators, more efficient cooling, superior robotics or cheaper manufacturing equipment. Yet the feedback loop becomes longer as it reaches further into physical production. Designs can move at computational speed; matter cannot.

When Intelligence Stops Being Scarce

The decisive threshold may therefore be less precise than the moment when a machine becomes “smarter than humans.” A more useful change to watch is the point at which skilled cognitive labour stops being one of technological development’s principal scarce resources.

Researchers have finite careers. Engineers can investigate only so many alternatives. Programmers can write and review only so much code. Scientists can read only a small part of the literature around their own fields. Organisations discard potentially useful questions constantly, not because the questions are worthless but because nobody has enough time to pursue them.

If increasingly capable AI systems can perform substantial portions of research, engineering and analysis in parallel, that constraint could weaken dramatically. A research organisation able to investigate a thousand ideas seriously where it previously investigated ten does not require an infinite intelligence explosion to behave very differently from any research organisation that came before it.

Its new abundance would make the remaining scarcities more visible. Compute matters more when researchers can productively consume enormous amounts of it. Laboratory capacity matters more when hypotheses arrive faster than experiments can test them. Electrical generation matters more when computation can absorb every additional megawatt. Manufacturing capacity matters more when designs improve faster than factories can implement them.

Even time cannot always be substituted away. Concrete still has to cure, biological systems still have to develop, patients sometimes have to be observed and fabrication lines still have finite throughput. More capable reasoning may find ways around some of these constraints, but success at one level exposes whatever is limiting the system underneath it.

Recursive improvement therefore does not have to approach infinity to transform technological development. It only has to make intellectual work abundant enough that systems previously organised around scarce human expertise begin running into different limits.

Superhuman machine intelligence might not abolish scarcity so much as reorganise it. For most of technological history, sophisticated reasoning has been expensive and slow. If AI makes it abundant, compute, energy, evidence, manufacturing capacity and physical time become correspondingly more important. The defining feature of advanced AI may not be that intelligence has escaped every limit, but that intelligence itself is no longer the limit.