Cunningham, Althoff, Halperin, Jabarian, Koh, Ramani, Trammell, Whitfill and Wu 2026, "The Economics of Recursive Self-Improvement" (arXiv:2609.15802v1)

The verdict compares two numbers: the threshold ln g_R / (ln g_A · ε_C,A) and the return to capabilities, ε_R,C ≈ 0.087. With algorithmic progress of 3× a year, research-input growth of 3× a year and a capability elasticity of 6.5, the threshold is 0.154 log points of research productivity per ECI point, and the core loop's gain (condition 5) is 0.56. The inputs come from different periods and scopes: a 2012–2023 pretraining rate of algorithmic progress (Ho et al. 2024), a cross-model slope pooling releases from February 2023 to January 2026, a February 2025 to March 2026 staff-poll uplift, and an assumed research-input growth (pp. 17, 28–29). The print reports this one cell, with no interval and no alternative pairing. The paper calls the calibration back-of-the-envelope and "very rough" (p. 28).

The gain stays below 1 under several changes. The capability elasticity's delete-one-organisation jackknife interval is [4.44, 8.45] (23 organisations), against the 11.54 that would flip the verdict. None of Epoch's 500 bootstrap draws of the index reaches a gain of 1. Taking algorithmic progress and the capability elasticity from one regression with organisation fixed effects gives 0.59–0.61 (jackknife [0.38, 0.81]). And the verdict holds at every research-input growth rate the paper's own component data imply, 2.3× a year and up; the printed route reaches 1 only at 1.86×. A one-regression pairing without fixed effects gives 0.89–0.92, but it assumes one industry-wide algorithm level per date, which organisation intercepts reject (F = 7.01 on (22, 61) df, p < 0.0001).

How far below 1 the loop sits is not shown. Under the paper's premise C = C(A·T), with training compute not falling, the capability path observed over the window (Claude 3.7 Sonnet, February 2025, to Claude Opus 4.8, May 2026) caps the gain at 1.006–1.027 on Anthropic's models and 1.146 on the industry frontier (1.21 in the October vintage), and at 1.200 for a window ending at the March 2026 poll, so the window by itself cannot place the loop below 1. The verdict flips at algorithmic progress of 7.0× a year; rates up to 7.4× (Anthropic's window) or 9.3× (the frontier) are consistent with the observed capability path, while 10× is not (14.97 ECI points a year from algorithms alone, against 13.02–14.53 observed). Propagating the cited sources' intervals gives a probability of a gain above 1 between 0 and 0.63 depending on the prior, 0.11 under a reference prior, and 0.035 under that prior once draws the observed capability path rules out are screened out.

"Currently" is an average over February 2025 to mid-2026; the current local value is unmeasured. The 15% threshold, which the introduction states as a finding (p. 4), is the value at this one pairing: other defensible pairings give 0.142–0.147 (one regression with organisation fixed effects) to about 0.25 (a date-controlled capability elasticity at 3×), and the paper's frontier-scale rate of 2.5× gives 0.185.

The parameter table on p. 17 prints five rates of algorithmic progress: 3× a year (Ho et al. 2024, pretraining), 3.5× and 10× "for the entire stack" (Whitfill, Snodin and Becker 2025; Ho 2026), and 1.5× and 2.5× by compute scale (Gundlach et al. 2025). The calibration uses 3× (p. 28). Its capability measure is exp(ECI), which counts post-training progress. The p. 29 remark that the rate "was estimated in 2024" is offered for the assumed research-input growth; the print gives no argument for 3× over its own entire-stack rates.

The verdict flips at 7.0× a year. The cited sources give a 95% interval of [0.667, 3.614] in log growth, or 1.95–37× a year (Whitfill, Snodin and Becker 2025), and an 80% interval of 2–50× (Ho 2026), both of which straddle the flip; the paper prints the point values only. Rates above about 7.4× (Anthropic's window) to 9.3× (the industry frontier) would have algorithms alone outpace the capability growth observed over the window, which rules out the printed 10× but not a band of 7.0–9.3× that reverses the verdict. Gundlach et al. (2025), also cited, write that a single rate is ill-defined without a reference algorithm and a compute scale.

Algorithmic progress carries about 55% of the loop gain's log-variance under independent priors on the four inputs, the largest share of any input.

The calibration takes the capability elasticity from "the slope of a regression of ECI on log(compute) across models using data from Epoch AI", which "yields an estimate of roughly 6.5" (p. 29). Pages 24–25 identify that elasticity with "precisely the object estimated by scaling-law studies" and treat it as "the best-measured of the four". On Epoch's public data the cross-model slope is 6.445 per natural-log unit of training compute (June 2026 vintage, 86 models, 23 organisations), and 6.33–6.91 across four vintages. Under the paper's premise C = C(A·T) the elasticity is the slope on effective compute; a cross-model regression on physical compute also picks up algorithmic progress, because later models have both more compute and better algorithms.

In the exact identity that splits the slope, 3.85 of the 6.445 is the date-trend term: the ratio of ECI's date trend to compute's, weighted by compute's R² on date. The other 2.60 is the slope within release date, weighted by the rest of compute's variance (compute and release date are correlated at 0.60). Holding release date fixed gives 4.04–4.10 in nine specification-vintage cells, each of whose jackknife intervals excludes 6.5; the difference from the uncontrolled slope is 2.37 (jackknife [1.10, 3.64], p = 0.0008). A noise regressor that shares only compute's date trend gets a slope of 4.1. Model-family fixed effects give 5.5 (June vintage) and 4.5 (October). Across 410 specifications, 6.5 sits at the 95th percentile, and no specification controlling for date or model family reaches it. Error in the compute estimates does not account for the gap: lifting 4.09 to 6.5 would take an error of 0.51 orders of magnitude on every model, which would put the uncontrolled slope at 8.4.

The paper attributes values "around 6.5" to Ho et al. (2025) and Epoch AI (2025) (p. 25). Ho et al. estimate within model families; converted with their own Table 1, their within-family slope is about 4.1 ECI units (2.9–4.9 across families), and their all-models slope about 5.4 [4.7, 5.9], a design they call biased upward "since algorithmic improvements would make you need less compute". Epoch AI (2025) is the index page, which prints no slope. The print does not report this difference in design. Keeping post-training progress in the slope, on the ground that exp(ECI) counts it, does not fit the pairing either, since the 3× rate paired with it is a pretraining rate (issue 2).

At the printed 3× rate, a date-controlled elasticity raises the threshold to about 0.25 and lowers the loop gain to 0.35, which strengthens the verdict there; with algorithmic progress re-estimated as well, the result turns on the pairing in issue 1.

The 9% on p. 29 is ln 4/16 ≈ 0.087: Anthropic staff's self-reported uplift of about 4× from Claude Mythos Preview, divided by the 16 ECI points between Claude 3.7 Sonnet and Claude Opus 4.8. The uplift is the median of a March 2026 poll of research staff (Favaro and Clark 2026; the Mythos Preview system card reports a geometric mean "on the order of 4×") about a model that has no public ECI; the span belongs to other models. Mythos Preview's system card places it above Claude Opus 4.6, so on any endpoint the public record permits, the per-point value is at most 0.101–0.105, and on the printed route the gain stays at or below 0.66–0.68. The system card's own conversion of the same uplift into an overall progress multiplier below 2× gives at most 4.3% per ECI point (ln 2/16).

The ratio is a total change over 15 months. Agent harnesses, user learning and inference compute all changed over the same window, so it is not the partial elasticity the condition needs. The uplift is a staff-poll statistic with no published dispersion, which its source (Favaro and Clark 2026) expects overstates the true uplift. Page 29 uses the figure as an upper case: the 4X is "very likely to be an overestimate, but even if it were true" it would imply 9%. The introduction states that "this return has been around 9% since the launch of coding agents" (p. 4), without that qualifier.

The system card says the gains it can identify behind Mythos Preview are "confidently attributable to human research, not AI assistance" (Anthropic 2026b, printed p. 43). The paper does not report that sentence. The 18 June draft recorded the attribution; it is absent from later drafts and from v1.

The balanced-growth identity on p. 20 sets the return to research equal to the ratio of algorithmic-progress growth to research-input growth, and with both at ln 3 the return is 1 (pp. 28–29). The identity is exact only on a balanced growth path; off it, a simulation gives 0.77. The inputs come from different periods and scopes: the 2012–2023 pretraining rate, 2022–2025 headcount growth and global chip production.

Erdil, Besiroglu and Ho (2024), which the paper cites, give posterior medians of 1.08–3.54 for software domains, measured with author counts rather than the paper's mixed research input; the naive ratio overstates threefold only for Stockfish (1.01–1.40× elsewhere). The plausible band therefore widens in both directions, although returns above about 1.8–2.0 would imply more algorithmic ECI growth than the window shows. The paper flags the balanced-growth caveat (p. 20) and the timing (p. 29) and lists off-path measurement as future work (p. 32).

The abstract's statement that "feedback loops are not currently strong enough" rests on condition (5), the core loop running from algorithmic efficiency to capabilities to research effort and back. The full conditions (4) (p. 14) and (8) (p. 37) add the economic loops and the inference-efficiency loop as further non-negative terms, each of which can only make acceleration easier. Neither term is sized. At the printed point the inference term would flip the condition if it exceeded 0.437, and the economic loops would flip condition (4) if capability caused roughly half of lab revenue growth per ECI point, a share no source sizes. The paper sets the economic loops aside as slower and for simplicity (p. 15), while calling their firm-level effects "self-evidently already occurring" (p. 13).

For the inference loop the paper writes that METR's expenditure-horizon results "suggest the inference loop is currently weak" (p. 38). That study (Cunningham et al. 2026, whose first author is the paper's corresponding author) measures autonomous agents on one task, speeding up NanoGPT training, in preliminary runs. Its cost axis is total agent cost, with experiment compute rather than inference making up about 70–90% of most runs' cost, and its horizons move up to 120-fold across its own human-cost assumptions. Its text describes the newer models' returns as still rising roughly log-linearly into the thousands of dollars. It does not measure the inference term in condition (8).

The introduction states "ample evidence that this return is not constant and has been increasing of late" and draws from it that the model "does not rule out" self-sustaining acceleration in the near future (p. 4). The return is research productivity per unit of capability, and the paper never estimates it at two dates. Rising uplift levels are what a constant return produces as capabilities rise. Fits of the uplift path through Anthropic's successive staff polls (the internal surveys in the Claude Opus 4.5 and Opus 4.6 system cards, Anthropic 2025a, 2026a, and the Mythos Preview poll, Anthropic 2026b) give a curvature of 0.54–1.16 against constancy at 1; the current local return reaches the threshold only at a curvature of 1.73–1.99, and no fit puts it above 0.100, against the 0.154 needed. Anthropic's August 2025 survey (Anthropic 2025b), in which staff recalled +20% a year earlier against +50% now, makes every fit that admits it convex (8 of 12 quadratic variants are convex, all 6 that include the recalled point); that point is a recalled figure on a different instrument.

The two figures printed beside "upward trend is clear" in the parameter table (p. 22), 1.4–2X from METR's survey (Becker 2026) and roughly 4X from the March 2026 Mythos Preview poll, measure early-2026 use by different populations with different instruments. They contain no time trend, and a population gap of 2.0–2.9× would remove the contrast. The same table cell's lines-of-code series (8X, 2024 to the second quarter of 2026) is a rise within one lab, which its source calls "almost certainly an overstatement" of the productivity gain. The comparison of GPT-1 to GPT-2 with GPT-4 to GPT-5 (p. 29) has no source, spans 2018–2025, and would need index values that Epoch does not publish for GPT-1 or GPT-2.

The hedged forms hold: the abstract's "appear to be strengthening" and p. 29's "these elasticities are likely increasing" are conjectures nothing in the record contradicts, and self-reported uplift and code output did rise within one lab. The wording changed between drafts: the 29 June draft said "we suspect", and the 13 July draft and v1 say "ample evidence"; the calibration inputs are the same in all three. The p. 4 "does not rule out" holds, but on the capability-path ceiling and the algorithmic-progress band in issues 1 and 2, not on the rising-return grounds the paper gives.

The paper cites evidence of AI's contribution to AI research in its parameter table (pp. 22–23) and in the list headed "AI systems are meaningfully contributing to AI progress" (p. 30). The figures come from their sources. The sources attach qualifications that the paper does not carry:

• the 64% for Claude Mythos Preview against humans (p. 30; against 51% for Opus 4.5 in November 2025, which the paper prints as 50% for Claude models released in 2025) is on moments the source "deliberately picked" where the human's choice had room for improvement, judged by a Claude model, and the source says it "isn't a like-for-like comparison" (Favaro and Clark 2026);
• the 8X code merged per engineer (p. 22) is, in the same source, "almost certainly an overstatement" of the productivity gain;
• the GPT-5.5 system card, cited for low pass rates on complex debugging problems (p. 23), says the model "did not meet our thresholds" for High capability in AI self-improvement (OpenAI 2026a);
• the Mythos Preview system card, the source of the 4X on pp. 29–30, attributes the gains it can identify to human research (Anthropic 2026b).

The 100-fold figure (pp. 29–30) is verbatim in OpenAI's 9 July 2026 GPT-5.6 launch post: the share of research compute devoted to internal coding inference "grew 100-fold" over six months (OpenAI 2026b). The same passage continues: "These adoption metrics do not measure research progress on their own". The paper reads the figure as "suggesting that the value of AI in R&D has dramatically increased" (p. 30), and the post has no entry in the reference list.

The paper balances the list with "Best evidence against", and its own Figure 8 (p. 38) shows autonomous algorithmic contributions rising across model generations in two labs, at speedups of 1–1.5%. With the qualifications restored, the evidence of a strong current AI contribution is weaker, which sits with the paper's verdict rather than against it.

The introduction names three biggest unknowns: the growth rate of algorithmic efficiency in labs, the shares of lab R&D expenditure going to each input, and direct measures of how much models contribute to research (p. 4). Under independent priors on the four inputs, algorithmic progress carries about 0.55 of the loop gain's log-variance, the research contribution 0.20, the capability elasticity 0.18 and research-input growth 0.07. The first and third named unknowns are among the leading ones.

The capability elasticity, which the paper leaves off the list and calls "the best-measured of the four" (p. 25), carries about as much as the research contribution, and its printed value is not established as the elasticity (issue 3). The input-share unknown ranks last or second depending on a prior for the inference share that has no anchor: last at shares up to 0.10, second at shares up to 0.30. Whether the requested data can feasibly be shared is untested, and the cost-share requests carry the labelling problem in issue 10.

The parameter table gives γ ≈ 0.15–0.3 from Villalobos and Atkinson (2023) (p. 22, footnote marker 7). The derivation, γ = (1–2)/6.5, exists only in the LaTeX source (3_Data.tex, line 303): the print shows the footnote marker with no footnote text. Page 22 does say that γ combines estimates from two separate studies. Because γ is (1–2)/6.5, γ times the capability elasticity equals 1–2 by construction. Villalobos and Atkinson's trade-off is convex, and its two arcs bracket γ at 0.077–0.31; the printed range is the upper half.

The route ε_R,C = γ S_K (pp. 21–22, 24), with S_K the inference share of research inputs, is algebraically correct for the automated-researcher channel, but it captures only that component of the return to capabilities. At plausible inference shares it gives 3.5–34.6% of the 9%, and it reaches 9% only if the inference share is at least 0.29 (at γ = 0.3) to 0.58 (at γ = 0.15). The parameter table and the data requests call it "one of the two components" of the return. No number in the verdict uses this route, which the paper introduces with "e.g." and "Suppose" (pp. 21, 24).

The paper's best guess is a superelasticity close to zero, because the capabilities-compute relationship has looked "close to log-linear across many orders of magnitude" of training compute (p. 25; also footnote 6, p. 15). Over the calibration's own 5.6 orders of magnitude of compute, log-linearity is not rejected. The full-sample curvature with date controls is negative (σ between −0.10 and −0.15), which would only strengthen the verdict, but it is carried by three runs in the top compute decile (Grok 3, GPT-4.5, Grok 4) and vanishes when that decile is dropped (+0.004 to −0.019). The data do not sign σ.

Ho et al. (2025) show log-linearity within three LLaMA families and name non-linearity as a way their model might fail; the index page (Epoch AI 2025) says nothing on it. The term that enters the condition, σ/(1 − ε_A,A), is never bounded, and Appendix B (p. 39) states the term without that scale and with the wrong sign at the scalar reduction.

Gundlach et al. (2025) find that measured pretraining efficiency gains at the frontier came mostly from scale-dependent innovations. Two innovations account for 91% of gains when extrapolated to the 2025 compute frontier, and the same sequence of innovations gives about 100× at small scale against about 10,000× at frontier scale. They describe their experiments as "relatively small and non-exhaustive", and the study measures training-compute efficiency of language models over 2012–2023 against an LSTM reference, leaving post-training and reasoning gains outside it.

The paper renders this study three ways: algorithmic improvements "may require increases in training compute" (p. 30); "algorithmic progress has historically depended on continued compute scaling" (p. 4, unhedged, with no citation on the page); and Gundlach et al. "show conclusively" that the scale-free model (6) is misspecified (Appendix A, p. 36). The calibration assumes the scale-free premise C = C(A·T) (p. 24), and the introduction does not note the tension. The p. 4 sentence points the same way as the verdict: the study's rates in place of 3× lower the core-loop gain to 0.470 (2.5×, the frontier-scale rate the paper attributes to the study on p. 17), 0.411 (the source's own 2.23× a year) or 0.208 (1.5×, small scale).

Page 29 illustrates the threshold: Claude Opus 4.8 "scores one ECI unit more than Claude Opus 4.7", so "Anthropic would meet the self-sustaining acceleration condition" if adopting Opus 4.8 raised research productivity by 15% over Opus 4.7. In Epoch's index the gap is 1.93 in the 22 July 2026 vintage and 2.30 in the 30 October vintage; the 10 June copy has no Opus 4.8, and 0.4% of Epoch's 500 bootstrap draws put the gap at or below 1.0. The authors' June 2026 copy was neither deposited nor archived, so the printed reading cannot be matched to its source. On the printed threshold the step would need about 35% (July vintage) to 42% (October) higher research productivity, computed as exp(0.154 × gap) − 1; reading log points as percentages, as the paper's 15% does, it would need 30–35%. The threshold's inputs are industry-wide, so the one-lab reading ("Anthropic would meet") is not derived from them. The verdict does not depend on the illustration.

The deposit is the arXiv LaTeX source (TeX, bibliography, TikZ macros and one figure). It holds no code, no data file and no calibration script; the calibration itself is arithmetic on printed values. The regression behind the capability elasticity of "roughly 6.5" is unverifiable as published: the print gives no sample, log base, standard error, vintage or code (p. 29). The authors' June 2026 copy of Epoch's index, from which they read the 16-point span and the one-unit gap, was not deposited and has no archived copy. Estimation error in the index is not carried into any printed number.

The authors' GitHub repository, linked from the title page as "Latest version", holds earlier drafts and an online appendix of survey and experimental instruments; arXiv v1 does not refer to that appendix, and no printed number rests on it.

The main conditions, the cost-share identity, the growth-spurt limits and footnote 9's invariance of the loop gain to the normalization of capabilities are correct; at a point, the invariance holds more widely than stated, for any increasing reparametrisation of the index; for the calibrated loop gain it is exact only for affine changes of the index, and approximate otherwise when both elasticities are measured over the same capability range. The defects are local:

• the restatement of Davidson et al. (2026) on p. 14 omits their irreducibility condition;
• Appendix B (p. 39) states the superelasticity term without its scale and gives the wrong sign at the scalar reduction;
• p. 6 labels ε_A,A "the rate of diminishing returns", inverting its meaning;
• the "revealed-preference bound" for firms that do not optimise (p. 20) has no stated direction;
• p. 28 writes the growth of inference compute, g_K, where the input is N = K·C^γ; the omitted term is conservative for the verdict.

• Future compute constraint (p. 4). Future compute growth "may become constrained by power, capital, or broader economic growth" is a conjecture with no citation on the page. By the paper's own definition (Table 1, p. 3), self-sustaining acceleration holds exogenous inputs, training compute among them, fixed, so a compute constraint bears on it only through scale-dependent algorithmic progress (issue 12).
• Political and regulatory constraints (pp. 4, 31). The support printed on p. 31 is at least 20 data-centre projects cancelled in the first three months of 2026 and $85 billion of projects cancelled over three years (The Economist 2026), with no denominator, and an unsourced sentence on policy reactions to Mythos. The figures concern compute build-out, not regulation of model deployment. The model has no political node, and p. 31 routes the effect through revenue and scaling, that is, through the uncalibrated economic loops. Page 31's "stop it in its tracks" is stronger than p. 4's "could be substantially slowed".
• Signs of a feedback loop (p. 2). The cited Anthropic piece (Favaro and Clark 2026) reports Anthropic data: a fixed small-model training-speedup task rose from about 3× (Claude Opus 4, May 2025) to about 52× (Claude Mythos Preview, April 2026), which the source says "should not be read as a real-world training speedup", and code shipped per engineer rose about 8×. The paper's sentence is more cautious than its source. The same developer's system card says the recent bend in its capability trajectory does not reflect AI-attributable 2× acceleration (Anthropic 2026b). All of this evidence is developer-reported.

Footnote texts 7 (p. 22), 8 (p. 23) and 10 (p. 27) are missing from the print; footnote 7 carries the derivation of γ (issue 10). The GPT-5.6 launch post quoted on pp. 29–30 has no reference entry. The 13 July draft listed an industry affiliation (Google AI) for one author; the v1 title page does not carry it. The v1 title page lists METR, Stanford, Virginia, Carnegie Mellon, Columbia, MIT, Stanford DEL and Epoch AI, and Yale, with all authors at the Elasticity Institute, administrative and financial support from METR and hosting from Constellation.