AI-assisted working assessment; human adoption pending
Initial evaluation: 2026-10-04; last substantive update: 2026-10-12; last attention check: 2026-10-10.

Targeted check on October 10 read the October 7 independent LessWrong discussion of the Cunningham paper. Earlier coverage is retained; this is not a fresh broad attention or citation audit. The October 12 evidence update added two Epoch AI pilot studies of AI research automation (7 and 8 October); it was not a new attention search or a model rerun, and the ratings are unchanged.

Download Markdown source · Editable repo source (collaborator access)

Human input and prioritization discussion · Companion paper discussion

Related pages: Interactive explorer of the Cunningham et al. models and calibration · Isitcredible AI review (blind, 30 Oct–2 Nov 2026)

AI-assisted working evaluation: The Economics of Recursive Self-Improvement + What If Automating AI R&D Triggers an Intelligence Explosion?

Evaluation date: 2026-10-04

Last substantive update: 2026-10-12 (evidence update; no model rerun)

Last attention check recorded: 2026-10-10 (targeted check of discussion of the Cunningham paper)

Storage reconciliation: 2026-10-12

AI-assistance disclosure. This working assessment supports The Unjournal’s prioritization process. Its judgments and ratings have not been adopted as a commissioned human evaluation or a team decision. A human evaluator should verify the factual claims, methods, and ratings before signing or submitting an official evaluation.

This update adds the case for evaluating the reliability and decision value of the economic framework, a threshold sensitivity example, and the distinction between historical average and current marginal productivity gains. The attention update checks one recent independent discussion; earlier coverage is retained. The underlying lab data have not been independently validated.

Evidence update, 12 October 2026. A new section summarizes two Epoch AI pilot studies of AI research automation published on 7 and 8 October. This is an evidence update, not a model rerun: no calibration was reproduced, and the provisional ratings are unchanged.

Papers

Executive assessment

The Cunningham et al. paper is a useful conceptual advance because it replaces the vague claim “AI can improve AI, therefore an intelligence explosion follows” with explicit feedback conditions. Its core point is correct and important: automation of AI R&D is not sufficient for self-sustaining acceleration. The relevant question is the total gain around the feedback loops, including diminishing returns and bottlenecks.

The paper's formal framework is much stronger than its calibration. The graph/elasticity representation is a clean way to organize a complicated system, but the empirical threshold is built from parameters that are only loosely measured. The paper itself is appropriately cautious, calling the calibration tentative and emphasizing large uncertainty and possible model misspecification.

The most decision-relevant number is the comparison between a roughly 15% AI-R&D-productivity uplift per one-unit capability increase needed under the paper's current calibration and a rough 9% observed uplift inferred from reported AI-engineer productivity since coding agents. This comparison should not be treated as a precise “distance to takeoff.” The 15% threshold depends on the mapping from algorithmic efficiency to the Epoch Capabilities Index, assumptions about effective R&D effort, and the production structure. The 9% estimate is not a clean causal elasticity: it combines changing models, tools, tasks, inference use, researcher behavior and lab processes.

The policy-facing Chan et al. paper is substantially more confident and urgent in tone. It points to fast-rising internal AI-R&D automation, including Anthropic's self-reported 26% “AI leads” share in August 2026, and argues that an intelligence explosion could compress years of AI progress into months or less. This is a plausible scenario that deserves analysis, but AI R&D automation share is not the same object as feedback-loop gain. A model can perform a larger share of current R&D tasks without causing proportionately faster frontier progress, especially if automated tasks are lower-leverage, if humans/experimental compute/data remain bottlenecks, or if productivity improvements substitute for rather than multiply total effort.

The paired evaluation therefore changes the weight I would place on the public argument: the technical paper strongly supports measuring the relevant elasticities and bottlenecks, while it does not establish that the threshold for self-sustaining acceleration is likely to be crossed on any particular timeline. The white paper's specific policy proposals rest on additional judgments about probability, timing, harms, institutional effectiveness and intervention costs that are not estimated by the Cunningham model.

Claim map

Claim 1: Full AI-R&D automation does not by itself imply self-sustaining acceleration

Assessment: strong.

Cunningham et al. explicitly distinguish R&D automatability from self-sustaining acceleration. In the simplest model, algorithmic efficiency can improve without its growth rate increasing if ideas become harder to find. In richer models, human labor, experimental compute, inference compute or data can bottleneck the feedback loop.

This is an important correction to popular formulations of recursive self-improvement. It is also broadly consistent with semi-endogenous growth theory: a technology can contribute to research without overcoming diminishing returns.

Claim 2: Feedback strength can be represented as products and sums of local elasticities around loops

Assessment: strong as a local organizing framework, with scope conditions.

In the baseline model, the total elasticity of new algorithmic improvements with respect to the stock of algorithmic efficiency combines a direct self-feedback term with a capability-mediated loop. The capability-mediated contribution is the product of the elasticity from algorithmic efficiency to capability and the elasticity from capability to algorithmic improvements.

This is mathematically natural under differentiable production relationships. It is useful because it exposes where a seemingly strong loop can fail: one weak edge can reduce the whole product.

The limitation is that these are local elasticities. Their values can change sharply with scale, task mix, complementary inputs and time lags. The graph is therefore not by itself a forecast. When edges change endogenously or production has strong complementarities/non-separabilities, static local elasticities can be a poor guide to dynamics far from the calibration point.

Claim 3: With the paper's current calibration, self-sustaining acceleration requires roughly a 15% increase in AI R&D productivity per one-unit increase in capability

Assessment: useful back-of-the-envelope threshold, low-to-moderate empirical precision.

The paper measures capability using the Epoch Capabilities Index and assumes algorithmic efficiency and training compute enter capability through “effective compute.” It estimates a capability elasticity with respect to algorithmic efficiency of roughly 6.5 and combines this with an estimate of returns to aggregate R&D effort.

The 15% threshold is not a structural constant. It moves if:

A particularly useful sensitivity figure would show the threshold for a broad grid of the three core elasticities rather than foregrounding one 15% crossing point.

For a simple illustration, the threshold for the research-effort response is 1 / (return to research × capability response). With the paper's return-to-research term of 1 and capability response of 6.5, this is 0.154; doubling the first term gives 0.077, approximately 8%, below the reported 9% uplift. This is an algebraic sensitivity example, not an alternative fitted estimate. A defensible recalibration must use compatible definitions, periods, and scales for all inputs.

Claim 4: Current AI-R&D productivity uplift is roughly 9%, below the 15% threshold but increasing

Assessment: directionally informative; numerically fragile.

This is the weakest empirical link in the headline calibration. The paper itself says there is almost no evidence that maps cleanly to the elasticity of R&D effort with respect to AI capability.

A reported engineer-productivity uplift can differ from the needed elasticity for many reasons:

The 9%-versus-15% comparison is therefore best read as an order-of-magnitude diagnostic, not a measured margin of safety.

There is also a distinction between the historical average and the current marginal response. Spreading an accumulated productivity gain across several model generations estimates an average response over that interval. It need not identify the response to an additional capability improvement today, especially when that response is changing. A below-threshold historical average cannot by itself establish that today's feedback loop remains below the threshold.

Claim 5: Anthropic's 26% “AI leads” statistic is evidence that the feedback loop is strengthening

Assessment: relevant evidence, but not a direct measure of the required elasticity.

Anthropic's September disclosure says Claude “leads” 26% of measured internal AI R&D work as of August 2026, meaning the AI completes most of a task end-to-end from a high-level prompt under human supervision. Anthropic also reports that no measured category is fully autonomous. The measure is internally constructed and, as reported publicly, not an independent audit of frontier R&D productivity.

The trend is important because it shows rapid change in task-level automation. But three quantities must not be conflated:

  1. share of tasks with substantial AI participation;
  2. average researcher/task productivity uplift;
  3. elasticity of frontier AI progress with respect to AI capability.

Only the third maps directly to the core feedback condition. A rise in the first can occur with a modest second, and a large second can still translate into a small third if automated tasks are not the bottleneck.

Claim 6: If the threshold is crossed, years of AI progress could be compressed into months or less

Assessment: possible within some model paths, but timing is not established by the threshold alone.

A local elasticity above one can imply an increasing growth rate in simplified models. How quickly this becomes an “intelligence explosion” depends on how far above the threshold the system is, starting levels, lags, reinvestment, bottlenecks and whether elasticities remain high as the system changes.

The Cunningham paper defines “intelligence explosion” very strongly as capabilities going to infinity in finite time, while the Chan et al. paper uses the term more operationally for extreme acceleration. The paired evaluation should keep these definitions separate.

The technical model therefore supports the logical possibility of rapid acceleration under strong enough feedback. It does not itself establish a date or a high probability for the white paper's “years to months” scenario.

Narrow versus broad capability: a major strength

One of the best features of Cunningham et al. is the explicit distinction between narrow AI-R&D capability and broad economically valuable capability. A system might become very effective at benchmarked algorithmic optimization without comparable gains in open-ended scientific judgment, robotics, deployment or social/economic impact.

This matters for the policy-facing paper because some downstream consequences require more than narrow R&D acceleration. Claims about large economic, cybersecurity, biosecurity or geopolitical effects need evidence about the transmission from narrow research capability to broad capability and real-world deployment.

The narrow/broad distinction should therefore be carried through more prominently into public summaries of “intelligence explosion.”

For safety decisions, broad economic output is an incomplete endpoint. An evaluation should check whether strategically important capabilities could improve rapidly before broad economic transformation. Limited aggregate economic impact would provide little reassurance in that case. This is a question for the evaluation, not a finding established by the calibration.

Bottlenecks and substitution

The model explicitly includes important bottlenecks: humans, inference compute, experimental compute and data. This is a strength. The main empirical problem is knowing the relevant substitution elasticities.

If research inputs are strong complements, a cheap and abundant AI-researcher input can have sharply diminishing value once experiments, human approval, data collection or training infrastructure bind. If substitution is easy, the feedback can be much stronger.

The current calibration does not pin this down. A serious follow-up should prioritize lab-level production data: expenditure shares by input, task-level throughput, experiment queues, bottleneck time, inference usage, and attribution of frontier-relevant advances.

New evidence: Epoch pilot studies of AI research automation (evidence update, 12 October 2026)

This section adds public evidence published after the 10 October update. It is an evidence update, not a model rerun. Both studies are early, small pilots, so the provisional ratings below are unchanged. The figures come from Epoch's public pages as reported in search results and secondary coverage; the full reports were not read end to end for this update.

InnovationEval: can an agent rediscover a recent ML method?

Epoch AI's *Can AI automate AI R&D yet?* (InnovationEval, early results, published 7 October 2026) tests whether frontier agents can independently devise a machine-learning improvement that matches a recent human-developed method they have not seen. The target was on-policy self-distillation (SDPO, posted to arXiv on 28 January 2026), scored against a tuned GRPO baseline. Each agent had about 3,000 GPU-hours.

Epoch reports the following:

What this means for the paired evaluation:

A repeated design could make this kind of test more informative for the calibration: several held-out innovations from different subfields; repeated seeds and attempts per model, with selection rules fixed in advance; blinded expert review of novelty; and human-only, AI-only and human-plus-AI arms with matched compute and time. Repeating it across model generations would give a time series closer to what the elasticity requires, though still measured at the task level.

Can AI automate Epoch?

Epoch's *Can AI automate Epoch?* (published 8 October 2026) gave six frontier models a small suite of assignments drawn from Epoch's own work, including graphics, data insights, data-centre research and research design. Epoch graded the outputs manually against its internal standards. Secondary coverage (Crypto Briefing) reports average scores between about 42% and 65%, with no model reaching the standard of Epoch's own staff on average.

Epoch reports that models handled many bounded computational tasks well but struggled with implicit editorial standards, with designing experiments that measure what they claim, and with interpreting flawed results: they treated flawed pilot results as key findings. One reported example is a research-design pilot in which many model responses hit a token limit and the model still treated the result as informative. This example and its exact counts were not independently verified against Epoch's report in this update.

Limits for this evaluation:

What this means for the paired evaluation: the study supports separating four quantities that the public debate often merges: tasks that can be delegated, researcher time saved, validated discoveries, and the resulting frontier progress. Neither Epoch study estimates the frontier-progress elasticity that determines feedback strength in Cunningham et al. Like Anthropic's 26% “AI leads” figure, both are task-level evidence. They are useful for designing the measurements an evaluation should ask for, not for locating the system relative to the threshold.

Relation between the two papers

The two papers are complementary but should not be treated as if the technical model directly proves the policy paper's premises.

Directly supported by the Cunningham framework:

Additional claims in the Chan et al. paper requiring separate evidence:

Those claims may be reasonable, but they are not outputs of the elasticity calibration alone.

Value of an Unjournal evaluation

The Cunningham paper is worth considering as a case study of whether economic modeling of AI scaling adds reliable, credible, and useful evidence for safety and governance. One paper cannot settle the credibility of the field, but it can reveal which modeling practices help and which produce conclusions that the measurements cannot support.

The central evaluation question is: Which conclusions about AI acceleration survive defensible changes in measurement and modeling assumptions, and what decisions can those conclusions actually support?

The strongest commissioning case is a focused assessment with complementary expertise in growth modeling and empirical AI-R&D measurement. It should:

  1. Check whether the evidence identifies the required relationship, separating coding throughput, perceived researcher productivity, completed research, and improvements to frontier models. Distinguish past average gains from current marginal gains.
  2. Test substantive assumptions: combining humans, AI labor, and experimental compute into one research input; inferring returns from historical steady growth; and treating algorithmic efficiency and training compute as interchangeable. Report which conclusions survive alternatives.
  3. Assess added decision value against simple trend extrapolation and an explicit bottleneck checklist. Specify a feasible measurement, forecast horizon, and decision, such as the frequency of safety evaluations as AI-assisted research improves. The feedback threshold alone does not supply a timeline or an optimal evaluation schedule.
  4. Assess governance interventions separately. Identifying a bottleneck does not establish the benefits or costs of restricting it, and weak broad economic effects do not rule out dangerous capability gains.

The useful deliverables are a reproduced calibration, a map of conclusions across plausible inputs and assumptions, and a short reliance assessment. Public data can support an initial robustness audit; stronger empirical conclusions may require author materials or lab access. Keep this as one coordinated assessment with the policy paper, while giving the economics paper its own measurement and model tests.

Current attention

The retained attention evidence includes the following coverage, published before the 3 October prioritization pass.

This is now clearly a high-attention public debate. That increases the value of checking the technical chain. It does not increase the evidentiary strength of the calibration.

A targeted check on 10 October also read “Evidence suggests RSI may decelerate”, published on 7 October. This independent discussion questions the productivity proxy and develops an argument about diminishing returns. Its conclusions have not been independently reproduced here. An Unjournal evaluation should add reproducible sensitivity analysis and a clear account of appropriate reliance beyond objections already raised in that discussion.

Highest-value robustness checks

  1. Reproduce every algebraic threshold and verify the graph rules against the underlying production functions.
  2. Publish a full sensitivity surface for the three core elasticities, not just the 15% benchmark.
  3. Replace point estimates with distributions for each elasticity and propagate them to a probability distribution over feedback gain.
  4. Separate task automation share, researcher productivity uplift and frontier-progress elasticity empirically.
  5. Re-estimate the AI-R&D productivity elasticity using lab experiments that randomly vary model/tool access where feasible.
  6. Measure lags between AI-assisted research, successful experiments, training runs and deployed capability gains.
  7. Stress-test strong complementarity with experimental compute, human oversight and data.
  8. Repeat the calibration under alternative capability measures, including METR time horizons and task-specific R&D benchmarks.
  9. Examine whether gains are concentrated in lower-leverage coding/engineering tasks or in advances that causally improve frontier models.
  10. Model the transmission from narrow R&D capability to broad capability separately.
  11. Treat task-level automation results, such as Epoch's InnovationEval, as validation inputs rather than elasticity estimates; specify repeated held-out tests with human-only, AI-only and human-plus-AI arms (see the 12 October evidence update).

Questions for Cunningham et al.

  1. How invariant is the headline threshold to alternative normalizations and alternative measures of capability?
  2. Can you provide the full calculation behind the 9% uplift estimate, including which model transitions, tasks and dates enter it?
  3. Which empirical input dominates uncertainty in the threshold? A probabilistic sensitivity analysis would be especially helpful.
  4. How do time lags between research, experiments and training alter the condition or the time-to-acceleration?
  5. How would the condition change under strong complementarity between AI research labor and experimental compute/human oversight?
  6. Can the graph framework accommodate task heterogeneity where AI automates a large mass of low-leverage tasks but not a small set of decisive bottlenecks?
  7. How should readers translate the Anthropic “26% AI leads” measure into, or deliberately not translate it into, your elasticity?
  8. What evidence would convince you that the loop had crossed the self-sustaining threshold?

Questions for Chan et al.

  1. Which quantitative claims in the white paper rely on Cunningham et al. versus independent evidence?
  2. What is the best current estimate or range for the probability and timing of a years-to-months acceleration, rather than simply the possibility?
  3. How much weight should be placed on Anthropic's internally measured automation index before independent validation?
  4. What evidence connects task-level R&D automation to frontier capability progress?
  5. Which conclusions depend on narrow AI-R&D acceleration versus broad capability acceleration?
  6. For each proposed institutional response, what evidence addresses effectiveness, implementation constraints, costs and unintended consequences? The technical RSI model does not estimate these.

Evaluation ratings

Cunningham et al.

Chan et al.

Bottom line

The strongest conclusion from the pair is not “an intelligence explosion is imminent,” nor “it is unlikely.” It is that the question can be decomposed into empirical objects that are much more precise than the usual debate suggests.

Cunningham et al. gives a good map of those objects. The present measurements are too weak to locate the system precisely relative to the threshold. Chan et al. is valuable for showing why the scenario matters and for assembling evidence of rapidly rising AI use in AI R&D, but some of its strongest implications require additional evidence about the feedback gain, timing, narrow-to-broad capability transmission and institutional effectiveness.

That makes the pair unusually well suited to a public Unjournal evaluation: the disagreement can be shifted from rhetoric about “takeoff” toward concrete measurements that labs and independent researchers can potentially supply.

Sources