All walkthroughs · Paper briefs

The Economics of Recursive Self-Improvement

Tom Cunningham, Lukas Althoff, Basil Halperin, Brian Jabarian, Andrew Koh, Arjun Ramani, Phil Trammell, Parker Whitfill, Cheryl Wu
Elasticity Institute paper; draft first posted June 2026, arXiv 2609.15802 v1 submitted 14 September 2026 (paper dated 15 September 2026)

Paper · Brief and rating · Dashboard entry

About 21 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of The Economics of Recursive Self-Improvement, by Tom Cunningham, Lukas Althoff, Basil Halperin, Brian Jabarian, Andrew Koh, Arjun Ramani, Phil Trammell, Parker Whitfill and Cheryl Wu. It is an Elasticity Institute paper, first circulated as a draft in June twenty twenty-six and posted to arXiv in September twenty twenty-six. This is an AI-generated walkthrough with two voices, a host and a discussant, and it is not an Unjournal evaluation or decision. The paper asks whether AI systems that help build the next generation of AI systems could produce a self-sustaining acceleration in AI capabilities. It sets up a family of semi-endogenous growth models, draws them as directed graphs, and shows that whether acceleration happens depends on products of elasticities around feedback loops. It then lists the parameters one would need, reviews what estimates exist, asks AI companies for specific data, and does a rough calibration. The headline: feedback loops do not currently look strong enough for self-sustaining acceleration, but they appear to be strengthening.
Discussant
Recursive self-improvement is a loaded term. What exactly do they mean by it?
Host
They mostly avoid using it as a technical term. They note that some people use recursive self-improvement, which I will call R S I, for any technology contributing to its own improvement, and others reserve it for full automation of AI research. They replace it with a defined object: self-sustaining acceleration, meaning AI systems accelerate capability growth without any growth in exogenous inputs like human labor or training compute. They separate this from an intelligence explosion, which they define as capabilities going to infinity in finite time, and from what they call R and D automatability. On that last one they write: "Automatability does not imply automation". Humans could stay employed in research because of slow diffusion or regulation. They also push back on the old I. J. Good argument that an ultraintelligent machine would unquestionably cause an explosion. In their words, "our models make it clear that such an explosion may not follow if there are diminishing returns" or if feedback loops become bottlenecked.
Discussant
So what is the core quantity?
Host
The authors call it the core feedback loop, and they phrase it as a question: "for a one-unit increase in model capabilities, how much do the capabilities of the next generation of models increase?" The starting point is the Jones nineteen ninety-five model. Let A be algorithmic efficiency, defined as the inverse of the training compute needed to reach a given capability level. The flow of improvements to A depends on R and D labor and on the current level of A. Hold labor fixed. If the elasticity of new improvements with respect to the existing stock exceeds one, the growth rate of A rises as A rises, and you get acceleration from the stock alone. Bloom, Jones, Van Reenen and Webb found this elasticity well below one in traditional industries, and prior studies of AI algorithms find the same without extra loops.
Discussant
And the AI-specific extra loop is capabilities feeding back into research.
Host
Yes. In their baseline model, capabilities, call them C, depend on algorithmic efficiency and training compute. Algorithmic improvements are now produced by human labor and also by AI capabilities. That creates a second loop: efficiency raises capabilities, and capabilities raise the rate of efficiency improvements. The strength of a multi-step loop is the product of the elasticities along its edges. The condition for self-sustaining acceleration in efficiency becomes: the self-feedback elasticity plus the product of the two core-loop elasticities must exceed one. Their graphical contribution is that you can read conditions like this off the diagram: each edge carries a partial elasticity, each loop's strength is the product along the path, and the total elasticity sums across all loops from A back to the flow of new A. Three regimes follow. Below one, growth fizzles toward sub-exponential. Exactly one is the knife-edge of exponential growth. Above one, growth accelerates.
Discussant
Constant elasticities would make that a statement about the whole future. Do they assume that?
Host
No, and they are explicit. Elasticities are local. Acceleration today need not mean acceleration forever. That leads to the bottlenecks model, which adds experimental compute and inference compute as inputs to research, and data as an input to capabilities. The loops stay the same, so the condition is formally unchanged, but the extra inputs give reasons the capability-to-research edge could weaken. Human labor could bind if AI lacks research taste; experimental compute could bind if it strongly complements capabilities; data and training compute could bind on the capabilities side. Inference compute here means the computing used to run AI models, as opposed to training them, so more inference compute means more AI researcher-hours.
Discussant
You mentioned narrow versus broad capabilities. Where does that come in?
Host
Section two point four splits capabilities into narrow ones, useful for improving algorithms, and broad ones, which produce economic output. The core loop only needs the narrow edge. So you could get acceleration in algorithmic efficiency and in narrow capabilities without acceleration in broad capabilities or output. Whether broad capabilities accelerate depends on what they call a superelasticity, which is how the pass-through elasticity from efficiency to broad capabilities itself changes as efficiency grows. If that pass-through falls, a narrow acceleration need not spill over. They are candid that the split is stylized: "The clean cut we draw between capabilities useful for algorithmic improvements and all other capabilities is a strong simplification." More realistically, progress concentrates in tasks that are easy to verify, like coding and parts of mathematics, and they write that "Such progress would have impacts on the real economy but would be exceptionally jagged."
Discussant
And growth spurts?
Host
Section two point five lets overall efficiency depend on two sub-algorithms, with AI helping only one. Suppose the compute costs of the two processes add. Then as the AI-assisted component becomes very efficient, the other becomes the bottleneck, and the elasticity of overall efficiency with respect to the assisted component goes to zero. So you get a temporary spurt that stalls unless AI also starts improving the bottleneck component. The same happens when a component is near a natural ceiling; their example is kernel efficiency at fifty percent, which can at most double. Kernels are the low-level routines that run model arithmetic on the chips.
Discussant
Then the economic loops, which are the macro part.
Host
Section two point six closes the model through output. Higher capabilities raise output, output finances more compute and data, and those feed back into capabilities and research. They stress that output can be read as G D P or as the revenue of a single AI firm, so the loop can operate inside a firm before it shows up in macro data. With several accumulating stocks, they use the Davidson, Halperin, Houlden and Korinek result that acceleration occurs when the dominant eigenvalue of a reduced-form elasticity matrix exceeds one. They state a closed form summing the core loop and two economic channels, assuming capital-like inputs accumulate one-for-one with output. They do not calibrate this richer model.
Discussant
So what do they actually calibrate?
Host
The bottlenecks model, simplified. They assume an R and D effort aggregator that combines human researchers, AI researchers running on inference compute, and experimental compute into one scalar, and that this scalar is the only input to algorithmic improvements. That lets them use historical human data to infer what AI inputs would do. The condition then factors into three terms whose product must exceed one. First, a return to R and D, the elasticity of improvements to effort divided by one minus the self-feedback elasticity. Second, the elasticity of research effort with respect to AI capabilities. Third, the elasticity of capabilities with respect to algorithmic efficiency. The authors give this product a lot of weight. They write: "In our view, estimating this quantity, tracking it over time ... is arguably the most important empirical exercise for assessing whether a self-sustaining acceleration will occur."
Discussant
Take the terms one at a time. How do they get the return to R and D?
Host
On a balanced growth path, the return to R and D equals the growth rate of algorithmic efficiency divided by the growth rate of research effort. Research effort growth is a spend-share-weighted average of input growth rates, using the standard result that under cost minimization and constant returns, cost shares equal output elasticities. They take algorithmic efficiency to triple each year, from Ho and coauthors' twenty twenty-four estimate for pretraining, and research effort to triple each year too, from headcount and compute growth. The ratio of the two logs of three is one. They acknowledge the balanced-growth assumption: "The smooth, steady historical rate of algorithmic progress makes this a reasonable benchmark, though not one we expect to hold if RSI pushes us off the balanced-growth path."
Discussant
And the third term, capabilities with respect to efficiency?
Host
They assume efficiency and training compute enter capabilities symmetrically, through effective compute, the product of the two. Then the elasticity of capabilities with respect to efficiency equals the elasticity with respect to training compute, which is what scaling-law studies estimate. Measuring capabilities as the exponential of the Epoch Capabilities Index, an index that places models from different benchmarks on one scale, a regression of that index on log compute gives an elasticity of roughly six point five. They also argue the superelasticity is near zero, because the capability-compute relation has looked close to log-linear.
Discussant
Which leaves the middle term, research effort with respect to capabilities.
Host
And that one is close to unknown. The paper says: "This is the hardest of the three to measure, since it captures how AI capabilities translate into real-world research effort." Also: "The science of AI capabilities is in its infancy." In the calibration section they put it more bluntly: "We have almost no evidence for this parameter because AI uplift studies do not report capability improvements in a form that maps cleanly to this elasticity." So they solve for a threshold instead. With the return to R and D at one and the capabilities elasticity at six point five, acceleration requires the middle elasticity to exceed about zero point one five. Because capabilities are the exponential of the index, that means each additional point on the Epoch index must raise research effort by fifteen percent. They note that Claude Opus four point eight scores one index point above Opus four point seven, so on their model Anthropic would meet the condition if that upgrade alone raised research productivity fifteen percent.
Discussant
And their estimate of where we are?
Host
A back-of-the-envelope calculation. Opus four point eight scores sixteen index points above Claude three point seven Sonnet, released in February twenty twenty-five alongside the coding agent Claude Code. They take uplift at that time to be roughly zero, citing the Becker and coauthors randomized trial that found AI slowed experienced open-source developers. Anthropic engineers surveyed in the Mythos Preview system card self-reported about a fourfold uplift. The authors say that figure is very likely an overestimate, but even taken at face value, a fourfold gain spread over sixteen points is about nine percent per point, below the fifteen percent threshold. Then they add: "Of course, these elasticities are likely increasing". They also sketch a cost-share route: the middle elasticity equals a quality exponent gamma times the inference-compute share of research spending, but they have almost no data on that share.
Discussant
The rest of the paper?
Host
A qualitative weighing of evidence. In favor of acceleration: the self-reported uplift, model wins over human researchers on some optimization tasks, and OpenAI's statement that the share of research compute going to internal coding inference grew a hundredfold in six months. Against: compute bottlenecks, including running out of G D P to fund scaling; fast diminishing returns to autonomous optimization; dependence on data; and political opposition to data centers. An appendix on inference adds a second sector and notes METR's expenditure-horizon results: "Agents find optimizations cheaply at first, but their returns diminish much faster than the returns to expenditure on human labor". And there is the data wish list: lab-level growth in algorithmic efficiency, spend shares across researchers, experimental compute and inference, and direct measures of how much of technical progress models produce.
Discussant
Before critiques, how has the paper been used and cited?
Host
Modestly so far, as far as we could find. The arXiv version has been up only a few weeks. OpenAlex lists no citing works yet, and the brief for this page records an Altmetric score of three for the arXiv version. Most discussion attached to a July research note on METR's site by two of the authors, Whitfill and Cunningham. In it they call the calibration loose and say it is "meant to be a first draft", and they also write that "we cannot rule out an extended and rapid acceleration in capabilities." There is short coverage in an AI news-alert service, which frames the data requests as putting the burden of proof on labs, and a summary on a Chinese technology news site. One related arXiv paper by Mikhail Burtsev, on recursive criticality, reaches a similar unit-threshold condition independently; we did not confirm that it cites this paper. The page's brief links the paper to a July letter from frontier-lab employees asking the United States government to help pace automated AI development; the letter's text does not mention the paper, so any influence there is unconfirmed. We found nothing supporting an earlier claim that this is the main economics input to that debate. A search miss leaves the question open.
Discussant
Who would act on it?
Host
Safety-focused funders and evaluators use forecasts of capability acceleration to set priorities. Regulators could adopt its wish list as a disclosure template, and labs could report the parameters it names. For an Unjournal evaluation, the attraction is that it turns a vague debate into named elasticities that can be challenged one at a time. The risk is that one threshold comparison, fifteen percent against nine, gets quoted as reassurance that acceleration is not happening.
Discussant
Then let me push on that number. The return to R and D is set to exactly one because two growth rates of log three cancel. The paper itself cites other estimates of algorithmic progress, three and a half times a year for the whole stack and ten times a year from a twenty twenty-six Epoch estimate. If algorithmic efficiency grows tenfold a year while effort triples, the return is about two, the threshold halves to roughly seven percent, and their nine percent estimate is above it. The conclusion flips under an estimate the paper lists.
Host
That is a fair reading of the arithmetic. It cuts both ways, since faster growth in inference compute raises research effort growth and lowers the return. An evaluator should ask for the threshold under the range of cited inputs, not a single point.
Discussant
Second, the nine percent. It uses a self-reported fourfold uplift that the authors themselves doubt, assumes zero uplift in early twenty twenty-five, and treats an engineer productivity gain as a gain in total research effort. Their own text says research output per capability point should be smaller than task uplift because labor is only part of research. So is nine percent an upper bound, or just a rough number in an unknown direction?
Host
The paper does not resolve that. It also spreads the gain linearly in index points over a period when, by its own account, the elasticity was rising, so the current marginal value could be above the average. An evaluator should check consistency between the cost-share formula, gamma times the inference share, and the uplift-based figure.
Discussant
Third, the identifying assumptions. The R and D aggregator assumes AI inputs interact with the stock of ideas exactly as human inputs did historically. Cost shares equal elasticities only under cost minimization and constant returns, and the paper notes lab budgets are financed by investor expectations. The authors argue that if firms do not optimize, "observed cost shares reflect the firm's own belief about marginal products, so they remain informative about the elasticities." Beliefs during an investment boom may be a weak guide.
Host
And the symmetry assumption behind six point five. The appendix acknowledges evidence from Gundlach and coauthors that algorithmic progress has been biased toward large scales. If so, the capabilities elasticity with respect to efficiency and with respect to compute can diverge, and scaling-law estimates identify only the second. The appendix even suggests that under fixed compute, "rapid RSI may still be feasible, but only via some kind of" directed technical change. Footnote nine says the loop gain is invariant to how capabilities are normalized, but the uplift numbers are expressed per index point, so an evaluator should check that the two pieces use the same normalization.
Discussant
What else should an evaluator check?
Host
Four things. First, whether the richer economic-feedback model, which is uncalibrated, would change the verdict, since firm-level revenue loops may already be operating. Second, the local nature of the condition: crossing one does not tell you how long acceleration lasts, given the growth-spurt and ceiling mechanisms. Third, data and code: none is posted yet, and the regression behind six point five should be reproducible from Epoch's public data. Fourth, independence: METR supported the work, the lead author and one coauthor work there, another is at Epoch, and several inputs come from lab system cards. Reviewers would ideally include a growth economist and someone who knows AI training data.
Discussant
A one-sentence summary?
Host
A clear framework for which elasticities govern AI-driven acceleration, with a useful data agenda, attached to a calibration whose bottom line depends on one nearly unmeasured parameter and on input choices the paper's own citations would let you vary.

Sources